Product-Market Fit for AI Startups

The Thesis

Product-market fit for AI startups is harder to find than it looks — because the technology creates false positives at every stage. Novelty drives signups. Demos convert skeptics. The first generation of output from a good model feels like magic, and magic makes people come back once or twice. But magic is not retention, and retention is not PMF.

The trap is that your top-of-funnel metrics look healthy even when you have nothing. High signups, decent D1 retention, enthusiastic Twitter threads — none of it means users are integrating your product into how they actually work. And in AI, “how they actually work” is the only thing that matters.

The core insight: AI PMF is not when people are impressed by your product. It is when people are dependent on it.

Why AI PMF Is Harder to Measure

Traditional B2B PMF signals are imperfect but reasonably reliable. If users are paying, using the product weekly, expanding seats, and getting upset when the product goes down — you probably have something real. The feedback loops are slow but legible.

AI products scramble those feedback loops in three ways.

The novelty effect. Large language models produce outputs that feel genuinely impressive to people who have never interacted with one. A first-time user who asks your product to draft an email, summarize a document, or explain a concept walks away thinking “that was incredible.” That impression is real. It does not predict whether they will use your product in six weeks.

The novelty effect means your Day 1 retention can be excellent not because you have PMF, but because everyone wants to see what the thing does. Day 7 and Day 30 retention are the numbers that separate novelty from value.

Demo-driven signups. AI products have unusually high conversion from demos because the demo is almost always impressive. A well-structured demo with good prompts and polished outputs will get a “yes” from buyers who have no particular problem your product solves. They liked the demo. They signed up. Six weeks later they haven’t touched it. This is not a sales problem. It is a PMF problem disguised as a sales win.

The “try it once” problem. For most knowledge work tasks, trying an AI tool once produces something useful. A user asks your product to write a first draft of a legal brief, generate a market analysis, or review a code diff. They get output that they can use. They do not come back next week, because the task was one-time, not recurring. Your product helped them, genuinely — but that is not the same as being in their workflow.

These three dynamics combine to produce a specific failure mode: a product with impressive demo-to-signup conversion, decent D1 retention, and collapsing D30. If you are seeing that shape in your cohorts, you do not have PMF. You have an impressive demo with a novelty hangover.

What Real AI PMF Actually Looks Like

The clearest proxy for AI PMF I am aware of is a direct adaptation of the Sean Ellis test: what percentage of your active users would be “very disappointed” if your product went away?

For a general SaaS product, Ellis found 40% as the benchmark. For an AI product, I would set the bar higher — not because the number matters precisely, but because in AI the “very disappointed” threshold is doing more work. It is distinguishing users who find the product useful from users who are genuinely dependent on it.

The difference in practice:

• Useful but not dependent: “I use it when I remember to, it saves me some time”
• Dependent: “I have changed how I work because of this tool and going back would meaningfully slow me down”

The second group is who you are looking for. They are harder to acquire but they are the only ones who drive retention, expansion, and referrals that compound.

Concrete signals of real AI PMF:

Users have changed their workflow, not just added a tool to it. This is the most important signal. A developer who ran tests manually and now uses your AI to write test cases first has restructured their process. A lawyer who used to spend four hours on due diligence and now spends 45 minutes with your product has restructured their process. Workflow restructuring is dependency. Tool addition is not.

Retention past 30 days holds above baseline. Your 30-day retained cohort should look meaningfully different from your general user base — higher usage frequency, more integrations enabled, more context provided to the model. If your 30-day retained users look the same as D1 users, you don’t have a core of power users yet. You have delayed churn.

Users are angry when you break something. Support tickets that express frustration rather than disappointment are a good sign. “I tried this and it didn’t work” is a disappointed user. “This is broken and I needed it today” is a dependent user.

Users are expanding into adjacent use cases without prompting. A product that has PMF gets used in ways the team never expected. Users find the seams of what your product can do and push into them. If your product is being used exactly and only for the use cases you designed for, you may have a product that works but not one that has found its full surface area yet.

The Retention Trap

AI products have structurally high top-of-funnel. The category gets attention, the demos work, and there is a baseline of curiosity that drives signups independent of product quality. This is a gift and a trap simultaneously.

The trap: high top-of-funnel makes it easy to avoid confronting bad retention. If you are getting 500 signups a week from content and paid, and your D30 retention is 8%, you are losing 460 of those 500 users in a month. But the inbox keeps filling up and the number looks like growth.

Most AI products in the 2023–2025 wave lived in this trap. Enormous top-of-funnel from the AI buzz, mediocre-to-bad retention hidden underneath it, and a business that looked healthy until the buzz subsided and the cohorts became impossible to ignore.

How to get out of the retention trap:

Fix activation before acquisition. The most common retention problem in AI products is that users never actually experience the core value. They sign up, play with the product in a sandboxed way, and leave before they do the one thing that would make them come back. Your activation flow should be designed around getting users to the specific action that predicts retention — not around impressing them in the first five minutes.

Narrow your ICP until retention holds. If your overall D30 retention is 12% but users in a specific job function or company size have D30 retention of 35%, the path forward is obvious. Stop trying to retain the broad market. Double down on the segment where retention is working, understand why, and build the product specifically for them.

Build for recurring problems, not one-time tasks. The easiest category of AI product to build is also the hardest to retain: task-specific tools that solve a well-defined problem the user has occasionally. You need to find the recurring workflow problem, not the one-time task. If your product helps users do something weekly or daily, retention is achievable. If it helps them do something twice a year, you have a feature, not a product.

Metrics That Actually Signal PMF

Vanity metrics in AI: signups, MAU (when driven by novelty), demo-to-trial conversion, press mentions.

Metrics that signal real PMF:

Workflow integration rate. What percentage of users have taken an action that indicates integration into existing tooling — API key setup, Slack integration enabled, IDE plugin installed, browser extension active? Users who integrate your product into their stack are significantly more likely to retain. This metric also filters for users with a real use case, since casual explorers rarely bother.

Sessions per week among 30-day retained users. Not total sessions. Sessions among users who have already been retained for 30 days. If this number is high (3+/week for a knowledge work tool), you have users for whom the product is in their regular rotation. If it is below 1 per week, you have users who remember you exist but have not made you habitual.

Feature expansion rate. Across cohorts, are users expanding into more features over time? Users who start with one capability and grow into three or four are integrating deeply. Users who stay on one feature for months either have a very specific use case (fine) if retention is good, or are not finding value beyond the initial hook (bad).

NPS among power users, segmented. Overall NPS for an AI product is almost meaningless because the population is so heterogeneous. NPS among users with 30+ sessions in their first 30 days is a leading indicator of whether your core value proposition resonates with people who have actually used the product. If this is below 30, you have a product problem. If it is above 50, you have something to build on.

Model-Dependent PMF vs. Product PMF

This is one of the most important distinctions in AI product building right now, and it is largely underdiscussed.

Model-dependent PMF: Users like your product because the underlying model (GPT-5.5, Claude Fable 5, Gemini 3.5) is impressive. If you replaced your product with a thin wrapper around the same API, your users’ experience would be roughly the same.

Product PMF: Users like what you built with the model. The model is infrastructure. Your prompting, UX, data integration, context management, and workflow design are what create the value.

The reason this matters: model-dependent PMF is not durable. As models commoditize and competitors can access the same underlying capabilities, your product becomes interchangeable. The moat is not “we use a good model.” The moat is “we have built something on top of models that our users cannot easily replicate elsewhere.”

Testing whether you have product PMF or model PMF: what would happen if OpenAI or Anthropic shipped a version of your core feature in their consumer product? If your users would immediately migrate, you have model-dependent PMF. If they would stay because of your integrations, your data layer, your workflow design, or your domain-specific tuning — you have product PMF.

Cursor is a good example of product PMF. The underlying models are the same ones developers can access through other tools. Cursor’s retention is driven by the editor integration, the codebase indexing, and the interaction patterns they’ve built around the model — not the model itself. When OpenAI releases a new flagship model, Cursor users stay because they’re buying the product, not the model.

Finding Your Wedge

PMF in AI is almost always found in a specific wedge: the particular job where AI is not incrementally better, but categorically better. 10x, not 10%.

Incrementally better does not produce dependency. If your product saves users 20% of the time on a task, they will use it when it is convenient. If your product makes a previously painful or impossible task tractable, they restructure their work around it.

The pattern across products that found real PMF:

Cursor: AI-assisted coding was incrementally better in IDE plugins for years. Cursor found the wedge in codebase-aware editing — the ability to write code with context about the entire project, not just the open file. For complex refactors and unfamiliar codebases, this is categorically better, not just incrementally better.

Harvey: Legal AI had a lot of incremental tools. Harvey found the wedge in complex contract analysis at scale — the specific task where the quality gap between AI and manual review was large enough, and the cost of manual review was high enough, that switching made obvious economic sense. Big Law firms started using Harvey not because it was slightly more efficient but because it let associates handle work that would previously have required senior time.

Sierra: Customer service AI was incremented to death with chatbots that annoyed users for a decade. Sierra found the wedge in enterprise-grade, on-brand conversational AI with genuine resolution capability — the difference between a bot that deflects and a system that actually resolves. The wedge was not “AI chatbot” but “the first chatbot your users won’t hate.”

Products that had buzz but not PMF tend to share a pattern: they were impressive demonstrations of what LLMs can do, without a specific wedge where the value was categorical rather than incremental. Summarization tools, generic writing assistants, and “AI for X” products without a specific hard problem to solve fell into this category.

How to Know When You Have It

PMF is not a milestone you hit. It is a phase you enter, and the entry is gradual enough that founders often argue about whether they are there yet.

The clearest indicator I have seen: the sales cycle inverts. Early-stage products without PMF are push sales — the team is convincing users that the product is valuable. Products with PMF are pull sales — users are coming in with a specific problem they need solved and asking whether your product solves it. When your users show up with the problem already articulated and are using your sales conversation to evaluate fit rather than discover value, you are in PMF territory.

The second indicator: churn becomes explainable. When you do not have PMF, churn is diffuse — users leave for a hundred different reasons and none of them add up to a clear signal. When you have PMF, churn is concentrated. Users who leave do so for specific, identifiable reasons: they outgrew the product, their company changed tools, they had a specific use case your product did not support. The churn is learnable, not random.

If you are building an AI product and your retention curves look like every other AI product — great top-of-funnel, D1 retention around 40–50%, D30 retention under 15% — you are not in PMF. You have work to do. The work is not more features. It is finding the specific users, jobs, and workflows where your product is not incrementally better but categorically necessary.

Find that. Everything else follows.