Product-market fit (PMF) is the condition in which a product satisfies real market demand with sufficient intensity that growth is organic and retention is high. The classic cycle (build, ship, iterate, scale when demand requires it) remains valid. What has changed with the proliferation of LLMs are the environmental conditions in which that cycle operates. The quality baseline users expect is higher, moats are different, and the signals distinguishing real PMF from illusory traction are also distinct.

Key takeaways

  • The quality baseline has risen: an LLM-powered MVP looks impressive in demo, making it harder to distinguish real adoption from technical curiosity.

  • Classic PMF metrics still work (retention, NPS, revenue expansion), but need to be complemented with LLM-specific metrics.

  • The thin wrapper problem: a product built only on an LLM’s API can be copied in days and has compressed margins from model cost.

  • Durable moats in AI are proprietary: accumulated user data, deep workflow integration, and network effects that improve the model with use.

  • The most reliable PMF signal in AI products remains the same as in any other: users who would refuse to stop using it if you removed it.

What has changed from before LLMs

The quality baseline has risen: ChatGPT and its successors have trained millions of users to expect high-quality responses. A product that would previously have impressed with its sophistication, such as an assistant that answers technical questions well, now faces high expectations from the first use. First impressions no longer suffice to differentiate. The minimum bar for generating genuine PMF, rather than just trial engagement, is higher than it was before 2022.

Competitor iteration is faster: if your product does something valuable with an LLM, a competitor with access to the same API can ship a basic version in days. The window to establish market position before copying appears is shorter than in traditional software.

Moats are harder to build and more durable when well built: a product purely based on API calls to OpenAI or Anthropic has compressed margins and minimal differentiation. A product that accumulates user data improving the model, integrates deeply into user workflow, and has network effects is hard to replicate.

Capital is more available: the abundance of capital in the AI sector means more competitors can get funded. Differentiation and the path to a durable moat matter more now than in previous software cycles.

Classic PMF metrics that still work

  • Retention cohorts: do users who adopted the product in month X still use it in month X+3, X+6? The retention curve that stabilises (does not converge to zero) is the most reliable PMF signal. Cohort analysis distinguishes products with genuine habit formation from those generating one-time curiosity.

  • DAU/MAU ratio: on how many days a month is the product the answer to a recurring problem? A ratio above 0.3 to 0.4 is the threshold for daily or near-daily use rather than occasional use.

  • NPS: do users recommend without being asked? Organic recommendation is the cheapest growth signal and one of the hardest to fake. It reveals whether users genuinely value the product or are merely tolerating it as a second-best option.

  • Revenue expansion: Net Revenue Retention (NRR) above 100% indicates the product solves problems that grow with use. It means existing customers expand their spend faster than churn erodes it, which is the definition of a sustainable growth engine.

LLM-specific metrics

  • Task success rate: does the model actually solve the problem the user came with? Unlike engagement metrics, task success requires explicit instrumentation or sampled evaluation, because it is not passively observable from clickstream data alone.

  • Prompt reformulation rate: if users reformulate more than once per session, the first response was not satisfactory. High reformulation rates signal either a quality problem in the model output or a mismatch between user intent and what the product expects them to ask.

  • Cost per valuable outcome: how many tokens (and how much money) to reach a response the user accepts as useful? This metric connects model efficiency with unit economics: a product where each valuable response costs $0.50 to produce has a fundamentally different margin profile than one where it costs $0.01.

  • Perceived latency: in streaming products, TTFT (time to first token) is what users feel as fast or slow. In products with streaming, high total latency with low TTFT feels faster than the reverse, because the streaming start gives the user immediate feedback that something is happening.

The thin wrapper problem

Signs a product is a thin wrapper:

  • COGS ≈ API markup: gross margin is low because the main cost is the model.

  • No proprietary data: the product does not accumulate data improving experience over time.

  • No workflow integration: the user can get the same result going directly to Claude.ai or ChatGPT.

  • The LLM is 100% of the value: the full value proposition can be replicated by anyone with API access.

A thin wrapper can generate initial revenue, particularly in segments of users who would not use the API directly. But it does not scale as a business: margins are compressed, differentiation is minimal, and the product can be copied in days.

Where real moats are built

Proprietary data with flywheel: the product improves with use because each interaction generates data refining the model, prompt, or experience. More users → more data → better product → more users. This flywheel is extremely hard to replicate by a competitor starting from zero, even using the same base model.

Deep workflow integration: a product embedded in user workflow has high switching cost. The user faces more than a learning curve: they must also migrate data, integrations, and habits. Cursor, for example, is embedded in how a developer writes code, not just how they ask questions.

Network effects: in some products, more users makes the product more valuable for each individual user: marketplaces, collaboration tools, shared knowledge bases.

UX the API alone cannot provide: the model is a commodity; how interaction is organised, what context is managed between sessions, and what workflow is built on the model is not.

Cases that found PMF and cases that did not

Products that found real PMF:

  • Cursor: developer excellence built on Claude/GPT embedded in the editor. Code completions, in-editor chat, and codebase-aware context are deeply embedded in how developers write code, not just how they ask questions. No standalone API call can replicate that level of workflow integration.

  • Perplexity: search synthesis with verifiable citations. Solves a concrete problem that ChatGPT only partially addresses: the need for sourced, up-to-date answers with traceable references.

  • Harvey (legal): vertical depth in specific legal workflows. The domain is sufficiently complex that specialisation genuinely adds value over a generic LLM.

Products that did not:

  • Generic writing assistants without differentiation: the same system prompt, no proprietary data, no workflow integration.

  • "ChatGPT for X" without real verticalisation: X is just the name in the title, not a functional specialisation.

  • Products that depended on a base model limitation that disappeared in the next LLM version.

Real PMF signals vs misleading ones

Not PMF:

  • Viral traction at launch: the retention curve converging to zero at 30 days is traction, not PMF. AI products generate significant curiosity at launch that rarely translates into retained usage.

  • Frequent use during free trial: free trial funds exploratory use. The true signals are paid conversion rate and post-conversion retention.

  • Feature requests of "make it like ChatGPT": if users request features the base model already has, the product has no real differentiation.

Real PMF signals:

  • Low churn despite available alternatives.

  • Spontaneous recommendation without incentive.

  • Growing use intensity: more sessions, longer, more integrated in user flow over time.

  • Free to paid conversion above 10%: real willingness to pay, not just curiosity.

Iteration and measurement

The pace of iteration in AI products can and should be fast, but it must be guided by metrics, not intuition:

  • Weekly retention cohorts: detect quality regressions. A new model version or prompt change that hurts retention shows up within a week if cohorts are tracked consistently.

  • Top use cases: identify which problem the product solves for the segment with the highest retention. Check which use case actually retains: it may not be the one the team originally designed for.

  • Instrumented friction points: where users drop off, where they reformulate prompts, what responses they reject. These signal where the product falls short of user intent, which is where to focus iteration.

  • Interviews with active users: quantitative data tells you what is happening; interviews tell you why. The two are not substitutes for each other.

The build-measure-learn cycle remains the same; what changes is that the ability to build is faster. With LLMs, an MVP can ship in days. But that also means competitors can replicate it just as fast. The advantage is no longer in building first; it is in learning faster from real use.

Conclusion

PMF for LLM-powered products retains classic fundamentals but operates in different conditions. Retention, NPS, and revenue expansion metrics still predict success. What is added: monitoring LLM-specific quality metrics, avoiding the thin wrapper trap, and building moats that are a consequence of accumulated use: proprietary data, workflow integration, network effects.

The winning combination for a serious AI startup in 2024 and beyond is vertical depth + workflow integration + proprietary data flywheel. The base model is a commodity; what is built on it, and the data accumulated through use, is not.

Frequently asked questions

How do I know whether my AI product is just a thin wrapper over the API?

There are four signs. COGS is essentially a markup on the API, so gross margin is low. You accumulate no proprietary data that improves the experience over time, and there is no workflow integration: the user gets the same result by going directly to Claude.ai or ChatGPT. And the LLM is 100% of the value, replicable by anyone with access to the same API: a thin wrapper can generate initial revenue, but it can be copied in days.

Does viral traction at launch count as product-market fit?

No. AI products generate significant curiosity at launch, but a retention curve that converges to zero at 30 days is traction, not PMF. Frequent use during a free trial does not count either, since the trial funds exploratory use: the true signals are free-to-paid conversion (above 10%) and post-conversion retention. The most reliable signal remains users who would refuse to stop using the product if you removed it, along with spontaneous recommendation without incentive.

Which metrics should I add to the classic ones when the product uses an LLM?

Four that conventional instruments do not cover. Task success rate, which requires explicit instrumentation or sampled evaluation because it is not passively observable. Prompt reformulation rate: if users reformulate more than once per session, the first response was not satisfactory. Cost per valuable outcome, in tokens and money, and perceived latency, where time to first token (TTFT) matters more than total latency in streaming products.