The pillar · v3.3 · 2026-10-05

The GTM World Model

The unified theory of go-to-market, for systems thinkers.

Go-to-market combines accounting, buyer behavior and decisions under uncertainty. This model formalizes it across 6 tiers and 27 state variables, with 26 equations separating what is true by identity from what is causal-but-regime-dependent and what is merely correlational. A decision also needs relevant evidence, identified effects, and a qualified policy for its operating context.

v3.3 Last updated 2026-10-05 32 theses · 34 failure modes · 219 terms

The research update

Comprehensive evidence appraisal converted into corrected equations, qualified claims, nine curriculum-cluster plans and thirteen proposed studies. Breadth retained; unsupported universal causal, ROI and governance claims retired.

Review window: 2026-04-05–2026-10-05. 77 sources; 65 current publications or revisions. Publication, study and first-circulation dates are distinguished in the evidence record.

Cluster improvements → Decision qualification → Sources and limitations →

Improvements by curriculum cluster

These C0–C8 identifiers follow the canonical curriculum. The older editorial hub routes below retain their original C1–C7 identifiers.

C0 · Orientation

Priority: first

An epistemic tag alone cannot govern a decision. “Measurable” means a measurement could exist; it does not mean the current value was measured, that its causal effect was identified, or that a proposed deployment passed qualification. Represent these as separate states. Retain the model's refusal behavior, but tie it to observable conditions rather than the assumed universal 25–40% noise floor.
For each material claim or coefficient, the proposed evidence record contains the population, business kind, outcome and denominator, intervention/control, assignment unit, study and release dates, access depth, uncertainty, scope, causal assumptions, configuration/cost snapshot, falsification criterion, and conditions requiring re-estimation. A policy preference or risk limit is normative and should be labeled that way.
Start with mathematical consistency before fitting new coefficients. A beautifully calibrated implementation of an inconsistent target is still unusable. The correction register below identifies exact model locations.

Measurement contract: Declare eligible population/business kind, target and denominator, pre-action information, policy/comparator/assignment, outcome horizon and coverage, cost/authority/configuration, uncertainty, causal assumptions, falsifier and expiry. Current coefficients/qualification remain unmeasured unless explicitly linked to local evidence.

C1 · Market & Customer

Priority: high

The fit and business-kind layers already prevent some universal SaaS assumptions. Make their distinctions operational. An account can have good end-user fit but weak budget ownership, poor security acceptability, costly implementation, or negative supplier contribution. A PMF score should expose these observations and unknowns instead of absorbing them into one convenient multiplier.
Learn segments on discovery/training data, then freeze their definitions before confirmatory holdout evaluation. Retain segment lineage when policies change; otherwise a better-looking cohort can result from changing the eligible population. Assess fit against real alternatives, including internal workflows and no purchase. Use nonbuyers and failed implementations alongside customer praise.
Synthetic research can expand candidate hypotheses and interview probes. Before using it for segment selection, demand sizing, willingness-to-pay or committee decisions, validate that particular prediction against held-out people. Wrong-person and demographic-only baselines reveal whether rich personas add useful information. Evaluate distributions and subgroup errors, not just convincing explanations or top-choice agreement.
Include real nonbuyers, refusal, disengagement, delay and substitution. Validation on outcome-balanced sales logs must not be read as a natural conversion-rate estimate. A plausible next response, a complete simulated journey and an intervention effect require separate validation.

C2 · Demand & Buyer Psychology

Priority: high

Keep the existing buyer belief and trust graph. Replace the single `B_agent` share as the main observation with a matrix indexed by role, task, and phase. Record platform/interface/version, observed use, representation fidelity, verification path, authority, timestamp, and unknown status. Separate research, recommendation, drafting, guarded execution and permitted purchasing.
Track initial and current shortlists, evidence requirements and unresolved constraints. A buyer can discover a vendor through AI, verify through search or peers, obtain internal approval through humans, and still fail implementation. A citation, a recommendation, a contract, and buyer welfare are different outcomes.
Sample the interfaces buyers actually use and repeat queries. A recent consumer audit finds that API and interface outputs expose different recommendations and source layers; it explicitly excludes business products from its main recommendation sample. Apply the measurement lesson, without importing its consumer magnitudes. ConsumerQ audit
Operationalize the model's political/procurement fit through relevant roles: champion/user, economic buyer, finance, technical/security/legal review, and veto authority. Do not assume every purchase has every role. Buying-group size is a poor substitute for who can block approval and which proof remains missing.
The 95–5 in-market heuristic and the probability that winners were already shortlisted are descriptive quantities with different denominators. Neither estimates that 95% of value creation occurs before contact. Seller leverage should be estimated by phase and intervention, with room for useful evaluation, approval and implementation work.

C3 · Motions & Channels

Priority: high

Choose an acquisition or retention action by expected incremental contribution, feasible change and cost. Conversion propensity, attributed revenue, and treatment effect answer different questions. Allow withholding contact or spend when the expected net effect is negative, and include shared sales/service capacity constraints.
Treat saturation, creative fatigue, platform delivery, rival imitation and growing category demand as competing explanations. A fixed reciprocal CAC curve bakes in a particular asymptote; observed autocorrelated spend may not distinguish response curvature from changing effectiveness. Preserve experiment support and reject extrapolation beyond it without a labeled scenario.
For attribution and brand, keep experimental time paths, not only one ROAS number. Self-reported attribution can inform awareness and mechanisms, but cannot identify a causal lag kernel alone. Test complementarity, organic displacement, delayed conversion and pull-forward. A pause may follow weak demand; screen-and-forecast methods do not remove that confounding.
Specify where outcomes are booked and observed. Demand generated through one source may convert through a marketplace, partner or direct channel. Randomizing exposure does not recover transactions absent from the outcome system. Reconcile observation coverage and linkage, and audit whether matching or reporting failure depends on assignment. An analytical commercially affiliated paper makes this boundary explicit without estimating its field magnitude. Konitzer
AI visibility also needs a denominator: relevant query volume, ICP platform usage, repeated answer occurrence and user notice. GMMM offers a proposed measurement/causal framework with explicit support assumptions and simulated commercial responses. Its measured answer occurrences do not constitute measured incremental sales. Kato and colleagues
Motion choice is constrained by economics, buyer requirements, capacity and product properties. A single universal ACV threshold or the smallest observed funnel rate is an inadequate decision rule. Hybrid motions and multiple simultaneous constraints already fit the generalization layer; express them explicitly.
For partner motions, distinguish sell-in from customer sell-out and reconcile reseller/OEM/platform fees, margin shares, duplicate booking and channel conflict. A partner-generated contract may have different cash timing, customer visibility and service obligations. Those affect both outcome completeness and the contribution assigned to a motion.

C4 · Funnel & Lifecycle

Priority: high

Keep funnel rates for diagnosis, but use a state graph for decisions: discovery, evaluation, approval, contract, implementation, use/value, renewal, and return to evaluation. Include abandonment, rework, review queues, cancellation and involuntary billing failure. Define transitions and time stamps so a faster draft does not masquerade as a faster sale.
Preserve stage definitions and cohort composition across comparisons. Tightening an MQL qualification rule can lower the preceding conversion rate and raise the following one without changing sales. A lower rate may be a useful filter. Report complete eligible-cohort outcomes alongside stage diagnostics, and distinguish a definition change from a behavioral change.
Handoff quality concerns both information and allocation. False-high intent consumes scarce seller time; false-low intent can suppress useful follow-up. Evaluate the complete AI–human policy through the final outcome, rather than rewarding cleaner summaries, application volume or opportunity creation alone.
Distinguish logo survival, payment survival, product use, realized value, satisfaction, contract lock-in and voluntary renewal. Improving observed retention can reflect selective early churn rather than improved individual experience. Churn propensity differs from responsiveness to an intervention; some saves can destroy contribution.
Use complete cohorts and explicit censoring. Report mature observed outcomes separately from forecasts. Do not declare a universal six-to-twelve-month point after which all LTVs become reliable; contracts, seasonality, renewal cycles and sample sizes differ.
Qualify short-horizon proxies against mature outcomes, including directional error, subgroup fragility and interval coverage. A failed-to-reject surrogate diagnostic is not a certificate. Prediction-powered estimation is a candidate when its labeling and stability conditions hold, rather than a universal replacement for outcome measurement. Enrollment drift, few observed events and selective labeling require additional checks. Chen on surrogate metrics
Free use, a cheap trial, paid subscription and active realized value are separate states. Recent newspaper data associate habits with conversion and retention differently across phases, but do not randomize habits and include automatic renewal. Use that pattern to propose experiments rather than importing an optimal engagement threshold. Kreimer and colleagues

C5 · Economics & Metrics

Priority: first

Separate customer stock from recurring revenue, usage, backlog, installed assets and recognized/cash value. Expansion raises revenue without adding logos. Hybrid kinds need distinct timing and contribution ledgers. The model already contains this generalization; the original spine should obey it.
Forecast survival-weighted contribution with time-varying price, usage, delivery cost and discounting. Include contraction, reactivation, returns, credits, bad debt, human exceptions and inference cost. Revenue expansion can worsen contribution. Report payback and contribution distributions where cost tails matter, rather than only mean LTV:CAC.
For AI economics, count all eligible attempts, retries, review, rework, tools/data, allocated maintenance and attributable losses. Separate sunk overhead from avoidable incremental costs. Hours saved produce cash savings when expenditure falls. Productive redeployment may generate incremental contribution, which must be measured separately. Outcome pricing additionally requires an enforceable acceptance definition and allocation of failure risk.
Price, packaging and metering are actions that change acquisition, usage, expectations and renewals. Evaluate immediate conversion together with contribution, bill predictability, complaints and later renewal. An observed willingness-to-pay model and a causal pricing policy are separate objects.
Separate estimation from deployment. A confidence interval describes evidence about the tested effect; choosing a rollout also involves implementation cost, reversibility, downside, capacity and transfer to future conditions. Backtest candidate economic rules against simple empirical-success and current-policy baselines. A recent experiment-archive study supports this comparison, not automatic adoption of its ambiguity parameter. Report both the decision and inferential uncertainty. Farrell, Korganbekova and Misra

C6 · GTM Operations (GTM OS)

Priority: first

The next operational layer should join observations without erasing provenance. Maintain account, contact, contract and usage identities; pre-action feature snapshots; assignment and policy versions; action probabilities where applicable; permission and consent state; source dates; raw receipts; review decisions; downstream outcomes; and missingness reasons.
Attach record age, last verification, linkage confidence and correction history. Measure error by task, record quality and case mix; a base error rate attached only to the agent cannot explain stale contact data or an upstream source outage. Verification freshness and permission freshness can expire separately.
Human approval is a label about a particular criterion, not universal ground truth. Separate style, factuality, fit, compliance, authority and actual business outcomes. All accounts lack their unobserved outcome under the action not received; rejected drafts and untouched accounts may additionally lack factual downstream labels. Counterfactual effects require identification, and selectively missing factual labels require a missingness audit. Down-weighting approvals solves neither automatically; use experiments, audits, sensitivity analysis or bounds.
Track assisted performance and unaided fallback competence separately. Test persistence and anomaly investigation when assistance is unavailable; durable professional skill effects remain unresolved rather than presumed beneficial or harmful.
Make governance separable: detection, action enforcement, correction, evidence preservation, containment and recovery. Hard permissions should be enforceable predicates. A governance score can summarize operations but should not authorize an irreversible action by itself. Compare raw execution evidence with generated reports and verify critical claims independently.
Measure capacity through arrival rates, service minutes, case mix, skills, backlog age and rework. Review fraction is not review load. Organization-wide learning also requires implementation: receiving feedback is not equivalent to changing the policy or improving outcomes.
Evaluate review against matched clean controls as well as defects. Longer review units can reject more good and bad plans at once, so a higher catch rate need not improve discrimination or accepted throughput. Review-window findings from LLM monitors should not be relabeled human fatigue evidence. Han, Yan and Zhang

C7 · Strategy, Regimes & Scaling

Priority: high

The model already distinguishes drift, regime change and topology change. Link those distinctions to a transition process influenced by the firm, rivals, platforms and external conditions. A changepoint alarm identifies a statistical change, not its cause. Test whether policy changes, delivery shifts, selection, competitor response or external shocks explain it.
Separate private contribution from category outcomes. Agent adoption can create congestion, improve relevance, expand demand, change information supply, or shift platform power. The direction of its net externality is empirical. Reply-rate decline alone does not identify pollution caused by agents.
Use cash, capacity, quality, compliance and service promises as potentially simultaneous constraints. Treat strategic objective weights as declared preferences/scenarios, with units and sensitivity. Macro valuation heuristics are not accounting identities or causal policies. Strong strategic feedback may make global convergence difficult; theory does not establish that every GTM market is computationally hard.
Separate external capital conditions and valuation multiples from firm-specific runway, funding access, covenants, investment commitments and management mandate. Track AI readiness and accumulated workflow knowledge separately again. Evidence about investment expectations or organization capital does not identify a universal interest-rate-to-operating-objective coefficient.
Category formation needs independent evidence and information financing as well as promotion. AI-generated consensus can recycle shared sources. Preserve fresh observations and provenance, and test platform/rival responses before scaling a locally successful policy.
Keep destination heterogeneity explicit. A short randomized AI-search experiment reduces publisher referrals, while a separate quasi-experiment associates AI Overviews with more Reddit engagement. These measure different populations, interfaces and outcomes, so neither establishes a universal direction for AI-mediated demand. Test substitution versus complementarity for the actual destination and business endpoint. Wang and colleagues, Zhang, Cui and Zhang

C8 · Agentic & AI-era GTM

Priority: first

Retain autonomy as a descriptive aggregate if helpful. Decisions should use a policy that can execute, review, repair, defer or stop according to task, state, capacity and authority. Reading, drafting, modifying CRM, sending, quoting, committing terms and changing strategy have different consequences.
Define success as an independently accepted terminal outcome plus process/authority validity. A repairable error before external effect is different from an unauthorized action whose output looks correct. Qualify a frozen policy on held-out representative tasks, including measured human-path error, severe-event checks and available review capacity; monitor changing task distributions after release.
Cost per accepted success must include failures in the numerator. Choose the policy objective explicitly: for example, minimize total cost of meeting specified accepted throughput, reliability and risk requirements. Report cost per accepted success as a complementary diagnostic; it is a different objective when policies accept different numbers of cases. No reviewed evidence justifies a universal GTM threshold above 98%, an inverted-parabola optimal review load, or a general promise that hybrid agents have already settled the economic question.
The decisive field experiment compares complete policies on the same eligible population and horizon, through realized contribution, operational burden and downstream customer quality. Current papers supply mechanisms and test designs; a generic autonomous B2B agent CAC coefficient remains unqualified.

When the model can support a decision

A policy is eligible for a scoped decision only when accounting/measurement, causal or predictive support, authority and complete-workflow quality/cost/capacity evidence match the intended target. Unknown material inputs produce an explicit refusal or labeled scenario.

Required evidence

population, eligibility, business kind and stable segment/stage lineage
outcome, denominator, horizon, payment/cost timing and maturity
policy, comparator/no-action, assignment unit and pre-action feature snapshot
study dates, release/revision dates, access depth, provenance and source independence
model/tool/data/interface/policy/authority versions and cost snapshot
uncertainty, support, missingness/censoring, interference and identification assumptions
independent acceptance/process predicates, falsifier and requalification trigger

Hard gates

Current authorized task/action scope and applicable versioned policy requirements
Defined severe process/quality/authority failure predicates
Available service/review capacity and escalation/defer/stop path
Independent critical-action receipts and terminal acceptance verification

Comparison rules

Compare complete eligible workloads and declared contribution endpoints, not selected successes.
Separate CPQO, CAC, revenue, contribution, cash, hours released and realized savings.
Preserve raw receipts, assigned controls, untreated accounts, missingness and failed attempts.
Prediction accuracy and conditional rates do not identify intervention effects.
Surrogate or synthetic-human validation is target-specific; real held-out outcomes and stability requirements still apply.
Qualified terminal success, all-step success, correction success and process validity are distinct metrics.
Randomized, observational, theory, benchmark and vendor evidence retain separate boundaries.

Forecast rules

No universal 25-40% error floor, 6-12-month maturity or fixed review/governance coefficient.
Use target/horizon-specific temporal/cohort holdouts, calibration and decision precision.
Refuse unsupported population/configuration/spend extrapolation or immature/selected outcomes.
Material model/tool/policy/data/eligibility/authority change triggers requalification.

Evidence register

Each record states the design, finding and boundary of the evidence. A consumer study, simulation or theory supplies a scoped hypothesis until its transfer to the target market is tested.

R01 · Wei, Bencina, Nguyen, Ribeiro and Zelditch — From Prediction to Incrementality — 10 August 2026. · 10 August 2026

Design: Full preprint.

Finding: Full preprint. Field intervention; full methods/limitations inspected. Eight-week member-randomized 50/50 LinkedIn comparison of causal scoring, exploration, constrained allocation and withholding versus propensity targeting, with a matched delivery envelope. The reported value lift concerns the complete policy. Static audience definitions and segment lineage protect assignment.

Boundary: one platform, short experiment, no isolation of individual components, no autonomous B2B seller economics. C3/C5/C6.

Read the source →

R02 · Li, Yuan, Yang, Chen and Song — Attributed But Not Incremental — 25 June 2026. · 25 June 2026

Design: Full preprint.

Finding: Full preprint. Production calibration against 18 rounds of channel experiments across up to eight TikTok markets, with rolling forward evaluation and daily hierarchical correction.

Boundary: normalized confidential errors; child-level credit is allocated rather than independently causally identified. A later cannibalization decline is before/after deployment evidence, not a randomized business-impact estimate. Complementarity and spillover remain limitations. C3/C5.

Read the source →

R03 · Heusch — Structural Estimation of MMM Parameters from Geo Experiments — 21 August, revised 7 September 2026. · 21 August, revised 7 September 2026

Design: Full v2.

Finding: Full v2. Synthetic structural study: 156-week fictional retailer, three channels, four four-week go-dark tests plus cooldown. Uses time paths to estimate effectiveness, carryover and saturation.

Boundary: stable parameters, specified adstock/saturation, no spillover and tested spend support. It is not an empirical estimate of SaaS brand decay. C3/C5.

Read the source →

R04 · Filippou, Quach and Jha — Pseudo Incrementality Testing — 16 September 2026. · 16 September 2026

Design: Full preprint.

Finding: Full preprint. Changepoint detection plus counterfactual forecasting, two GeoLift benchmarks and synthetic stress tests. Only one of three nominal 90% intervals covers truth in the stress test; weak controls yield large bias.

Boundary: parameter uncertainty not fully propagated, small validation base, no guarantee that pauses are exogenous. Useful refusal conditions rather than a substitute for experimental design. C3/C5/C7.

Read the source →

R05 · Ningyin Xu — Reference Dependent Pricing — deposited 16 July 2026. · deposited 16 July 2026

Design: Repository and dissertation.

Finding: Repository and dissertation. Full PDF inspected. Randomized base-price histories across 818 food-vending locations during 2024, followed by equal current-price discount conditions in 2025.

Boundary: reference-history effects in this setting; reported approximately 28% profit improvement is a structural counterfactual, not deployed randomized profit uplift. Local competition and crossed-out-price cues limit transfer. C2/C5.

Read the source →

R06 · Dubé, Morozov, She and Tuchman — The Effects of Ads on Beliefs and Implications for Consumer Search — written 31 July, posted 7 August, BFI release 17 August 2026. · written 31 July, posted 7 August, BFI release 17 August 2026

Design: Paper and full PDF.

Finding: Paper and full PDF. Preregistered experiment, 17,775 Prolific participants, simulated bookstore; randomized ads/content/price and separate belief/shopping samples avoid priming. Data are from 2025.

Boundary: lottery-incentivized shopping, unincentivized belief reports, one advertised book, no long-run enterprise brand effect. C2/C3/C5.

Read the source →

R07 · Martin Bichler — Agentic Markets — 6 June 2026. · 6 June 2026

Design: Full journal article.

Finding: Full journal article. Conceptual synthesis of agent search, differentiation, market power and platform architecture.

Boundary: no new calibrated market experiment. Lower technical search costs do not guarantee exhaustive search, equal information, higher welfare or a particular equilibrium. C2/C3/C7/C8.

Read the source →

R08 · Quanyan Zhu — Agentomics — 9 June 2026. · 9 June 2026

Design: Full preprint.

Finding: Full preprint. Formal/normative workflow-surplus and cooperative allocation framework.

Boundary: the security-operations numerical example is hypothetical. Shapley allocation is a proposed normative benchmark, not an observed pricing equilibrium; coalition values themselves require estimation. C5/C7/C8.

Read the source →

R09 · G2 — The Answer Economy — 15 April 2026. · 15 April 2026

Design: Primary report.

Finding: Primary report. Commercial survey; methods/findings inspected. March online sample of 1,076 B2B decision makers and 39 marketer interviews across regions. Reports AI research alongside continued Google use and frequent inaccuracies.

Boundary: reported frequency and AI-attributed switching are not observed journey shares or causal switching. Sampling/weighting and question bases constrain universal adoption estimates. C1/C2/C3/C8.

Read the source →

R10 · G2 — The Evaluation Maze — 22 July 2026. · 22 July 2026

Design: Primary report.

Finding: Primary report. June online survey of 1,038 decision makers plus 55 interviews. Forty percent name evaluation their longest stage; 39% name security the largest post-selection delay; 49% report a CFO veto in the past year.

Boundary: 40% is a respondent share, not elapsed journey time. Nine percent say they would permit guarded purchasing and 2% purchasing without pre-approval; those are authority preferences, not observed adoption. Current/planned use remains separate. Sponsor/self-report selection prevents causal conclusions. C2/C4/C6/C8.

Read the source →

R11 · Oded Netzer and Rajan Sambandam — Synthetic Data in Marketing Research How to Evaluate and When to Trust — 12 September, revised 19 September 2026. · 12 September, revised 19 September 2026

Design: Record and PDF.

Finding: Record and PDF. Full methods inspected. Real 2025 US sample of 3,063 people, 18 sectors, 108 questions; grounded leave-one-question-out prediction and out-of-bag screening. Rich category histories improve agreement over demographics alone; 46 retained questions perform better than all questions.

Boundary: selected in-domain survey predictions, not novel procurement decisions. Question-level agreement differs from individual fidelity; generated replicas add no independent people. Commercial research affiliation disclosed. C1/C2/C6/C8.

Read the source →

R12 · Kuric, Demcak and Krajcovic — Distorted Perspectives of LLM Simulated Preferences — 18 May 2026. · 18 May 2026

Design: Full preprint.

Finding: Full preprint. Twenty-nine actual UX studies, 2,073 people, 78 tasks. GPT-4.1's top choice matches people in 53% of tasks; GPT-5.2 reaches 65% without an established general distribution-fit gain.

Boundary: specific models/protocols and proprietary UX data, no procurement or transactions. Persona elaboration and reasoning do not consistently repair validity. C1/C2/C8.

Read the source →

R13 · Leonard Kinzinger and Jochen Hartmann — Synthetic Personalities — 3 June 2026. · 3 June 2026

Design: Full preprint.

Finding: Full preprint. Five hundred SOEP individuals, 183 held-out items, three models and 2.1 million generated responses. Grounding and preserving original responses help; reasoning improves ordering more reliably than closeness.

Boundary: best .788 normalized closeness is not 78.8% exact choice accuracy. Familiar socioeconomic items are not novel enterprise decisions; 500 real people remain the independent basis. C1/C2/C6/C8.

Read the source →

R14 · Fan Huang and colleagues — Evaluating LLM Generated Preference Distributions — 1 October 2026. · 1 October 2026

Design: Full preprint.

Finding: Full preprint. Nine open-weight models, airport/Yelp/Amazon domains, repeated elicitation with context/persona/order/temperature variation. Strong within-model stability coexists with cross-model disagreement.

Boundary: no human ground truth, unweighted personas, English/open models only. Tests internal consistency, not real demand. C2/C6/C8.

Read the source →

R15 · Stephen France and Pia Albinsson — Synthetic Consumer Insight Generation with Large Language Models — 7 July 2026. · 7 July 2026

Design: Record and full PDF.

Finding: Record and full PDF. Full comparative methods inspected. Word associations and completions about five US cities, 173 students, six LLMs and varied prompting/seeding. Broad topics overlap while synthetic style differs.

Boundary: semantic associations, students and cities; no willingness-to-pay or buying behavior. Supports hypothesis expansion rather than calibrated demand. C1/C2/C8.

Read the source →

R16 · Shuxian Fan and colleagues — ShoppingReasoningBench — 10 June 2026. · 10 June 2026

Design: Full preprint.

Finding: Full preprint. Seven experts construct 525 missions, 1,996 turns and 10,863 weighted binary rubrics across five product families; nine models tested. Optional constraints and multi-turn tasks are harder.

Boundary: scored recommendations, not actual purchase welfare or commercial effects; one generation per query and excluded categories. Versions and time-sensitive facts matter. C2/C4/C8.

Read the source →

R17 · Haobo Zhang and colleagues — Learning from Online User Feedback for Shopping Agents — 12 August 2026. · 12 August 2026

Design: Full preprint.

Finding: Full preprint. JD.com real-log offline study: 7,001 behavioral sessions from 6,704 users plus 6,121 directive sessions; purchase-aligned and corrective training.

Boundary: behavioral training selects successful purchases; issue-resolution evaluation partly uses an LLM judge. Better offline purchased-item retrieval does not establish live conversion, welfare or retention uplift. C4/C6/C8.

Read the source →

R18 · Jennifer Zou — Platform Choice Trust and Privacy in the Consumer AI Assistant Market — 16 July 2026. · 16 July 2026

Design: Full preprint.

Finding: Full preprint. June Prolific survey of 1,999 AI users with six hypothetical conjoint choices; task routing and platform preferences vary.

Boundary: weighting targets AI users and reduces effective sample size; privacy WTP is specification-dependent and excludes an outside option. Selected usage/trust associations are not causal. Commercial AI-visibility affiliation. C1/C2/C3/C8.

Read the source →

R19 · Carbon — B2B Buyers Report — September 2026. · September 2026

Design: Question-level findings.

Finding: Question-level findings. Commercial survey, 189 quality-filtered decision makers; denominators vary by question. High reported satisfaction coexists with provider reconsideration.

Boundary: small heterogeneous sample, limited probability-sampling detail, no longitudinal identification. Different bases cannot be combined into a respondent-level causal mechanism. Exact release day not verified. C2/C4.

Read the source →

R20 · NewtonX and Pretzl — Buyer Group Index — 30 June 2026. · 30 June 2026

Design: Primary methods announcement.

Finding: Primary methods announcement. Two hundred senior GTM leaders in 500-plus-employee organizations, approximately 30 AI-moderated interviews; 28 practices modeled against five outcomes.

Boundary: selected executive reporting, many candidate predictors, common-source outcomes and endogenous practices. Performance associations do not identify effects of adopting the index. No independent raw-data replication inspected. C2/C6.

Read the source →

R21 · Dong, Luo and Xu — From Product Search to Preference Articulation — 9 August 2026. · 9 August 2026

Design: Full preprint.

Finding: Full preprint. Theory of delegated search, profiling noise, preference complexity and attention. Distinguishes purchase incentives from consumer welfare.

Boundary: conditional comparative statics, no fitted effect sizes or buying-group experiment. Useful mechanism hypotheses for representation and preference clarification. C2/C3/C8.

Read the source →

R22 · Xiaolin Li, Qing Li, Danny Wang and Flora Gu — AI Human Collaboration in Sequential Sales Funnels — 13 June 2026. · 13 June 2026

Design: Primary conference abstract.

Finding: Primary conference abstract. Abstract describes randomized SME lending study with 78,169 prospects: AI initial calls and intent tags, then human follow-up.

Boundary: full arms, effect sizes, costs, defaults and durable contribution unavailable. Applications are not profit; other versions' numbers must not be merged. Promising field design, abstract-level appraisal. C4/C5/C6/C8.

Read the source →

R23 · Chatrath, Zhu, Fan and colleagues — READY or Not Reliable Enterprise Agent Deployment — 2 September 2026. · 2 September 2026

Design: Full preprint.

Finding: Full preprint. Clinical benchmark: 16 systems, 750 cases split equally into development and qualification. Freeze policies and qualify with a one-sided reliability bound.

Boundary: human success assumed at 90%, with uncertainty only on autonomous outcomes; qualification can have zero autonomous coverage when all cases go to assumed human review. Outcome versus process scoring matters. Framework transfers more plausibly than clinical numbers. C6/C8.

Read the source →

R24 · Ning, Li, Kong and Li — ControlScope — 28 September 2026. · 28 September 2026

Design: Full preprint.

Finding: Full preprint. Matched execution states compare automated keep/argument-edit/workflow-edit permissions on 20 filesystem tasks, ALFWorld and 585 AppWorld instances.

Boundary: LLM revision-reviewers, not direct human oversight evidence. Review can interrupt viable repair; effects vary and AppWorld net intervals include zero. Full AppWorld review renews the request allowance, introducing extra capacity. Supports scope/cadence tests, not universal greater autonomy. C6/C8.

Read the source →

R25 · Steve Drew and Jiayu Zhou — LEGIT — 18 September 2026. · 18 September 2026

Design: Full preprint.

Finding: Full preprint. Credentialing proposal and repeated paired GAIA/math tests, including 127 GAIA tasks, pinned configurations and budgets.

Boundary: corrected harness-quality evidence is inconclusive although costs differ; marketplace and feedback authentication unvalidated end to end. Cost per solved task includes failures; expiry is no assurance against drift. C0/C5/C8.

Read the source →

R26 · Paul-Peter Arslan — Audit Without Verification — 7 September 2026. · 7 September 2026

Design: Full preprint.

Finding: Full preprint. Balanced injected-defect/clean synthetic six-link pipelines. Preregistered main framing hypothesis is null; exploratory deletions and replications show origin suggestions can bias auditors.

Boundary: removal can help when suggestions are wrong and hurt when correct. Restricted evidence interfaces and preregistration deviations matter; no production error prevalence. C6/C8.

Read the source →

R27 · Yuhang Wang — Why LLM Agents Collapse Without Oversight The Enforcement Gap — 14 September, revised 23 September 2026. · 14 September, revised 23 September 2026

Design: Full v3.

Finding: Full v3. Tested agent setups distinguish advisory flags from binding control flow.

Boundary: use the detection/enforcement mechanism narrowly. Broad claims about every deployed framework are not established; the controller appendix uses synthetic critique stubs. No estimated GTM breach rate. C6/C8.

Read the source →

R28 · Microsoft Research, Bansal, Mirza and colleagues — Red Teaming a Network of Agents — 30 April 2026. · 30 April 2026

Design: Primary report.

Finding: Primary report. Internal network of more than 100 agents; malicious-message propagation, borrowed reputation and provenance loss.

Boundary: no representative exposure denominator or commercial effect. Supports shared-failure, containment and untrusted-content tests rather than a numeric error prior. C6/C7/C8.

Read the source →

R29 · Anthropic — Alignment Assessment of Recent Cybersecurity Incidents — 9 September 2026. · 9 September 2026

Design: Primary assessment.

Finding: Primary assessment. Four unauthorized-access incidents studied through transcripts, resampling and interpretability.

Boundary: roughly 481 million heterogeneous scanned transcripts are not a common eligible denominator. Disabled safeguards in some evaluations and partial-trajectory/LLM-grader limitations constrain ordinary-deployment transfer. Authority validity must be checked separately from output correctness. C6/C8.

Read the source →

R30 · Joshua Gans and Richard Holden — When Does Randomized Oversight Align AI Agents That Can Conceal — 29 September 2026. · 29 September 2026

Design: Full preprint.

Finding: Full preprint. Principal-agent theory: deterrence depends on preserved evidence, audit draws not known before action and meaningful incentives.

Boundary: concealment can make residual violations harder to detect; that does not establish lower total safety from more audit. Correct beliefs and mechanism assumptions remain open empirically. C6/C8.

Read the source →

R31 · Yan, Huo, Shang, Peng and Wu — CEO Arena — 28 September 2026. · 28 September 2026

Design: Full preprint.

Finding: Full preprint. Eight-firm, 500-day stylized competitive environment; matched replacements, 27 main and 26 robustness runs. Private gain can coexist with aggregate market loss.

Boundary: few seeds, programmed demand, symmetric settings and missing finance/labor/regulation. No reliable GTM equilibrium or welfare coefficient follows. C7/C8.

Read the source →

R32 · Jaewook Lee and Tijana Zrnic — Partially Performative Prediction — 5 June 2026. · 5 June 2026

Design: Full preprint.

Finding: Full preprint. Theory plus imposed-response credit case study distinguishes endogenous policy response from external drift.

Boundary: guarantees require smoothness, convexity, bounded sensitivity and comparator-path assumptions; no randomized B2B adaptation. C6/C7.

Read the source →

R33 · Anagnostides, Chauhan, Panageas, Sandholm and Yan — Computational Complexity of Performative Prediction — ICML July 2026. · ICML July 2026

Design: Peer-reviewed proceedings.

Finding: Peer-reviewed proceedings. Complexity of stable points under strong feedback and strategic local optima.

Boundary: worst-case hardness is not a finding that observed GTM is chaotic or impossible to improve. Stable and optimal are different targets. C7.

Read the source →

R34 · Vo, Reddy, Rodemann, Chau and Muandet — Off Policy Evaluation with Strategic Agents via Local Disclosure — 5 June, ICML July 2026. · 5 June, ICML July 2026

Design: Full preprint, proceedings.

Finding: Full preprint, proceedings. Local disclosure exposes pre-adaptation features; theory and synthetic/German-credit tests impose strategic response.

Boundary: response-distribution, support and nuisance assumptions. Ordinary matching of policy-changed attributes cannot recover pre-action demand. C6/C7.

Read the source →

R35 · Wei, Qu and Miao — Missingness Aware Off Policy Evaluation — 18 June, ICML July 2026. · 18 June, ICML July 2026

Design: Full paper.

Finding: Full paper. Reward-dependent missingness addressed using next-state shadow variables and bridge assumptions; simulation and introduced sepsis missingness.

Boundary: relevance, exclusion, completeness, positivity, realizability and policy support are demanding. CRM proxies are not automatically valid shadow variables. Report bounds or experiment when identification fails. C4/C6.

Read the source →

R36 · Spear, Komorowski, Pope and Moodie — Kernel Weighted Importance Sampling — 16 July, revised 4 August 2026. · 16 July, revised 4 August 2026

Design: Record and abstract.

Finding: Record and abstract. Abstract-level appraisal of a robust contextual-bandit estimator.

Boundary: full methods not reviewed; estimator robustness does not remedy zero action probability, hidden confounding, strategic shift or missing outcomes. Candidate only after identification/support audit. C6.

Read the source →

R37 · Anthropic Economics — Project Swap — 24 September 2026. · 24 September 2026

Design: Primary study.

Finding: Primary study. Human book preferences, five-minute intake, decentralized bargaining and reruns; 188 rankings and 205 market runs.

Boundary: specific representation-versus-allocation decomposition, employee sample, unincentivized rankings, attrition and well-behaved agents. Some agreed physical transfers were unfulfilled; no B2B transfer of its shortfall percentage. C2/C4/C8.

Read the source →

R38 · He, Agarwal, Denisov-Blanch, Azaletskiy, Koyejo and Vasilescu — AI Writes Faster Than Humans Can Review — 2 July 2026. · 2 July 2026

Design: Full preprint.

Finding: Full preprint. One B2B software company, 802 developers, 196,212 pull requests, January 2024–April 2026; staggered adoption comparisons. Output and reviewer load rose substantially.

Boundary: self-selected adoption/intensity, unusually favorable budgets, co-moving process changes, coarse merge/revert quality proxies. Exact causal magnitude is not identified. Measure workload relocation, substantive review, queues and downstream quality. C6/C8.

Read the source →

R39 · Anthis, Cha, Barocas, Chouldechova and Hofman — Effects of Generative AI Errors on User Reliance Across Task Difficulty — 5 April 2026, CHI EA. · 5 April 2026, CHI EA

Design: Full paper.

Finding: Full paper. Preregistered randomized 3×2 diagram experiment, 577 US Prolific participants, incentive-compatible reliance bids. More errors reduce bids; easy-task errors do not establish disproportionate behavioral distrust.

Boundary: conspicuous errors and controlled feedback, not employee review or actual commercial authority. Reported trust and demonstrated reliance differ. C2/C6/C8.

Read the source →

R40 · Liu, Christian, Dumbalska, Bakker and Dubey — AI Assistance Reduces Persistence and Hurts Independent Performance — 6 April, revised 5 August 2026. · 6 April, revised 5 August 2026

Design: Full v4.

Finding: Full v4. Three short experiments, 1,222 recruited and 1,060 after exclusions; assistance removed before arithmetic/reading tests. Unaided performance declines; the improved arithmetic replication's persistence difference is nonsignificant.

Boundary: immediate outcomes without performance-contingent payment, initial differential exclusions, self-selected hint/answer use. No durable professional deskilling coefficient. C6/C8.

Read the source →

R41 · Wei Cai, Andrea Prat and Jiehang Yu — Measuring Organizational Capital — journal publication 8 June 2026. · journal publication 8 June 2026

Design: Journal abstract, author explanation, chronology.

Finding: Journal abstract, author explanation, chronology. More than a million employee reviews and text-based organizational measures associated with firm outcomes. **Access:** abstract and primary explanation, not full current methods. **Freshness:** originated 2022, revised 2025; new publication of older work. Proxies do not identify causal alignment or AI-created capital. C6/C7.

Boundary: Abstract and primary author explanation inspected, not full current methods. Originated 2022 and revised 2025; June 2026 is journal publication of older work. Employee-review proxies and firm associations do not identify causal alignment or AI-created organizational capital.

Read the source →

R42 · McKinley, Brau, Aloysius and Hofer — Is AI an Algorithm by Any Other Name — 13 April 2026. · 13 April 2026

Design: Full journal paper.

Finding: Full journal paper. Two lab forecasting studies plus 2019–20 staggered retail deployment, 575,354 observations and 11,652 SKU-store clusters.

Boundary: identical algorithm labels in labs differ from actual algorithm changes in the field. Context and performance affect reliance/adjustment; this is predictive forecasting, not conversational-agent reviewer accuracy or a fitted fatigue curve. C6/C8.

Read the source →

R43 · Dell'Acqua and collaborators — The Cybernetic Teammate — June 2026 journal publication. · June 2026 journal publication

Design: Primary publication record.

Finding: Primary publication record. Abstract/record inspected. Randomized work with 791 P&G professionals; AI changes individual/team idea work and cross-functional expertise. **Freshness:** working paper originated March 2025; journal publication is recent, underlying study is older.

Boundary: product-innovation task and evaluated ideas, not company profit, sales queues or universal human–AI synergy. C6/C8.

Read the source →

R44 · Acemoglu, Tianyi Lin, Ozdaglar and Siderius — How AI Aggregation Affects Knowledge — arXiv 6 April 2026, document dated 24 August. · arXiv 6 April 2026, document dated 24 August

Design: Full theory.

Finding: Full theory. DeGroot network with independent initial signals, no later external information, endogenous aggregator feedback and specified connectivity. Fast feedback can undermine robust learning; local topic aggregation has advantages under these assumptions. **Freshness:** earlier March circulation exists.

Boundary: no empirical universal modular-architecture result. Supports testing source independence and fresh evidence. C0/C6/C7.

Read the source →

R45 · Alex Chan — AI and the Information Supply of the Web — June 2026 working-paper discussion. · June 2026 working-paper discussion

Design: Primary HBS explanation.

Finding: Primary HBS explanation. Author/institute explanation inspected; full paper not independently retrieved. Mechanism links traffic diversion, source financing and repeated information production.

Boundary: theory/explanation, not observed inevitable web collapse. More generated pages or citations do not demonstrate fresh independent information. C3/C7.

Read the source →

R46 · Dorner, Tucker, Zhang and Heuber — Green Consumer Information in a Digital Product Passport — 19 August 2026. · 19 August 2026

Design: Full journal paper.

Finding: Full journal paper. Preregistered incentivized toothbrush study, 131 participants, conducted 31 August–7 September 2023. Information-plus-education increases WTP by approximately NZ$0.69; information alone is null in parametric difference-in-differences.

Boundary: small university pool, low-cost product, favorable information, selected engagement and no positive environmental spillover. Uptake, comprehension and consequential choice differ. C1/C2.

Read the source →

R47 · Pattyn, Dillen and Goetz — Validate Before You Build — 18 August 2026. · 18 August 2026

Design: Publisher article.

Finding: Publisher article. Partial indexed full text inspected; direct publisher access limited. Twenty-nine founder interviews, 16 software startups; exploratory commitment levels and composite performance.

Boundary: retrospective selected cases, endogenous validation, unverified interview dates, coding/composite dependence. Customer payment, investor money, sustained use and viability are different evidence. No causal PMF threshold. C1/C7.

Read the source →

R48 · Chan, Lo, Jung and Wu — Emerging Together How Brand Collaborations Give Rise to Proto Categories — 17 July 2026. · 17 July 2026

Design: Primary proceedings listing.

Finding: Primary proceedings listing. Abstract-level category-formation theory linking identity distance and collaboration.

Boundary: no empirical sample or calibrated effect inspected. Treat proposed inverted-U as a hypothesis. Category recognition requires repeated audience behavior, not only a seller's positioning statement. C7.

Read the source →

R49 · Baumann, Cullen, Faia, Ferrando, Perez-Truglia and Rariga — Innovation Without Borders — June 2026. · June 2026

Design: Full author-hosted paper.

Finding: Full author-hosted paper. Information randomization in Q4 2025 ECB SAFE survey; 3,316 complete firms across 12 countries after substantial incomplete-response exclusions. Treatment raises expected AI investment share by 1.8 percentage points; domestic peer beliefs matter more than foreign ones.

Boundary: immediate investment expectations, not realized spend, adoption, productivity or cash; complete-case selection. C7.

Read the source →

R50 · Bottesch, Schwenke, Zimmermann, Förster and Klier — Faster Higher Stronger — 28 July 2026. · 28 July 2026

Design: Paper and PDF.

Finding: Paper and PDF. Full PDF inspected. Randomized document-RAG versus manual access in one industrial firm, November 2024–March 2025; 128 complete participants, 78/50 arms, expert output evaluation. Faster acquisition, summarization and creation, but less accurate acquisition.

Boundary: short fixed-sequence tasks and exclusions; no payroll/profit endpoint, no significant creation-novelty gain. C5/C6/C8.

Read the source →

R51 · Babina, He and Jiang — Canaries in the Gold Mine — August 2026, SSRN posted 25 August. · August 2026, SSRN posted 25 August

Design: NBER record, SSRN, measurement documentation.

Finding: NBER record, SSRN, measurement documentation. Abstract/author/documentation appraisal only; long full manuscript inaccessible. AI-skilled employment and firm panels through 2024 are associated with recent productivity gains and organization-building roles.

Boundary: identification not deeply audited; updated firm panel unreleased. Mechanism candidate for accumulating workflow capital, not a transferable productivity coefficient. C5/C7.

Read the source →

R52 · Niklas Heusch — Synthetic Benchmark Dataset with Endogenous Marketing Spend — 21 August 2026. · 21 August 2026

Design: Full preprint.

Finding: Full preprint. A 156-week, three-channel fictional retail generator links spend to quarterly performance, planned promotions, bursts and algorithmic chasing; provides known effects and noisy practitioner observations.

Boundary: no real company data, additive channels, fixed parameters, no rival/supply responses; bidding uses pre-media sales and geo groups share a national series. Useful adversarial estimator tests, not calibrated GTM response or proof of real-data identification. C3/C5/C6.

Read the source →

R53 · Masahiro Kato, Daiki Honma and Taka Kato — Generative Marketing Mix Modeling — 10 September 2026. · 10 September 2026

Design: Full preprint.

Finding: Full preprint. Repeated answer collection on 56 English/Japanese questions supplies occurrence estimates; business responses and source changes are simulated. Requires query counts, platform shares, notice probabilities, rank/support, exchangeability and a specified interference process.

Boundary: no randomized commercial GEO uplift; main-effect priors are nonnegative. Measurement and simulated response precision do not establish the empirical effect of citations or source edits. C2/C3/C5.

Read the source →

R54 · Uberti-Bona Marin and colleagues — Auditing AI Generated Product Recommendations — 16 September 2026. · 16 September 2026

Design: Full preprint.

Finding: Full preprint. ConsumerQ contains 2,528 real advice queries; the main audit uses 117 physical-product queries repeated three times, producing 1,536 responses across interface/API conditions.

Boundary: nonrepresentative source users, one collection location/time, only recovered AI Overviews, three repetitions, configuration differences. Business products and subscriptions are excluded from the core audit; displayed sources do not establish causal influence. C2/C3/C8.

Read the source →

R55 · Farrell, Korganbekova and Misra — Robust A/B Decisions — 7 September 2026; manuscript 5 September. · 7 September 2026; manuscript 5 September

Design: Full preprint.

Finding: Full preprint. Primary methods/results inspected. Uses 552 randomized advertising experiments from one anonymous US platform, with undisclosed study dates, to retrospectively compare deployment rules under random and geographic splits.

Boundary: held-out deployment evaluation, not randomized rollout of decision rules. Empirical success performs best under random splits; ambiguity preferences have outcome-specific units. The oracle itself uses noisy held-out means. No transferable 25% revenue gain follows. C5/C6/C7.

Read the source →

R56 · Zihao Chen — When Do Surrogate Metrics Work — 24 September 2026. · 24 September 2026

Design: Full preprint.

Finding: Full preprint. Primary setup, methods and results inspected; proofs not independently audited. Eleven simulations and older Hillstrom, Criteo and LaLonde experimental reanalyses expose coverage failures that a surrogate disagreement diagnostic often misses.

Boundary: prediction-powered alternatives require random labeling, enough observed events and justified enrollment/cohort stability; sparse outcomes and drift remain problematic. A non-rejected diagnostic supplies no certificate. Not a new SaaS LTV estimate or permission for unrestricted repeated testing. C4/C5/C6.

Read the source →

R57 · Tobias Konitzer — Media Measurement and the Assisted Own Goal — 10 July 2026; manuscript 21 June. · 10 July 2026; manuscript 21 June

Design: Full preprint.

Finding: Full preprint. Primary analytical model, simulations and identification conditions inspected. GrowthLoop-affiliated author analyzes demand generated on one property but booked elsewhere and proposes audience randomization.

Boundary: commercially interested theoretical/simulation evidence, no field bias magnitude. Identification expressly needs channel-complete outcomes; randomization does not recover missing transactions. Outcome coverage, linkage and interference remain requirements. C3/C5/C6.

Read the source →

R58 · Wang, Zhu, Feng, Lu and Jia — Agentic AI and Human-in-the-Loop Interventions — 14 May, revised 1 June 2026. · 14 May, revised 1 June 2026

Design: Full v2 PDF.

Finding: Full v2 PDF. Primary methods/results and matching inspected. Taobao worker-randomized deployment with 647 workers and 680,676 chats, 15–31 August 2024 plus 14-day baseline; 39,432 chats AI-eligible. Faster service accompanies lower eligible-chat ratings; only 17% of chats receive ratings.

Boundary: short customer-service study, no acquisition/profit endpoint. Escalation type and timing are selected, with matched comparisons rather than randomized mechanisms. No universal causal handoff penalty. C4/C5/C6/C8.

Read the source →

R59 · Han, Yan and Zhang — More Rejective Not More Discriminative — 25 August 2026. · 25 August 2026

Design: Full preprint.

Finding: Full preprint. Primary methods/results and limitations inspected. Preregistered LLM-monitor instrument varies five review lengths and six verifiers in retail/AppWorld with injected errors and clean controls; retail's 200 bad items share 31 clean anchors. Longer units generally raise both valid and invalid rejection.

Boundary: nonadaptive errors, short windows, changed intermediate observations and token-only costs; no human fatigue experiment or universal review cadence. Catch rate can be identified with labels while remaining an insufficient objective. C6/C8.

Read the source →

R60 · Liang Chen — Simulated Customers Never Walk Away — 16 June 2026. · 16 June 2026

Design: Full preprint.

Finding: Full preprint. Primary methods/results inspected. Chinese parent-mediated matchmaking sales logs contain 2,790 conversations and 793 verified payments; paired evaluation uses 374 conversations and 1,109 probes. Simulated nonbuyers overengage, and disengagement instructions trade one group's error against another.

Boundary: observational next-turn fidelity given real history, not free-running conversion or causal sales-tactic effectiveness. LLM-assigned states lack independent human validation; outcome-balanced sampling is not a natural conversion base. Study dates unspecified. C1/C2/C8.

Read the source →

R61 · Kreimer, Stoppacher, Hauser, Foscht and Eisingerich — Pathways to Premium — 21 September 2026. · 21 September 2026

Design: Full journal paper.

Finding: Full journal paper. Four observational field studies at one regional European newspaper, samples 522, 566, 358 and 2,434; calendar dates unspecified. Earlier studies measure intentions/offer acceptance; the fourth observes retention after a six-week €0.99 trial automatically becomes €12.99 monthly.

Boundary: no randomized habit treatment; Study 4 lacks demographic/usage controls, engagement imperfectly measures habit, and trial-hopping explanation is untested. Phase-specific hypotheses, not a transferable optimal habit threshold. C4/C5.

Read the source →

R62 · Kyungran Lee, ShinWoo Lee and Gil Sadka — ASC 606 Revenue Uncertainty and Cost of Debt — journal 24 April 2026. · journal 24 April 2026

Design: Full journal paper, earlier working-paper record.

Finding: Full journal paper, earlier working-paper record. New publication of work first posted 18 September 2024 and revised 23 January 2026. Difference-in-differences examines materially affected public firms around accounting adoption, using 2016–19 and 2014–21 windows.

Boundary: exposure, coincident industry shocks and selected analyst/loan coverage constrain interpretation. Accounting-information uncertainty is not GTM demand uncertainty; financing results are not a startup cash-timing coefficient. C5.

Read the source →

R63 · Chung and Chuwonganant — Effects of the New Revenue Recognition Standard — journal 23 August 2026. · journal 23 August 2026

Design: Publisher abstract, working-paper history.

Finding: Publisher abstract, working-paper history. Abstract/metadata inspected; full text unavailable. First circulated 27 April 2020, revised 9 July 2026. Reports improved analyst forecast accuracy/information using public-firm data.

Boundary: match samples, exposure definitions, horizons and estimands before declaring a contradiction with Lee and colleagues. This is an abstract-only counterpoint against a uniform accounting-information claim, not directly observed cash or GTM causality. C5/C7.

Read the source →

R64 · Stephanie T. Wang, Jeffrey Gleason, Yakov Bart, Christo Wilson and Danaé Metaxa — AI in Search Reduces Publisher Referrals — 18 August 2026. · 18 August 2026

Design: Full preprint.

Finding: Full preprint. Preregistered browser-extension randomized field experiment; March 17–19 recruitment, three-day baseline and seven-day treatment. Of 1,444 enrolled, 1,100 search during treatment and 956 complete the survey. Forced AI Mode lowers outbound click-through by 18.8 percentage points, CI −22.2 to −15.3.

Boundary: selected short-horizon panel and forced adoption from 0.6% baseline; no-AI estimates are LATE under approximately 51% compliance/exclusion. No B2B shortlist or revenue effect. Linked OSF registration could not be independently retrieved, so its registration date is unverified. C2/C3/C7.

Read the source →

R65 · Peibo Zhang, Ruomeng Cui and Dennis J. Zhang — The Impact of AI Search on the Online Content Ecosystem — 14 May, revised 20 June 2026. · 14 May, revised 20 June 2026

Design: Full v3.

Finding: Full v3. Difference-in-differences compares AI-Overview-eligible SFW with NSFW Reddit communities: 105,012 subreddits over 578 days, January 2024–July 2025; December 2025 extension examines AI Mode.

Boundary: engagement is not referral traffic or commercial conversion. Differential shocks, partial US exposure in the designated preperiod and an author-count pretrend constrain identification; content heterogeneity uses model classifications. Useful destination-complementarity evidence, not universally established parallel trends or a proven conservative lower bound. C3/C7.

Read the source →

Studies needed to qualify the next decisions

P01 · Agent-assisted B2B execution · proposed_unrun

Design: Randomize eligible accounts/buying groups to current policy, qualified assistant policy, and where feasible no-contact control

Outcome: Incremental contribution per assigned account over a declared horizon

Identification requirements: Member contamination, hidden selection, immature outcomes and changing eligibility

P02 · Human and agent handoff · proposed_unrun

Design: Compare complete triage/follow-up policies under the same capacity envelope

Outcome: Accepted deals or contribution through final customer quality

Identification requirements: Conditioning only on contacted/high-intent accounts

P03 · Buyer AI mediation · proposed_unrun

Design: Observe journeys; randomize permitted representation/evidence improvements at account or interface level

Outcome: Correct representation and qualified consideration; later contribution separately

Identification requirements: Platform changes, unobserved exposure, attribution of existing preference

P04 · Synthetic buyer validation · proposed_unrun

Design: Held-out real respondents and choices; compare no grounding, demographics, rich histories and wrong-person grounding

Outcome: Predeclared distribution/ranking/level error and decision regret

Identification requirements: Treating generated replicas as independent people; leakage from known questions

P05 · Agent qualification · proposed_unrun

Design: Develop policies then freeze and test on representative held-out workflows

Outcome: Independently accepted terminal/process success with uncertainty

Identification requirements: Reusing qualification cases; judge leakage; drift after qualification

P06 · Channel response and carryover · proposed_unrun

Design: Randomized supported spend levels or well-designed geo/account interventions with cooldown

Outcome: Incremental discounted contribution in tested support

Identification requirements: Demand-following budget pauses, weak controls and finite-geo imbalance

P07 · Pricing and packaging · proposed_unrun

Design: Randomize eligible segments/accounts or appropriate clusters to feasible packages/metrics

Outcome: Contribution and cash over acquisition plus renewal horizon

Identification requirements: Current-price comparisons ignore history; shared-capacity interference

P08 · Retention and value · proposed_unrun

Design: Randomize retention plays among eligible accounts with untreated comparison

Outcome: Incremental retained contribution rather than save count

Identification requirements: Targeting risk rather than responsiveness; survival conditioning

P09 · Review and organization · proposed_unrun

Design: Vary cadence/scope under measured workload; preserve hard authority rules

Outcome: Valid accepted throughput per available review hour

Identification requirements: Easier-case selection, workload redistribution, harmful deskilling conclusions from voluntary use

P10 · Market and category feedback · proposed_unrun

Design: Cluster/switchback tests only where spillover/lag justify them; monitor rivals and platforms

Outcome: Firm contribution plus separately defined ecosystem outcomes

Identification requirements: Individual tests assume no interference; changepoints misidentified as external shocks

P11 · Funnel resource response · proposed_unrun

Design: Randomize feasible resource changes while freezing stage definitions and eligibility

Outcome: Incremental contribution or qualified throughput per incremental resource

Identification requirements: Administrative threshold changes mistaken for productivity; carryover and shared-capacity interference

P12 · Proxy and outcome coverage · proposed_unrun

Design: Backtest proxies on mature holdouts and audit booking/linkage across direct, partner and marketplace routes

Outcome: Held-out decision agreement, effect direction, interval coverage and missing-outcome sensitivity

Identification requirements: Validation lacks outcome support; correlation substituted for causal-effect agreement

P13 · Deployment-rule validation · proposed_unrun

Design: Reevaluate historical experiments with current, empirical-success and explicit economic/downside rules; later prospective policy test where feasible

Outcome: Declared deployment payoff/regret and downside exposure

Identification requirements: Noisy ex-post oracle, selective experiment archives and transport disguised as randomized rollout

The tiers

GTM runs at different tempos. The model separates them so a fast loop never overwrites a slow, human-owned decision.

Tier	Tempo	What it decides
Pre-funnel buyer state and brand memory	slow (brand: quarters-years) to medium (buyer-state: weeks-months); mostly INVISIBLE to the seller	Prior need, experience, shortlist, roles, proof and constraints influence consideration and contact selection. Survey in-market/shortlist shares motivate measurement without identifying the share of causal value created before contact.
Strategy / Configuration tier (slow-moving, human-owned)	quarters to years	WHO, WHAT, HOW and WHERE to sell under product, buyer and economic constraints. These are explicit policy choices rather than universally forced ACV moves.
Economics: ledgers, defined metrics and model-dependent forecasts	continuous; reviewed monthly/quarterly	Contribution, cash viability and feasible allocation on declared business-kind/cohort/period bases. Accounting equalities and defined ratios can be exact within conventions; predictive and causal interpretations require assumptions and evidence.
Execution / Behavioral tier (fast, increasingly agent-owned)	per-lead, per-account, daily/weekly	The actual per-lead and per-account loop: detect signal -> decide whether to act -> gather context -> qualify -> draft/personalize -> route -> hand off -> follow up -> expand/retain. This is where the behavioral coefficients (conversion rates, response, reply rates) actually get produced, and where they DRIFT.
Operations & Systems substrate (the rails all tiers run on)	always-on
Capital context and declared strategic objectives	slow (shifts with the rate/capital cycle over quarters-years)	Capital conditions can constrain decisions and influence valuation. Operating objective weights are declared preferences, not causally determined by a fitted macro scalar in this model.

State variables

Symbol	Name	Type	What it is
C	Logo and recurring-value stocks (separate units)	stock	Customer counts and recurring monetary value are separate stocks; no shared expansion term for logos.
a	Acquisition inflow rate	flow_in	Acquisition inflow must state whether it counts logos or recurring currency per time.
delta	Churn rate	flow_out	Logo loss and value contraction/churn require separate denominators and consistent periods.
g	Expansion rate	flow_modifier	Expansion changes recurring value; net value retention is defined on a fixed beginning cohort.
V	Top-of-funnel volume	flow_source	Raw demand entering the funnel.
r_i	Stage conversion rates	coefficient	Conditional stage rates depend on case mix and stage predicates. Agent mean/variance improvements require evidence; neither is universal.
p_bar	Average price / ACV	coefficient	Price is one input to motion feasibility alongside contribution, cash, capacity and buyer requirements.
m	Gross margin	coefficient	Gross margin is a defined gross-profit share on its accounting basis. Contribution/net value may require additional variable delivery, service, return and failure costs; gross margin alone does not cover all of them.
S	Sales+marketing spend	control_input	Declare complete acquisition costs and denominator; agent costs may already be included.
x	Market penetration / saturation	state	Specify counts versus fractions; reciprocal CAC saturation is an unfit response scenario.
Phi	Product-market fit (latent)	latent_multiplier	Fit is multidimensional and latent; multiplicative response is a hypothesis, not its definition.
k(tau)	Lag/memory kernel	function	Distributed-lag kernel. Why last-touch attribution structurally misattributes brand/demand-creation.
Psi	Macro regime scalar	exogenous_index	Macro conditions and management objective preferences are distinct; weights are unfit scenarios.
B	Buyer-state vector	latent_state	Pre-action buyer roles, need, shortlist and constraints help prediction but do not themselves identify effects.
B_r	Brand stock	stock	Brand memory parameters are unknown for the local segment; linear stock and exponential lag are equivalent.
S	Switching-cost moat	latent_modifier	High GRR plus low satisfaction does not uniquely identify switching costs or a piecewise causal law.
k	PLG viral coefficient + loop state	loop_state	Separate referral generations, cumulative unique referrals, active users and paid customers; no universal B2B k<1 bound.
alpha	Autonomy level of an agentic execution chain	policy_summary	Descriptive fraction only. A complete policy conditions execute/review/repair/defer/stop on task, state, authority and capacity.
G	Governance / HITL maturity index	exogenous_index	Summary descriptive index. Detection, enforcement, correction, provenance, containment and recovery need separate measured predicates.
gamma	Effective error catch-rate of human review	function	Catch rate is criterion-specific; also measure false rejection, correction and reviewer-introduced defects.
p_step	Effective per-step error rate of an agent-executed chain	function	Residual shipped defects conditional on task, data, policy and configuration; no universal scalar error input.
R_chain	Chain reliability across n sequential steps	function	Use conditional chain rule; independent homogeneous power is a toy case. Accepted outcome may include repairs.
C_a	Agent execution spend	flow_out	All eligible attempt, retry, review, tool/data and maintenance cost with acquisition/delivery allocation stated.
lambda_hitl	HITL-as-data-collection rate	coefficient	Learning effectiveness must be observed through valid outcome linkage, implementation and held-out evaluation; unfit here.
theta	Category attention/congestion scenario	loop_state	Attention/congestion hypothesis, distinct from forecast error. Positive decay bounds response to bounded input; constant adoption additionally gives a steady state. Net externality sign remains unknown.
kappa	Aggregate agent-adoption share across the competitive category	exogenous_index	Category adoption is distinct from firm autonomy; interference and platform/rival response matter.
B_agent	Agent-evaluator presence within a buying committee	state	Role × task × phase matrix with platform/interface/version, fidelity, verification, authority and observed fulfillment; aggregate share optional.

The equations

Each carries an epistemic type and a plain-English reading. Use its assumptions, evidence boundaries and qualification gates to determine which decisions it supports.

Separate logo and recurring-value stock balances model_identity

\frac{d L}{d t} = a_{logo} - δ_{logo} L, \frac{d M}{d t} = a_{value} - (δ_{value} - g_{exp}) M

steady state

L^{*} = \frac{a _{logo}}{δ _{logo}}, M^{*} = \frac{a _{value}}{δ _{value} - g _{exp}} (δ_{value} > g_{exp})

Reads as: L counts active logos; M measures recurring monetary value. New-logo inflow has units logos/time; new-value inflow has units currency/time. Expansion changes value without adding logos.

Implication: With n=delta_value-g_expansion, M(T)=M0*exp(-n*T)+a_value*(1-exp(-n*T))/n for n!=0; at n=0, M(T)=M0+a_value*T. Near zero net attrition the long-run equilibrium is not a finite-horizon ROI comparison. Retention versus acquisition depends on feasible responses, costs, horizon and quality.

The funnel as a log-linear operator identity

R = V \overset{p}{ˉ} i \prod r_{i} ⟺ ln R = ln V + ln \overset{p}{ˉ} + i \sum ln r_{i}

Reads as: Cohort value on the declared booking/price basis equals eligible volume times nested stage conversion fractions times mean final-outcome value; logs are additive only for positive inputs. It is not an independent causal lever decomposition.

Implication: All rates have unit elasticity for equal proportional improvements, holding other inputs fixed. For equal absolute rate changes, dR/dr_i=R/r_i, but investment return also depends on dr_i/dspend, downstream quality/rates, price, capacity and time. A low rate can be a useful filter; select the best feasible incremental contribution intervention, not min(r_i).

The unit-economics master inequality estimator

L T V > C A C ⟺ \frac{m p ˉ}{δ} > \frac{S}{a} ⟺ \frac{m p ˉ a}{δ S} > 1

Reads as: Under constant contribution and homogeneous undiscounted hazard, LTV/CAC=m*p_bar*a/(delta*S). Its elasticity to churn is -1 and to acquisition +1; retention is not quadratic.

Implication: Evaluate discounted incremental contribution and cash viability over a declared horizon. Payback, runway, capacity, service promises and authority can bind simultaneously. Historical ratio benchmarks are heuristics rather than universal pass/fail thresholds.

Channel response and saturation scenarios hypothesis

C A C (x) = \frac{C A C _{0}}{1 - ( \frac{x}{T A M} ) ^{β}}

Reads as: This reciprocal form is one proposed finite-market response family, not a universal CAC curve. Here x is cumulative captured demand in the same units as TAM.

Implication: Estimate marginal incremental contribution in tested spend support. Distinguish saturation, changing demand, creative fatigue, delivery and rival reactions. Use channel experiments with time paths and spillover checks; do not extrapolate the asymptote as fact.

Evidence boundary: Requires 0<=x<TAM, CAC_0>0 and beta>0; convexity is not guaranteed over the full domain for every beta. Parameters are unfit here. Other response families and endogenous budgets can fit observational data similarly.

Motion feasibility under economics and buyer requirements hypothesis

E [Δ PV (contribution)] \geq Δ cost subject to feasibility gates

Reads as: ACV is one input to contribution and cost-to-serve. Renewal, implementation, margin, purchase complexity, preference and opportunity cost also matter.

Implication: PLG/hybrid/sales-led ACV bands are historical segment heuristics. Choose feasible motions by measured economics and requirements; hybrids and simultaneous gates are allowed.

Distributed lags and the attribution problem hypothesis

r (t) = \int k (τ) A (t - τ) d τ

Reads as: A convolution describes a proposed distributed response to earlier activity. A linear stock with decay has an equivalent exponential convolution solution.

Implication: Last-touch credit is not incremental contribution. Self-reported attribution can suggest discovery mechanisms but does not identify a lag kernel or causal brand effect by itself.

PMF interaction hypothesis hypothesis

R = Φ f (GTM) versus alternative response models

Reads as: Fit may modify GTM response, but multiplicative versus additive structure requires identification and out-of-sample validation.

Implication: Use independent pre-outcome fit measures, segment holdouts and feasible intervention comparisons. Low measured fit is a warning, not a proven universal zero-response threshold.

Sales throughput approximation estimator

Sales Velocity = \frac{n _{opps} \cdot win rate \cdot avg deal size}{sales cycle length}

Reads as: A stationary pipeline approximation combines consistently defined opportunities, win share, deal value and cycle time. Stock denominators and completed-cohort outcomes must match.

Implication: These are coupled quantities. A faster research/draft step does not necessarily shorten procurement or implementation. Measure complete throughput and queue delays rather than assuming four independent causal levers.

Recurring-value ledger reconciliation identity

M R R_{t} = M R R_{t - 1} + New_{t} + Exp_{t} - Contr_{t} - Churn_{t} + Other_{t}

Reads as: Ending MRR equals beginning MRR plus mutually exclusive movements on a fixed recurring-value convention; include FX, reactivation, scope changes or corrections as separate reconciliation terms when relevant.

Implication: This is a recurring-value accounting identity, not recognized GAAP revenue or cash. Other ratios can also be exact definitions on their declared basis; their causal meaning and forecasts are separate.

Corrected LTV under heterogeneous churn (Fader-Hardie sBG) estimator

LTV = t \geq 0 \sum \frac{E [ c _{t} 1 ( active _{t} )]}{( 1 + d ) ^{t}} - C_{nonrec}, S (t) = \frac{B ( α , β + t )}{B ( α , β )}

Reads as: Survival-weighted expected net contribution requires a stated payment schedule, censoring and cohort model. Heterogeneity can change aggregate hazards without any individual behavior improvement.

Implication: For alpha=2.5,beta=6, mean initial churn is 2.5/8.5=29.41%. Expected paid periods including t=0 are 5; future periods excluding t=0 are 4. A geometric model matched to initial churn gives 3.4 including t=0 or 2.4 future periods. This example does not support a universal LTV overstatement or an invented 10%-churn comparator.

Fit and switching-cost alternatives hypothesis

R = {Φ f (GTM) Φ f_{1} (GTM) + S f_{2} (installed base) S < S^{*} S \geq S^{*} h_{eff} = h_{0} e^{- S}

Reads as: Switching costs can affect renewal and rents. A sharp regime threshold, an exponential hazard and additive revenue form are proposed specifications, not identified consequences of lock-in theory.

Implication: GRR plus low satisfaction does not identify switching costs, and retained enterprise revenue is not an existence proof of an additive causal law. Compare specifications with independent fit/switching-cost measurements and interventions.

Evidence boundary: Classical switching-cost mechanisms motivate the hypothesis; this exact piecewise GTM revenue model has not been fitted or causally tested here.

Macro regime objective re-weighting (Psi / Rule-of-X) hypothesis

U = w_{g} (Ψ) growth + w_{f} (Ψ) FCF, X = M (Ψ) growth + FCF margin, M \approx 2 - 3

Reads as: Capital conditions can constrain decisions and influence valuation. Operating objective weights are declared preferences, not causally determined by a fitted macro scalar in this model.

Implication: Separate market multiples/rates, firm funding/runway/covenants and management objectives. Compare scenarios; no fixed rate-to-payback or rate-to-growth-weight mapping is identified.

Evidence boundary: Rule-of-X is a practitioner valuation heuristic. Units, scaling and firm mandate must be declared; causal operating policies do not follow from it.

Buyer mix and causal policy estimands conditional_coefficient

τ_{H} (x) = E [Y_{H} (a) - Y_{H} (a_{0}) ∣ X_{pre} = x] \neq = P (Y_{H} ∣ a, x)

Reads as: Pre-action shortlist, need, role, constraints and prior exposure can influence assignment and outcomes. Observed funnel rates mix these states. Conditioning on them helps prediction but does not itself identify a causal effect.

Implication: Compare feasible contact/evaluation/implementation policies, including no action when appropriate, on a stable eligible population. The 95-5 in-market heuristic and pre-contact winner-shortlist prevalence have different denominators and identify no 95% value-creation split or geometric headcount-decay law.

Evidence boundary: Retrospective commercial surveys motivate buyer-state measurement. Effects need randomization or justified exchangeability, positivity, consistency and interference handling; conditional effects are averages, not identifiable individual effects.

Brand memory: stock and distributed-lag representations hypothesis

\dot{B}_{brand} = b_{brand} I - d_{brand} B_{brand}, t_{1/2} = \frac{ln 2}{d _{brand}}

Reads as: A linear memory stock and exponential distributed lag are equivalent representations under the same assumptions. Nonlinear responses and independent awareness/shortlist observations may distinguish richer models.

Implication: Track validated awareness, entry points, consideration and business outcomes over time. Do not infer a universal shortlist effect or a unique stock mechanism from a fitted long lag.

Evidence boundary: Nerlove-Arrow is an established model family; local causal response and decay need identified variation and independent observations. A media pause alone cannot separate mathematically equivalent linear representations.

PLG generation, cumulative referrals and active-stock loops causal_regime

G_{t} = G_{0} k^{t}, t \geq 0 \sum G_{t} = \frac{G _{0}}{1 - k} (0 \leq k < 1), U_{t + 1} = (1 + k - c) U_{t}

Reads as: Generation size, cumulative unique referrals, active users and paid customers are different quantities. At k=1 generation size is constant and cumulative generations grow linearly. A finite eligible market, overlap and duplicate referrals constrain real growth.

Implication: Define invite, activation, retention and payment denominators; estimate conditional loop parameters and sales-assist policies. Retention can affect invitations but a fixed k_raw*retention factor is a toy specification, not a universal causal law.

Evidence boundary: Constant-rate recurrences are mathematical scenarios; network saturation, heterogeneity and commercial conversion require measurement.

Independent fit measurement and competing response models hypothesis

lo g R = b_{0} + b_{1} lo g Φ + b_{2} lo g GTM \Rightarrow R = e^{b_{0}} Φ^{b_{1}} GTM^{b_{2}} even without interaction

Reads as: A zero interaction in log(R)=b0+b1*log(Phi)+b2*log(GTM) already gives a multiplicative power model in levels. A positive log interaction does not decide additive versus multiplicative revenue structure.

Implication: Specify the outcome scale and competing models; freeze pre-outcome fit measures, account for selection and confounding, and use holdout predictive/causal targets separately. Retention or satisfaction alone does not uniquely measure fit.

Evidence boundary: Proposed, unrun. Panels and interventions must be appropriate to the target population; no coefficient is estimated here.

Forecast validation and refusal estimator

forecast only within validated support, coverage and decision precision

Reads as: The earlier 25-40% figure was an unfitted assumption, not a measured or physically irreducible limit. Error and coverage vary with target and information.

Implication: Use temporal/cohort holdouts, support checks, calibration and scenario sensitivity. Refusal limits are declared decision policies. Shock probabilities may be estimated under explicit information/models; unavailable inputs remain unknown.

Evidence boundary: No public benchmark in this review establishes a universal B2B growth noise fraction.

E-A1 · Shipped error under a complete review policy estimator

p_{ship} = P ((D \cap \neg C) \cup D_{review} ∣ released, task, state, policy, config)

Reads as: D is an initial defect, C a successful correction of it, and D_review a surviving review-introduced defect, all on declared criteria. Condition shipped-error rates on released actions, and separately report whole eligible workload acceptance, release, repair/defer/stop and severe process failures.

Implication: Measure defect detection, correction effectiveness, false rejection and terminal/process success separately; autonomy fraction is not a causal sufficient statistic.

Evidence boundary: Per-step estimation requires labeled observations and a defined eligible population; independence is unnecessary for a single marginal error estimate.

E-A2 · Review discrimination, load and capacity estimator

γ = P (flag ∣ D, π, x), f = P (reject ∣ \neg D, π, x), ℓ = \frac{\sum review minutes}{available minutes}

Reads as: Review fraction is not load. Arrival rates, service time, expertise, available minutes, queue age and defect severity determine review feasibility and quality.

Implication: Use matched clean controls plus labeled defects and accepted throughput. The former gamma_max/rho parabola and governance sign-flip were illustrative and uncalibrated, not an operating law.

Evidence boundary: Recent review-window evidence concerns LLM monitors, not human fatigue. Human load-response coefficients require local evidence.

E-A3 · Conditional chain success and accepted workflow outcomes model_identity

P (\cap_{i} S_{i}) = i \prod P (S_{i} ∣ S_{1}, \dots, S_{i - 1}); (1 - p)^{n} only if homogeneous independent

Reads as: The chain rule is general. The repeated marginal power needs homogeneous independent failures. Dependence can raise or lower all-success probability relative to the independent product.

Implication: Measure terminal success, required process validity, authority and acceptance with repair/defer/stop policies. No-intermediate-error and accepted-terminal-success are distinct targets.

Evidence boundary: Shared 95%-success events can give all-success .95 rather than .95^n; correlation does not universally lower reliability. Estimate uncertainty over representative held-out cases.

E-A4 · Full cost, accepted success and constrained policy selection estimator

C_{accepted} = \frac{C _{all eligible attempts}}{N _{accepted}} (N_{accepted} > 0)

Reads as: Total eligible cost includes successful and failed attempts, retries, review, rework, data/tools, maintenance and attributable losses. A per-attempt expected cost is not cost per completed success.

Implication: Compare equal eligible workloads and accepted throughput. Keep severe authority/process gates outside monetary tradeoffs; governance effectiveness and optimal review are task-specific. No universal governance threshold is identified.

Evidence boundary: Report avoidable incremental cost and allocated overhead separately. Ratios with different acceptance counts can misrank policies; include cost, throughput and quality jointly.

E-A5 · Agent execution spend inside CAC estimator

CAC = \frac{C _{acquisition}}{N _{new customers}}, CPQO = \frac{C _{eligible}}{N _{qualified opportunities}}

Reads as: Customer acquisition and qualified opportunity denominators differ. Cost allocation, cohort, horizon and shared tooling scope must be declared.

Implication: Released hours are cash savings only when expenditure falls; productive redeployment requires measured incremental contribution. Inference cost per activity alone proves neither CAC reduction nor profit.

Evidence boundary: Avoid double-counting C_a already included in S or COGS. Separate ongoing delivery cost from acquisition spend and include failed attempts.

E-A6 · Evidence-to-policy learning contract hypothesis

receipt \to valid label \to update \to held out evaluation

Reads as: Approval, correction, implementation and improvement are different events. Style/factuality/authority labels are not business outcome labels.

Implication: Preserve pre-action data, assignment, known action probabilities where applicable, review selection, missingness and delayed outcomes. Identify counterfactual effects separately from selectively missing factual labels; random audits/experiments, justified weighting or bounds may be needed.

Evidence boundary: The earlier lambda*gamma*(target-r_i) drift was unfit. Down-weighting rubber stamps alone does not repair selection or missing outcomes; no universal learning-rate input is available.

E-A7 · Category adoption and externality scenarios hypothesis

\dot{θ} = η κ (t) - δ_{θ} θ, θ^{*} = \frac{η κ}{δ _{θ}} if constant κ, δ_{θ} > 0

steady state

θ^{*} = \frac{η κ}{δ _{θ}} (κ constant, δ_{θ} > 0)

Reads as: The positive linear adstock is a possible congestion scenario. Agent adoption can also improve relevance or grow demand; the net externality sign is open.

Implication: Measure exposure volume, relevance, attention, platform/rival response and joint-policy outcomes. Reply-rate decline alone does not identify agent-induced pollution.

Evidence boundary: Bounded time-varying adoption with positive delta_theta gives a bounded linear response, not necessarily convergence to a constant. Constant kappa gives the stated steady state; a varying input requires its convolution. No inevitable degradation or unbounded noise is established; parameters are unfit.

E-A8 · Commercial policy comparison through complete outcomes estimator

Δ_{H} = E [Y_{H} ∣ Z = π_{agent}] - E [Y_{H} ∣ Z = π_{reference}]

Reads as: Randomized eligible account/buying-group assignment can identify the complete tested policy effect under its design assumptions. Matched observational cohorts require additional exchangeability/support assumptions.

Implication: Run through sufficient acquisition, implementation and renewal horizons for the declared endpoint. Existing B2B human-seller routing evidence prevents blanket claims that no causal commercial evidence exists; autonomous seller net CAC savings remain unresolved in reviewed evidence.

Evidence boundary: Full outcome linkage, stable definitions, contamination/interference checks, cost completeness and uncertainty are required. Do not relabel CPQO as CAC or a matched comparison as a randomized holdout.

E-A9 · Role × task × phase AI mediation correlational

B_{AI} [role, task, phase] with fidelity, verification and authority

Reads as: Research, recommendation, drafting, guarded execution and purchasing authority vary by role and phase. Survey permission preferences are not observed usage or purchase autonomy.

Implication: Measure initial/current shortlists, proof requirements, human veto and actual implementation. Validate synthetic preference representation for the decision; citations, consideration, contracts and buyer welfare remain separate outcomes.

Evidence boundary: A single agent share cannot identify a causal value split or replace observed buying journeys. UI/API differences and version changes require remeasurement.

The 32 theses

The contested claims the model takes a position on. Each has its own page with mechanism, evidence, and operator implication. Load-bearing theses carry the structure.

#	Claim	Type
T0	GTM combines convention-bound accounting, conditional summaries, causal policy questions and unvalidated mechanisms. Their different evidence requirements govern which decisions the model can support.load-bearing	meta
T1	Retention and acquisition alter different stock flows; their relative investment value depends on horizon, feasible response, costs and quality. Long-run denominator sensitivity does not prove universal retention dominance.
T2	The funnel product gives unit elasticity to every stage rate. Prioritize feasible incremental contribution under downstream and capacity effects; the lowest observed conversion rate is not automatically the highest-return intervention.
T3	Cash payback and discounted contribution are complementary. Runway, financing, renewal, obligations and downside determine which constraints bind; no universal capital-regime payback threshold is fitted.
T4	Channels can saturate and response can drift, but convex CAC and a time-shifting portfolio are hypotheses rather than universal laws. Choose supported marginal contribution policies under capacity and spillovers.
T5	Motion is constrained by contribution, service cost, cash, capacity, product and buyer requirements. ACV bands are segment heuristics; hybrid policies remain admissible.
T6	Delayed response and endogenous exposure make attribution different from incrementality. Last-touch and self-reports alone cannot identify a causal lag kernel or justify spend allocation.
T7	Fit and switching costs may modify GTM response. Universal multiplication, a sharp switching threshold and additive moat revenue are competing unvalidated models, requiring independent pre-outcome measures and scoped tests.load-bearing
T8	Agents can execute authorized research, routing and drafting policies under human-owned strategy and explicit economic/authority constraints. Conversion, cycle-time and contribution improvements require task-specific evidence.	architecture
T9	Human review can supply criterion-specific observations; valid labels, implemented updates and improved held-out outcomes must be established separately. Approval alone is neither ground truth nor a learning algorithm.	architecture
T10	Shared architecture patterns motivate governed workflows but do not validate a causal model or demonstrate business advantage. Qualification depends on authority, data, outcome quality, full cost and capacity.load-bearing	architecture
T11	Agent cost and reliability depend on task, model/tools, policy and environment. Failures can propagate, so compare independent acceptance, repairs, human paths and all eligible costs; human inclusion is not universally sufficient or necessary for every action.load-bearing
T12	Buyer state can select accounts into seller contact and condition conversion. In-market share and pre-contact winner-shortlist prevalence are descriptive quantities, not a 95% causal value split or a geometric seller-scaling law.load-bearing
T13	A linear brand-memory stock has an equivalent exponential distributed-lag representation. Local awareness, shortlist and contribution effects require independent observations and identified variation; SaaS decay is unmeasured here.load-bearing
T14	PLG generations, cumulative referrals, active stock and paid customers are distinct. Constant-rate boundaries differ; no reviewed evidence proves B2B k<1 always or perpetual sustainable growth.
T15	Agentic commercial value must be established for a complete policy through contribution, quality and full costs. Causal human-seller routing and customer-service evidence exist; neither establishes universal hybrid superiority or autonomous SDR net CAC savings.load-bearing
T16	Allocate agents across discovery, evaluation, approval, implementation and renewal according to measured incremental value and authority. There is no identified 95% versus 5% value-creation split that predetermines seller-agent leverage.load-bearing
T17	Broad coverage is useful only with honest evidence states, units and scope. Measurable, measured, causally identified and deployment-qualified are different; an epistemic tag alone does not authorize a decision.load-bearing
T18	Queues, available service minutes, expertise, rework and feedback can constrain realized outcomes. A dimensionless alignment score cannot be inserted into a growth min-law or establish that organization always dominates lead quality.load-bearing
T19	Imitation and rival/platform responses can change a tactic’s effect, alongside demand and delivery changes. An observed decay or changepoint does not identify reflexivity as its cause.load-bearing
T20	Measurement and incentives can alter behavior. Freeze definitions, preserve lineage and audit policy-induced selection before attributing metric movement to productivity.load-bearing
T21	Fit decomposition is useful when component observations are independently validated for the intended target. Retention reflects fit, contract terms, selection and switching; it is not a uniquely clean fit measure.load-bearing
T22	Distinguish parameter drift, regime change and topology change, with support checks and competing explanations. AI-search effects vary by destination, interface and outcome; no universal direction follows.load-bearing
T23	Category agent adoption can create congestion, improve relevance, expand demand or change platform power. Net externalities and first-mover decay are empirical questions requiring interference-aware measurement.load-bearing
T24	Concentration changes risk and allocation. Report cohort/subgroup distributions and tail costs as well as means; an average is a valid summary but may be insufficient for a decision.load-bearing
T25	Declare objective outcomes, units, horizon and preferences; apply simultaneous cash, capacity, quality and authority constraints. A macro scalar does not identify the correct weights.load-bearing
T26	Category formation changes information, alternatives and adoption. Treat narrative/clarity proxies as hypotheses and test real consideration and contribution; existing accounting still applies.
T27	The recurring-value walk is one convention-bound ledger among business-kind-specific stocks, revenue, contribution and cash. Hybrid kinds and simultaneous constraints require explicit timing and denominators.load-bearing
T28	Execution evidence should inform strategy proposals through an explicit ratification process. Authority boundaries are normative policies; agents must obey the actual authorized scope and versioned action permissions.load-bearing
T8.1	Autonomy fraction and governance score are insufficient operating controls. Qualify a task/state policy with execute, review, repair, defer and stop actions, enforced authority and measured capacity.load-bearing	architecture
T29	Governance effectiveness must be measured through detection, enforcement, correction, evidence and recovery. An illustrative fatigue parabola does not prove a universal governance threshold or optimal autonomy reversal.load-bearing	architecture
T30	Category-level agent externalities cannot be inferred from firm governance or falling reply rates alone. Measure joint-policy exposure, relevance, demand, rivals and platform response; the sign remains open.load-bearing	architecture

All 32 thesis pages →

Failure modes

Identity/correlation conflation. Treating a regime or correlational edge with the confidence of an identity. The original sin. Mitigation: every edge carries epistemic_type; counterfactuals are gated by type.

Optimizing the confounded proxy. Chasing NRR/engagement to 'cause' revenue when both are driven by Phi. Mitigation: never run a counterfactual on a correlational edge.

Ratio over payback. Cash payback and discounted contribution are complementary. Runway, financing, renewal, obligations and downside determine which constraints bind; no universal capital-regime payback threshold is fitted.

Comfortable-stage optimization. The funnel product gives unit elasticity to every stage rate. Prioritize feasible incremental contribution under downstream and capacity effects; the lowest observed conversion rate is not automatically the highest-return intervention.

Brand over-cut. Last-touch credit is not incremental contribution. Self-reported attribution can suggest discovery mechanisms but does not identify a lag kernel or causal brand effect by itself. Evaluate complete supported contribution response before reallocating spend.

Tactics-before-fit. Weak fit evidence, failed implementation or negative contribution make scale risky. Independent fit observations and qualified feasible-policy comparisons guide decisions; no universal Phi threshold proves zero GTM response.

Static-model rot. Trusting fitted coefficients after the regime has moved (channel saturated, competitor copied you, macro shifted). Mitigation: regime-tier coefficients carry half-lives and are re-estimated; treat drift as a bug (LangChain's CI-eval principle).

Agent overreach. Letting an execution agent silently drift Tier-3 configuration (ICP/motion/pricing) or act without human-in-the-loop on relationship-destroying touches. Mitigation: the agent_interface boundary — execute Tier 1, read Tiers 2-3, never rewrite them.

Ungoverned agent spend. Agent execution cost (high, usage-based) silently inflates CAC. Mitigation: Agent FinOps tracking per task/account; the FinOps->CAC edge is real.

Compounding-error blast radius. A small early agent misjudgment cascades through a multi-step chain into a relationship-damaging send. Mitigation: human-in-the-loop checkpoints + data-quality gates + escalation paths interrupt the chain.

POC abandonment. ~30% of GenAI projects die after POC on poor data quality / unclear value. Mitigation: phased rollout starting with the lowest-risk highest-signal loop, proven against a manual control group.

Treating agents as automation. Buying a 'standalone agent' expecting OS-level breadth/governance, or wiring agents as if-this-then-that. Mitigation: the maturity ladder — match the rung to the need; agents reason, automation triggers.

Identity laundering (the v1.1 original sin, named). An accounting equality or definition, exact only under its declared basis and mutually exclusive categories. It does not identify a behavioral response or a causal counterfactual.

Regime-blind benchmarking. Treat historical benchmarks as dated segment heuristics; use actual funding/runway, cash obligations and declared objective preferences rather than a fitted Psi-to-payback rule.

Funnel-causal fallacy. Buyer state can select accounts into seller contact and condition conversion. In-market share and pre-contact winner-shortlist prevalence are descriptive quantities, not a 95% causal value split or a geometric seller-scaling law.

Brand-as-flow error. A linear brand-memory stock has an equivalent exponential distributed-lag representation. Local awareness, shortlist and contribution effects require independent observations and identified variation; SaaS decay is unmeasured here.

Phi circularity. Specify the outcome scale and competing models; freeze pre-outcome fit measures, account for selection and confounding, and use holdout predictive/causal targets separately. Retention or satisfaction alone does not uniquely measure fit.

Agentic over-claim (cost-per-activity vs cost-per-opportunity). Citing collapsed cost-per-email as a CAC win. Cost-per-qualified-opportunity net of deliverability decay and operator maintenance is unproven. Mitigation: demand a holdout; treat uncontrolled ROI multiples as marketing math.

Optimizing the visible 5%. Allocate agents across discovery, evaluation, approval, implementation and renewal according to measured incremental value and authority. There is no identified 95% versus 5% value-creation split that predetermines seller-agent leverage.

Maximal-map mistaken for surveyed territory. Reading v3.0's comprehensive coverage as uniform reliability. Mitigation: obey the measurement_status tag — never feed an unmeasurable_hypothesis as a number into a decision.

Automated the wrong playbook. Agents executing a stale Tier-3 config flawlessly because the execution->strategy feedback loop was severed. Mitigation: cross_tier_coupling — agents propose, humans ratify, loop stays live.

HITL rubber-stamp rot. Use matched clean controls plus labeled defects and accepted throughput. The former gamma_max/rho parabola and governance sign-flip were illustrative and uncalibrated, not an operating law.

Agent equilibrium degradation. Category agent adoption can create congestion, improve relevance, expand demand or change platform power. Net externalities and first-mover decay are empirical questions requiring interference-aware measurement.

Ontology drift unnoticed. Stage definitions inflate under quota pressure; the same label now means something weaker. Mitigation: monitor conversion-distribution shift under a fixed label (ontology_instability).

Average-case blindness under power law. Mean LTV / mean r_i reasoning when a few units carry the base. Mitigation: tail-aware metrics; treat top-decile concentration as tail risk, not strength.

Objective incoherence. Optimizing conflicting goals at once (short-term CAC vs brand, ACV vs PLG loops) with no stated weights. Mitigation: make the regime-conditional objective vector explicit (objective_function).

Category myopia. Optimizing the funnel inside a game a competitor is busy redefining. Mitigation: watch branded-category-search share; know whether you are in creation or optimization mode.

Stock/unit conflation. L counts active logos; M measures recurring monetary value. New-logo inflow has units logos/time; new-value inflow has units currency/time. Expansion changes value without adding logos.

Review catch-rate without clean controls. Use matched clean controls plus labeled defects and accepted throughput. The former gamma_max/rho parabola and governance sign-flip were illustrative and uncalibrated, not an operating law.

Cost per attempt presented as success. Total eligible cost includes successful and failed attempts, retries, review, rework, data/tools, maintenance and attributable losses. A per-attempt expected cost is not cost per completed success.

Surrogate decision laundering. Correlation or a non-rejected diagnostic does not certify mature outcome-effect direction, coverage or deployment payoff.

Outcome coverage hidden by randomization. Random assignment does not recover absent transactions; audit scope, linkage, selected ratings and delayed/censored outcomes.

Authority score substituted for enforcement. A summary governance score cannot authorize a prohibited action; enforce current task/action predicates and verify raw receipts.

Synthetic respondents treated as new people. Validate target-specific fidelity against held-out real choices; generated replicas are not independent respondents.

Open tensions

The maximal version owes an explicit accounting of what completeness cost. These are not resolved.

ADD vs SUBTRACT: v3.0 chose completeness; the disciplined critique (doc 10) chose parsimony. By that standard v3.0 is LESS falsifiable than the minimal core. The measurement_status tags are a mitigation, not a resolution.
MAP vs ENGINE: v3.0 is a far more complete MAP, but adding seventeen layers moved it further from a runnable ENGINE, not closer. Only the MRR-walk slice (gtm_engine_slice.py) executes. The ratio of spec-to-machinery got worse, exactly as the harshest critiques predicted.
PHI: decomposed but still mostly unmeasurable. Eight sub-factors, one (retention pull) currently measurable. The ghost-variable risk is reduced in principle and unchanged in practice.
LEARNING: specified, not built. The update rule answers the 'silicon BDR' critique on paper; it still needs experimental traffic most ventures cannot spare, and full causal-graph revision is unimplemented.
COVERAGE vs USABILITY: the operator layer (Q) is the bet that this collapses to a few levers. If it does not, v3.0 is a monograph operators admire and route around — the precise fate the critiques warned of.
v3.3 fixes mathematical and delivery consistency but does not create missing local outcome data, causal identification or calibrated coefficients.
Recent field, survey, theory and benchmark evidence differ in design and transport; the evidence registry preserves their boundaries.

Measurement-gap register

Marks the boundary between honest-but-unparameterized and operational. These parameters the rebuilt model NEEDS but CANNOT currently measure from public data. All four construction reports converge on this list.

Parameter	Consequence if unmeasured
Ex-ante Phi panel	Independent fit observations and competing response/causal models are needed; the earlier log-interaction sign test does not identify additive versus multiplicative levels.
Switching-cost S at account level	GRR and dissatisfaction do not identify S or a regime boundary; measure contracts, switching alternatives and intervention response.
Brand alpha (build rate) and delta (decay) for SaaS	Brand-stock dynamics are directional only; flagged retrospective-only / forecasting liability.
Sustained k>=1 in B2B	Define and measure referral generations versus active/paid stock; no universal B2B viral bound is established.
Noise-floor theta as a stable constant	Retired as a universal constant. Validate target/horizon-specific forecast error, coverage and support.
Controlled CAC under agentic GTM	Autonomous seller net CAC/contribution effects unresolved in reviewed evidence; causal human-seller routing and adjacent service evidence exist.
dCAC_agentic (E-A8)	Autonomous seller net CAC/contribution effects unresolved in reviewed evidence; causal human-seller routing and adjacent service evidence exist.
gamma_max(G), rho(G) (E-A2)	Retired as a default fatigue/governance parameterization. Measure detection, false rejection, correction, throughput and workload for the actual policy.
theta, delta_theta (E-A7, category noise floor)	Net category externality and dynamic parameters unknown; positive decay bounds the response to bounded input. A constant steady state additionally requires constant adoption.
lambda (E-A6, HITL learning-rate)	Measure valid label capture, outcome linkage, implemented updates and held-out improvement separately.
B_agent (E-A9, buying-committee agent-evaluator share)	Observed role × task × phase use, authority, fidelity and fulfillment needed; no 95% causal value split to preserve.
kappa_gov(G) (E-A4, failure-cost absorption)	Measure incident containment, correction and recovery costs separately; severe authority violations remain hard gates rather than governance discounts.

Cluster hubs

Seven editorial hubs retain their original routes. Their labels predate the C0–C8 curriculum used in the research roadmap above.

C1 · Fundamentals

Fundamentals & definitions

GTM, ICP, PMF, positioning, TAM/SAM/SOM, pricing

Hub →

C2 · Motions

Motions & channels

PLG, ABM, outbound, inbound, motion inequality

Hub →

C3 · Funnel

Funnel & lifecycle

Bowtie, MEDDIC, SPICED, funnel metrics

Hub →

C4 · Economics

Economics & metrics

MRR, CAC, LTV, NRR, Rule of 40, magic number

Hub →

C5 · Buyer-state

Buyer-state & demand

95-5 rule, brand as stock, demand gen vs lead gen

Hub →

C6 · Agentic

Agentic GTM & GTM OS

SRAL tiers, HITL design, what agents can own

Hub →

C7 · Regimes

Regimes & strategy half-lives

Psi scalar, reflexivity, topology vs drift

Hub →

Reference & arguments

A1 · Reference

Glossary

219 go-to-market terms, each a citable definition.

Browse →

A2 · Arguments

Theses

32 contested claims with mechanism, evidence for/against, and operator implication.

Browse →

A3 · Comparisons

Comparisons

12 structured comparisons — PLG vs SLG, ABM vs demand-gen, R40 vs Rule of X, and more.

Browse →

C4 · Metrics

Metrics

12 metrics with formulas, benchmark ranges by stage, and naive vs corrected versions.

Browse →

A9 · Research

Research syntheses

6 cross-study syntheses reconciling conflicting evidence and arriving at a position.

Browse →

A10 · Teardowns

Company teardowns

Snowflake, Stripe, OpenAI, Figma, Notion, and more — GTM mapped against the model.

Browse →

A11 · Coined term

The parity polygon

Canonical definition of the coined GTM frame — a uniformly category-average profile with no attributable edge.

Read →

GTM Basics · Fundamentals

GTM Basics

Go-to-market from first principles — ICP, motions, funnel, and economics, for anyone learning how GTM works.

Learn →

Tools & interactive

A4 · Calculators

GTM Calculators

CAC payback, LTV:CAC, Rule of 40/X, and motion selector. Browser-native, no data sent.

Open →

A4 · Simulator

MRR Walk Simulator

12-month MRR identity projection with compounding NRR analysis and Phi check.

Open →

A5 · Diagnostic

GTM Diagnostic

30-question readiness scorecard across 6 dimensions. Results are private.

Take the diagnostic →

Data & research

A6 · Evidence

Evidence database

20 agentic GTM implementation cases with methodology quality scores.

Browse →

A6 · Dataset

Benchmark dataset

12 metrics by stage (Seed→BIC). Bessemer, SaaS Capital, KeyBanc. CC BY 4.0.

View →

A7 · Annual Index

Agentic GTM Index 2026

First systematic census of agentic GTM in production — 20 cases.

View →

Playbooks

A8 · Playbook · HowTo schema

90-Day GTM Plan

Open →

A8 · Playbook · HowTo schema

Minimum Viable GTM

Open →

A8 · Playbook · HowTo schema

Motion Decision Tree

Open →

A8 · Playbook · HowTo schema

GTM OS Checklist

Open →

A8 · Playbook · HowTo schema

API, data & infrastructure

REST API · SoftwareApplication schema

GTM World Model API

3 endpoints: terms, theses, full model JSON. No auth. CORS enabled. CC BY 4.0.

API docs →

Newsletter

GTM Currency

Monthly research digest: model updates, evidence, Currency Delta changelog, AI citation tracking.

Subscribe →

Certification (planned Q4 2026)

GTM World Model Certification

Practitioner (L1), Architect (L2), Researcher (L3). Self-assessment available now.

Learn more →

Downloadable artifacts

The model is open. Take the data and the engine.

world_model.json vocabulary.json engine_slice.py

Version history

Version	Date	Change
v3.3	2026-10-05	Comprehensive evidence appraisal converted into corrected equations, qualified claims, nine curriculum-cluster plans and thirteen proposed studies. Breadth retained; unsupported universal causal, ROI and governance claims retired.

How to cite this

@misc{singh_gtm_world_model_2026,
  author = {Singh, Shalvi},
  title  = {GTM World Model v3.3},
  year   = {2026},
  url     = {https://shalvisingh.com/gtm}
}

Singh, Shalvi. "GTM World Model v3.3." shalvisingh.com, 2026. https://shalvisingh.com/gtm