Product-market fit for an AI product is two conditions, on two axes, that fail independently. The first is the one Marc Andreessen named in 2007 and Rahul Vohra operationalized at Superhuman in 2018: would users be very disappointed if the product disappeared. That axis still binds. The second is new in 2026: does the model carry the load at acceptable margin and reliability, and does it continue to do so as the frontier moves. Most of the 2024–2026 AI startups that died at the demo-to-revenue gap died because their founders measured only the first axis. This article names the parallel definition, the failure quadrants, the weekly signals, and the three traps that hit AI products specifically.
This spoke sits inside the idea validation playbook and the broader idea-to-product manifesto. Where the playbook describes the first weeks of work that produce a PRD, this article names the operating definition founders run against the shipped product weekly.
Table of Contents
- The Classical PMF Definitions, Restated
- Why AI Products Complicate PMF
- The Two-Axis AI PMF Frame
- Five Weekly Signals an AI Founder Should Track
- The Three PMF Traps Specific to AI Products
- A Measurable Weekly Definition of AI PMF
- Frequently Asked Questions
- Closing
The Classical PMF Definitions, Restated
Two definitions remain load-bearing for any AI product founder.
The first is Marc Andreessen’s, from his 2007 essay “The Only Thing That Matters”: “You can always feel product/market fit when it’s happening. The customers are buying the product just as fast as you can make it… and money from customers is piling up in your company’s checking account.” Qualitative, market-pull-shaped. A founder shipping to a market that pulls the product out of the company knows it; a founder who is not, should not be reassured by a vanity dashboard.
The second is Rahul Vohra’s, from his 2018 First Round Review essay on the Superhuman PMF engine. Vohra borrowed Sean Ellis’s question — “How would you feel if you could no longer use this product?” with four choices: very disappointed, somewhat disappointed, not disappointed, no longer using — and named 40% “very disappointed” as the threshold for early PMF, a leading indicator that does not require waiting for a hockey-stick retention curve.
For SaaS products from 2007 to 2022 these two definitions, applied together, were close to sufficient. The shipping team only had to keep the product working — a deterministic question with a deterministic answer. AI products break that last assumption.
Why AI Products Complicate PMF
Three structural reasons make the classical definitions necessary but not sufficient for a 2026 AI product.
The model layer changes underneath the product. Frontier models update on cycles of weeks. Over the past 24 months OpenAI shipped the GPT-5 family; Anthropic shipped Claude Opus 4 through Opus 4.8 and Claude Sonnet 4.6; Google shipped Gemini 2.5 Pro; the open-weights frontier moved from Llama 3 to Llama 3.3 (Artificial Analysis LLM Leaderboard). Each transition silently changes the behavior of any system pinned to a model alias. A SaaS product that passed PMF on Tuesday still works on Wednesday; an AI product may quietly regress overnight.
The same prompt today and six months ago produces a different answer. This is a property of the substrate, not a flaw of one provider. PMF for an AI product cannot be a one-time measurement; it is a continuous measurement against a frozen evaluation set.
The user’s reference for “good” is moving faster than the product. A user trying an AI product in mid-2026 is comparing it to the latest frontier model in their pocket, not to the SaaS competitor in the adjacent tab. The PMF baseline is the SOTA reference, which lives on the leaderboard, not in the founder’s head. Sequoia’s “AI’s $600B Question” essay and a16z’s 2024–2025 builder essays both name this dynamic; neither prescribes a measurable response.
The takeaway: AI PMF is classical PMF plus a model-axis measured separately and continuously.
The Two-Axis AI PMF Frame
The parallel definition is a two-axis frame.
User-PMF (the Andreessen / Vohra axis). Would the user be very disappointed if this product disappeared? Are they returning? Are they pulling new use cases out of the product without prompting?
Model-PMF (the new axis). Does the model carry the workload at acceptable margin and reliability? Is the eval pass-rate stable or improving across model updates? Is the cost-per-successful-task curve flat or declining? Is the ratio of model-handled to human-fallback requests above the threshold the unit economics require?
| Strong model-PMF | Weak model-PMF | |
|---|---|---|
| Strong user-PMF | Real PMF. Ship, scale, defend. | Loved but unaffordable or unreliable. Re-engineer the model layer or the workflow degrades into a costly novelty. |
| Weak user-PMF | Capable but unloved. The model works; the workflow is wrong. Re-scope the user job. | No PMF on either axis. Do not raise on this product. |
The lower-left quadrant (strong user-PMF, weak model-PMF) is the most common 2024–2026 failure mode. The founder sees retention. The Vohra survey hits 42%. NPS reads positive. The founder concludes PMF and raises a Series A. Three months later the inference bill outpaces revenue at every defensible price point, the model degrades after a provider update because no eval suite was checked in, and the company is burning growth-stage capital on revenue that cannot grow because every new user is unprofitable.
The upper-right quadrant is also instructive: the model handles the job, eval pass-rate is high, cost-per-task is low — and users do not return. The product is technically capable but the workflow is wrong. The fix is user work in the stage-2 falsification sense, not more model work. A founder running only the Vohra survey cannot see this matrix. A founder running both axes can.
Five Weekly Signals an AI Founder Should Track
The two-axis frame becomes operational as a weekly dashboard. Five signals, all founder-runnable:
-
Retention by use-case, week-by-week. Not a single number — a breakdown by named jobs-to-be-done. A blended number masks PMF on one job and absence on another.
-
Eval pass-rate on a frozen test set, run on every meaningful change and every model update. This is the model-PMF anchor: 50 to 150 representative cases with expected behaviors and a quality threshold, checked into version control. A dropping pass-rate is a model regression even when retention has not yet caught up.
-
Cost-per-successful-task curve. Inference dollars divided by successful tasks (defined by the eval rubric, not the user clicking “OK”). Should be flat or descending. A rising curve is a model-PMF failure even when users are happy.
-
NPS broken out by use-case. A blended NPS hides the same heterogeneity blended retention does. A use-case tag shows that workflow-A users sit at NPS 58 and workflow-B users at NPS 12 — the signal that points to the next PMF win or rescue.
-
Ratio of model-handled to fallback-handled requests. Most AI products in production have a human-in-the-loop fallback or a rule-based escape hatch. The ratio of clean model completions to fallback completions is the operational read on model-PMF. A drifting ratio is the leading indicator of every quiet model regression.
These signals are a Friday-afternoon dashboard the founder builds in a spreadsheet from production logs in week one.
The Three PMF Traps Specific to AI Products
Three traps are recognizable in the field and worth naming so founders can avoid them by name.
Trap 1: Model-version-pinned-PMF
The product was strong when the founder pinned to a specific model version at launch. No eval suite, no testing against new releases, no migration budget. Six months later a competitor ships the same workflow on a newer frontier model. The user’s SOTA reference has moved. The product has not. Retention bends down two quarters in. Honest diagnosis: model-PMF eroded silently with no instrumentation to see it.
Trap 2: Demo-PMF
The product works perfectly on the inputs the founder used for the YC demo, the seed pitch, the launch video. These inputs were curated, often unconsciously. Production traffic looks different. The pass-rate on a representative sample of real traffic is meaningfully lower than on the demo set. The founder runs the Vohra survey on a survivorship-biased sample of remaining users and reads 42%. Honest diagnosis: demo-PMF on a curated input distribution is not PMF. The fix is a representative eval set sampled from production logs.
Trap 3: Free-tier-PMF
Users return. Retention is strong. NPS is positive. The founder has not asked anyone to pay because the seed round is subsidizing the inference bill. When the willingness-to-pay test is finally run — typically two quarters in — conversion-to-paid is at a level that does not survive the cost-per-task curve. Honest diagnosis: free-tier-PMF is not user-PMF. A user returning to a free service whose marginal cost is absorbed by venture capital is reporting one fact: the service is free. Charge before you celebrate.
Naming each trap is half the defense. The other half is the weekly dashboard above, instrumented to surface the trap signal before the founder mistakes it for a win.
A Measurable Weekly Definition of AI PMF
Pulling the frame and the signals together, here is a founder-runnable weekly definition of PMF for an AI product:
A 2026 AI product has product-market fit when, four weeks in a row:
- Retention by primary use-case is above the threshold the founder set at launch
- The Vohra “very disappointed” survey clears 40% on a non-survivorship-biased sample
- The eval pass-rate on a frozen production-representative test set is at or above the PRD threshold, with a flat or improving four-week trend
- The cost-per-successful-task curve is flat or declining
- The model-handled-to-fallback ratio meets the threshold unit economics require
- Model-handled tasks stay at threshold across the most recent update from at least one frontier provider the product depends on
Four weeks is intentional — PMF is a steady state, not a Tuesday. Six conditions are intentional — each closes one quadrant or trap. Founders who run this will sometimes find three clear and a fourth slowly fail. The honest read: they are at “PMF on the conditions I am measuring,” not yet at PMF.
Frequently Asked Questions
Is the 40% “very disappointed” benchmark still valid for AI products?
Yes, as one condition of six. Vohra’s Superhuman methodology and Sean Ellis’s 40% threshold remain a reliable read on the user-PMF axis. They do not measure model-PMF at all. Clear 40% on Vohra and the four model-PMF conditions in the weekly definition above.
What is the difference between user-PMF and model-PMF?
User-PMF is the classical condition: would users be very disappointed, are they returning, is the market pulling. Model-PMF is the AI-specific condition: does the model carry the load at acceptable margin and reliability, is the eval pass-rate stable across updates, is the cost-per-successful-task curve flat or declining, is the fallback ratio inside the bounds the unit economics require. Both axes fail independently. For the upstream framing of what counts as the “product” being measured, see what is an AI MVP, actually.
Why isn’t the Vohra survey enough for an AI product?
The Vohra survey measures user affect at one point in time. It does not see model behavior, inference cost, or model-version risk. An AI product can clear 40% on a survivorship-biased sample and quietly lose model-PMF when the frontier moves, a provider update changes behavior, or the cost-per-successful-task curve bends the wrong way. Necessary, not sufficient.
How often should an AI product re-measure PMF?
Continuously on the model-PMF axis, weekly on the operational signals, monthly on the Vohra survey. The classical SaaS pattern of measuring once and banking the result does not survive the model substrate. Run the eval suite on every commit and every model update; review the Friday dashboard weekly; re-run the Vohra survey after each meaningful release.
What is the most common AI PMF failure I should expect?
Strong user-PMF, weak model-PMF — the lower-left quadrant. Users return, NPS is positive, the Vohra survey reads above 40%, and either the inference bill outpaces revenue at any defensible price, or the eval pass-rate is silently regressing across model updates. The product feels real and is operationally underwater. Instrument model-PMF before the user signal convinces the team there is nothing left to debug.
How do I run a model-PMF check if I cannot write code?
Build the eval set in a spreadsheet — 50 to 150 representative inputs, expected behaviors, and a quality rubric the founder can score by hand. Run the inputs through frontier models in their chat consoles weekly. Score the outputs against the rubric. Track the pass-rate over time on the same sheet. The sibling spoke how to validate an AI product idea before you write a single prompt walks through this exercise step by step.
What is the connection between AI PMF and the eval suite?
The eval suite is the only instrument that measures model-PMF directly. Without one in version control and run on every model update, the founder cannot distinguish a quality drop from a flake, a model regression from a prompt regression, or a real PMF erosion from noise. See decoding the AI agency case study for what a credible eval-suite artifact looks like in a partner’s portfolio.
Should I aim for PMF on the cheapest model or the best model?
Measure PMF on the model the unit economics require, not the one that produces the prettiest demo. A demo-PMF on Claude Opus 4.8 that does not survive the move to a cheaper model is a forecast that cost-per-successful-task will permanently exceed revenue. Evaluate the workflow on both a frontier and a workhorse model; set the PMF threshold on the model the price point can absorb; only use the more expensive model where the eval pass-rate gap justifies the spend.
What is the relationship between AI PMF and willingness-to-pay?
Willingness-to-pay is the falsification test for free-tier-PMF. A user returning to a service whose marginal cost is absorbed by venture capital is reporting one fact: the service is free. Run a paid pilot at the price point the founder needs for sustainable unit economics, with the eval pass-rate shared with the customer, before treating user-PMF as real.
Closing
The classical PMF definition is a 2007 instrument used on a 2026 substrate. Still useful, no longer sufficient. AI products need a parallel definition with a second axis — model-PMF — measured separately and continuously, with a cost-per-successful-task curve and a fallback ratio next to the retention dashboard. Founders who run the two-axis frame, the weekly definition, and the three named traps will catch the lower-left-quadrant failure mode that killed most of the 2024–2026 AI startups that died with positive Vohra readings and negative bank accounts.
The next step inside this cluster is the operational sequence that produces the PRD this instrument plugs into: the idea validation playbook.
— Arthur Wandzel, CEO, SFAI Labs
Arthur Wandzel
