Why Most PMF Research Is Built on Team Recall, Not Evidence

What team recall actually produces

Every early-stage SaaS team has a working theory of their market. They know who their best customer is. They know what job the product does. They know which features drive retention and which ones are just nice-to-haves. They have heard it from customers directly — in demos, in support tickets, in customer success calls.

The problem is that "heard it from customers" and "measured it systematically" are two different things, and most teams treat them as equivalent.

Memory is not a neutral recording of customer conversations. It is a selective reconstruction that favours the dramatic over the mundane, the recent over the historical, and the voices that were most confident or most frequently encountered. When a founding team sits down to build a JTBD framework or define an ICP, they are drawing on this reconstructed memory — which means they are drawing on a biased, compressed, and partially invented account of what customers actually said.

This is not a critique of any particular team. It is a structural property of how human memory works, and it has predictable consequences for the research outputs that teams build from it.

How large is the gap?

A concrete illustration: a product team built a 27-job JTBD framework through a standard recall-based workshop process — collective team memory, customer quote harvesting, and synthesis into a prioritised list. The team was confident in the output. The framework passed internal review.

When the same product's 60 recorded sales calls were systematically coded against the same JTBD categories, the analysis found more than 85 jobs — more than three times the team's recall-based estimate. The recall process had not missed fringe edge cases. It had missed entire categories of jobs that came up repeatedly in the calls but had not registered as "important" in the team's memory.

The priority ranking gap was larger. The team's second-highest-priority feature — the one they most frequently cited as a retention driver — appeared in 43% of calls when systematically counted. The team's recalled estimate was 88%. That is not a rounding error. That is a feature prioritisation decision built on a frequency belief that was twice the actual rate.

A second common pattern: teams often have a confident belief about who their primary user is. When this is validated against actual login and usage data, the segment the team identifies as primary frequently accounts for a much smaller share of active users than believed. In one analysis, a user persona that was described as accounting for 38–52% of users in the team's ICP definition accounted for 17% of actual product logins. The segment the team thought was secondary was primary in the data.

These are not unusual findings. They are the expected output of comparing recall-based research against evidence-based research. The gap varies by team and product, but it is consistently present and consistently in the same direction: the team overestimates the frequency of jobs and features they believe are most important, and underestimates everything that did not make it into the team's dominant narrative.

PMF signals that are not PMF signals

Beyond recall-based research, there is a second category of PMF evidence problem: signals that feel like PMF confirmation but are not measuring what teams think they are measuring.

High NPS from early adopters

Early adopter NPS is a self-selecting signal. The people who respond to NPS surveys in a product's first year are disproportionately the people who already have a strong positive relationship with the product. They found it, adopted it through friction, and stayed — which means they are the least representative sample of the broader ICP you need to capture to scale. A high early-adopter NPS tells you the product works for the most motivated subset of users. It does not tell you whether it works for the median customer you will need to acquire in year two.

Strong engagement from power users

Power user engagement metrics — session length, feature breadth, login frequency for the top decile — do not predict cohort-level retention. The behaviour of the users who engage most intensely with a product is not a reliable guide to what happens when you expand into a broader ICP segment. Power users tolerate more friction, discover more features independently, and self-solve more problems than the median user. Building for power user engagement often means building against median user retention.

Positive customer interviews

Customers in research interviews are polite. They tell you the product is useful, the team is responsive, and they are happy with the value. Then they churn. The interview context creates social pressure toward positive framing that does not reflect the internal calculus the customer is making about whether to renew. Positive qualitative feedback is weakly correlated with retention at the cohort level — and essentially uncorrelated with expansion.

What real PMF evidence looks like

PMF evidence is cohort-based, not aggregate. It connects a specific behaviour at a specific point in the user journey to a measurable outcome downstream. The four most reliable forms:

Cohort retention curves with a behavioural split

A retention curve shows what percentage of users return at 30, 60, and 90 days. A retention curve split by activation behaviour answers a different question: do users who complete activation event X retain at a materially higher rate than those who do not? If yes, you have identified a behaviour that is worth building around. If the retention curves are flat regardless of activation, the activation event you defined is not predicting the outcome you care about.

Job frequency from systematically coded sales calls

Recording sales calls and transcribing them is not enough. Systematic coding means defining a set of job and pain categories in advance, then counting how often each category appears unprompted across a call set large enough to produce stable frequency estimates — typically 40 to 60 calls minimum. The output is a frequency ranking that can be compared directly against the team's recalled priorities. The discrepancies are the research finding.

Feature adoption by declared ICP segment

If the team has defined an ICP — a specific segment they believe is the best-fit customer — the question is whether that segment actually uses the features built for them at the expected rate. If the ICP segment's feature adoption is lower than non-ICP users, the segment definition and the product build are misaligned. This is a PMF problem, not an onboarding problem.

Revenue expansion signals

Retained customers who naturally expand — who upgrade, add seats, or increase usage without a sales intervention — are the clearest PMF signal available. Expansion means the product is creating enough value that the customer spontaneously invests more. Flat retained customers have found an equilibrium where the product is worth keeping but not worth growing. Declining retained customers are churning slowly. The distribution across these three groups is a direct read on PMF health at the cohort level.

Why this matters more at certain stages

The gap between recalled and evidenced PMF research is a persistent problem, but its consequences change with product stage. At pre-seed, building on recall is reasonable — you do not have enough data to do otherwise, and speed matters more than precision. The JTBD framework and ICP definition from a founding team workshop are good enough to guide the first 12 months of product decisions.

The problem appears at Series A and beyond, when the team is making resource allocation decisions that assume the recall-based research is still accurate. By the time a product has 60+ sales calls recorded and 12 months of retention data, the team has the raw material for evidence-based research. But most teams continue to use the original recall-based framework because updating it is uncomfortable — it means finding out that the priorities the team has been executing against are based on inflated frequency estimates.

The investment case for PMF validation is not that the original research was wrong from the start. It is that the original research was a best-available approximation, and the data now available is substantially better. The teams that act on that data make different roadmap decisions, different hiring decisions, and different messaging decisions — and those decisions compound.