GTM research compendium

1 · the words

The vocabulary

One term per concept, in the order the chain produces them. Every later section uses these words and only these.

2 · fit

ICP: the method, and what each source gave us

An ICP is a filter over observable attributes, derived from payers first, and it is only real if it passes three tests. That is the synthesis; the sources below each contributed one piece.

The three tests an ICP must pass

The four parts of the ICP object

• Observable attributes, each checked one of a fixed set of ways (a page check, a directory field, a feed field, a search hit, a numeric range). The model fills in the parameters; it never invents a new kind of check.
• The trigger it is tied to (lives in the timing layer, not the fit layer).
• The pain, verbatim, with its evidence tier.
• Exclusions: who looks like a fit and is not.

Who said what

Three things the sources disagree on, and how we resolved them

• Do trigger events belong in the ICP? Kellogg no, Bowery and Winning by Design yes. Resolution: two layers in one object, a stable fit layer that filters and a perishable timing layer that queues.
• How narrow? Lenny "comically narrow," Dunford "as narrow as your near-term sales number allows," Moore "small enough to lead." Resolution: the starting segment is sized by the denominator test, and it must still produce enough candidates to measure (about 30 a month to read a rate).
• Where do you start with zero customers? Kellogg regression needs payers; Hoy/Hillman observation does not. Resolution: tier A from payers when they exist, tier C from the ICP's and competitors' public words when they do not, and the tier is recorded so the derivation knows how much to trust itself.

3 · timing

Triggers and timing

Fit tells you who; timing tells you when. The timing layer is a queue of people inside the fit filter who just said something that means they are looking. It is perishable and re-read daily.

The window. Day 3 to 45 after the triggering event, peak day 5 to 21, a second smaller spike at day 28 to 35 (the first billing cycle). Day 0 to 2 is euphoric and unreachable; after day 60 the founder is rationalising the exit.

The seven detectable states (each maps to a phrase pattern a detector looks for):

The compound phrase pattern. Elapsed time + a zero ("4 days", "0 signups") is the highest-precision signal in the corpus. Either half alone is noise.

Nine noise classes the detector must reject at zero false positives: retrospective wins ("a year ago I had 0, now…"), advice-givers quoting the pattern, engagement bait built on it, hiring posts, agencies fishing, launches with paid promotion, product-hunt hunters posting for others, people describing someone else's launch, and pre-launch "about to launch" posts (outside the window).

Evidence tiers decide how much a phrase is trusted: A is what paying customers did and said, B is how signups behave, C is the market (what buyers and competitors' customers say in public), D is claims and guesses, including a model's paraphrase. The tier comes from the source the phrase was found in, checked by matching the phrase against the source text; a model never sets a tier, and D is never used as a live query.

4 · where we read

Data sources, verified

Where the seed stage reads. Each was opened and its terms checked; limits are as found on 2026-09-19/20.

Directory-level facts that shaped the design: Show HN's real volume, once measured, was about three times the first estimate; the launch check must be evaluated relative to a frozen as_of date or every fixture row ages out in a month; a promo price displayed larger than the list price (a $6 promo shown over a struck $29) and prose figures near the real price ("burns $500/month", a "$40,000" ad-budget illustration) are the three ways a naive price check fails.

5 · learn

Campaigns, the funnel, the loop

Campaign = one segment × one playbook × one offer × one window, with a written hypothesis and a kill date. Several can run at once under the contention rules; each is measured alone.

The funnel equation

paying = reached × visit% × signup% × activate% × pay%

Every term is labelled measured or assumed with its inputs and an as-of date. The constraint is the weakest measured term, or the first term with no data at all; an unmeasured step outranks a small one. Capacity binds past about 40 candidates a month, which is why the starting segment is sized to what one founder can hold.

Read tiers (how a rate is allowed to be stated):

No bandits under 100 per arm. No "winner" on 12 clicks.

Attribution

• Touch id in utm_content on every link the engine sends; carried through visit, signup, payment.
• The Stripe webhook is the sole writer of paid rows.
• Three columns per customer, first touch / last touch / self-reported, never summed. Self-report ("how did you hear about us") is the primary instrument because most of the funnel is dark.

The three-speed loop

Deterministic weekly triggers (Monday read). Prompts never rewrite themselves; every change is a decision with a date, a reason and a number. The Monday report names the funnel, the constraint, this week's one test, and what was retired.

Build order: stages 1 to 4 (harvest, derive, gate, seed) get built and run against a real brief, not skipped. The gate has four preconditions and refuses to start a campaign the funnel cannot measure.

6 · data and experiments

Data capture and the experimentation engine

Every stage of the pipeline is a decision, and every decision is training data if it is recorded in the right shape from day one.

The data spine: eight rules every stage follows

1. Log the decision, not just the result: every choice by the system, the founder or the lead is an append-only event naming who made it.
2. Version whatever decided: model, prompt, rule, scorer, ICP and detector versions on each row.
3. Freeze the inputs: store the features as they were at decision time; recomputing them later leaks the future into training data.
4. Keep the negatives: discarded matches, skipped leads, blocked sends and "no" replies, each with its reason.
5. Log the odds: the probability of each choice, plus a small random slice, so a new model can be judged on old data.
6. Every human touch is a label: edits, corrections, skips and never-contact marks, with before, after and reason.
7. One chain of ids: detector → signal → person → lead → move → touch → account → Stripe customer.
8. Keep the raw, with a retention rule: full text, the url, posted and captured timestamps; public data only, deleted on request.

The experimentation engine sits across all 12 stages: any stage's rule, prompt, threshold or play is a knob it can test.

The per-stage detail (what each stage records, the label it creates, and the model it trains) is on the pipeline overview.

7 · reference

Sources

Every page opened during the research, grouped, with what it gave us. Section 2's table has the per-source method notes; this is the flat list for citation.

ICP method

• Kellogg, Your ICP starts as an aspiration and becomes a regression — the fit/timing split, regression on outcomes
• Dunford, quickstart guide to positioning — best-fit by behaviour, positioning order
• Dunford on Lenny's podcast — don't over-commit pre-PMF
• Obviously Awesome notes, Nat Eliason and Will Patrick — the exercise
• Dunford, positioning exercise on HackerNoon
• Crossing the Chasm notes and Moore on Lenny's — beachhead, referenceability, Documentum's 40 departments
• Lenny Rachitsky, how to identify your ideal customer — founder survey, nine attributes, outbound as the instrument
• First Round Review, Superhuman's PMF engine — the 40% survey as segmentation; PG quote
• Kyle Poyar, ICP funnel note — the denominator test
• Jason Cohen, Selling to Carol — pain in their words, the 3-to-10-second test
• Alex Hillman on Sales Safari and Stacking the Bricks EP42 — zero-customer observation
• Craig Elias, SHiFT preview chapter — trigger events, Window of Dissatisfaction
• Rob Walling, Start Small Stay Small notes — niches congregate
• Four Forces, jobstobedone.org
• Unusual Ventures field guide, Clearbit, TAM/ICP/personas, Bowery Capital founder's guide, Winning by Design blueprint, Bob Apollo, no fit no deal — hierarchy and qualification order
• HXC, secondhand; PG notes, How to Get Startup Ideas; YC library, how to talk to users (not extractable)

Data feeds

• HN Algolia search_by_date?tags=show_hn; PeerPush /api/v1/launches; Product Hunt Atom feed; Uneed; BetaList; TrustMRR /api/ai, /api/ai/discovery, /recent