ProjekinAja. Founder's Playbook Stage 2 · Build The Right Thing Lesson 2.6
● Stage 2 · Lesson 2.6

Testing the MVP — What Good Feedback Looks Like

Showing an MVP to customers produces mostly noise — here is the observation protocol that separates genuine usability signal from politeness.

⏱ ~22 min read 🎛 Interactive sorter 🔬 Signal vs. noise

🎯 By the end you'll understand

Why behavioral feedback (what the customer actually did) carries fundamentally different weight than stated feedback (what they said they would do), how to run a think-aloud protocol that produces usable signal, the three-tier taxonomy of feedback quality — feature requests, behavior changes, and unprompted referrals — and why the post-demo survey is structurally incapable of producing the signal you think it is. You will also sort 18 real-world feedback items from Forge's first pilot and calibrate your ability to distinguish signal from noise before it influences a product decision.

KEY INSIGHTCustomers are polite and that is your problem

An MVP session that ends with "this is really impressive, you should definitely keep going" has told you approximately nothing. The signal was drowned in social grace the moment the customer realized you built this yourself and care about what they think.

You shipped the MVP. You booked sessions with five pilot customers. You walked them through the product. They nodded, they smiled, they asked a few questions, and at the end three of them said it looked promising. You left feeling validated. You should not feel validated — you have collected social data, not product data. The distinction between these two is the entire subject of this lesson, and getting it wrong is one of the most expensive mistakes in startup execution.

🔑 The core move of MVP testing

Replace the question "do they like it?" with the question "what did they actually do, and what does that behavior tell you about the product's usability and value delivery?" The first question produces polite answers. The second produces evidence. Liking and using are not the same thing. A customer who "loves" your product but cannot complete the core workflow without help has told you the product is broken, not that it is good. Your observation protocol must be designed to surface the doing, not the saying.

CORE CONCEPTBehavioral feedback vs. stated feedback

The cleanest distinction in MVP testing is between what the customer does and what the customer says. Behavioral feedback is a record of actions — where they clicked, where they paused, where they asked for help, what they skipped, what they went back to, how long they spent on each step. Stated feedback is a record of opinions — "this is easy to use," "I'd want it to do X too," "this reminds me of [competitor]." Both are useful. They are not equally useful, and treating them as equivalent is how founders end up building feature lists driven by customer enthusiasm rather than customer friction.

Feedback type Example Signal weight Why the weight differs
Behavioral: task completion Customer completes the core defect-review workflow without assistance in under 4 minutes High — direct usability evidence They either did it or they did not. No amount of politeness changes the log.
Behavioral: hesitation & backtracking Customer pauses for 22 seconds on the threshold-calibration screen, re-reads labels, adjusts, then reverts High — specific friction signal Identifies exact locations in the workflow where comprehension breaks down.
Behavioral: unprompted feature use Customer opens the export panel without being shown it and downloads a CSV High — job-to-be-done signal Reveals real workflows the customer is trying to integrate your product into.
Stated: general approval "This is really clean, I can see our team using this." Low — social desirability Easy to say; costs nothing; does not predict actual adoption or continued use.
Stated: feature request "It would be great if you could add a mobile app." Low — wishlist item Feature requests are almost always stated in the context of the demo, not derived from actual workflow pain.
Stated: behavior change "I showed your alert digest to my QA lead and she's already asking when we can deploy this line-wide." Very high — organizational spread signal A change in the customer's actual behavior (sharing with colleagues) is behavioral evidence of value, regardless of what was said in the sentence.

The key structural insight: behavior is harder to fake than opinion. When a customer struggles to complete a task they were shown how to do, that is data. When a customer says "great product, looking forward to using it" and then never logs in again, that is also data — but it only becomes legible if you track usage after the session, not during it. The post-demo survey asking for a 1–5 rating is measuring the emotional state of the room, not the product's ability to deliver value in the customer's real environment.

💡 Separate the session from the follow-up

The two most informative data points from an MVP test often come 48–72 hours after the session. Did the customer log in again unprompted? Did they share the product with a colleague? Did they ask a question that implies they are already planning to use it for a real task? These behaviors are observed, not stated, and they are structurally immune to the social pressure that distorts in-session feedback. Build the follow-up into your testing protocol before you run the first session.

CORE CONCEPTThe think-aloud protocol

The think-aloud protocol is the most widely used usability research technique for a reason: it converts the opaque internal state of a user ("they looked confused, but I can't tell why") into observable, analyzable data. The instruction is simple and must be given verbatim before the session begins: "As you use this, please say out loud whatever you are thinking — what you notice, what you are trying to do, what confuses you, what you expected to happen. Don't worry about being polite or helpful. I want to know what is actually happening in your head, not whether the product is good."

That last sentence is important and most founders skip it. Without explicit permission to be critical, customers self-censor. The think-aloud protocol only produces signal if the customer has genuinely internalized that critical comments are more valuable than approving ones. In B2B contexts — where your customer is a professional who knows you need this to succeed — the social pressure to be supportive is intense. Countering it requires explicit framing, not just the general instruction to think aloud.

1

Frame the session as testing the product, not the customer

Say: "If something is hard to figure out, that means we designed it badly — it's a bug in the product, not a gap in your understanding." This reframes mistakes from customer failure to product failure. Customers are much more willing to articulate confusion when they are not implicitly being evaluated for how quickly they figure things out.

2

Give them a task, not a tour

Do not walk the customer through the product. Give them a task they would actually need to perform in their real workflow: "Imagine you just got an alert that Line 3 had an anomalous defect rate this morning. Walk me through how you'd figure out what happened." Then stop talking. Every word you say after handing them the task contaminates the data. The point is to observe them discovering the product, not to demonstrate it.

3

Record observations, not reactions

Your notes during the session should be timestamped logs of what the customer did: "T+1:40 — opened the wrong tab looking for threshold settings," "T+3:15 — asked unprompted if alerts can be routed to Slack," "T+4:50 — completed the defect review without asking for help." Not: "seemed a bit confused in the middle," "asked about Slack integration (good sign?)," "finished okay." The first set of notes is analyzable. The second is a mood board.

4

Ask the five-second follow-up, not the five-minute debrief

After each significant hesitation or error, ask: "What were you expecting to happen there?" Not: "Is this confusing?" The first question produces diagnostic information — what the customer's mental model predicted. The second produces a binary self-assessment that is usually distorted by social grace. The five-second follow-up is the most information-dense intervention available in a testing session; use it consistently and sparingly.

Real task no demo, no tour Observation think-aloud active Behavioral log actions, hesitations Verbal log expectations, confusion Analysis pattern across sessions behavioral log: higher weight verbal log: useful with caution
The think-aloud protocol produces two parallel logs. Behavioral observations carry higher evidentiary weight; verbal logs are useful for understanding expectations but must be interpreted knowing that social desirability pressure is always present.

CORE CONCEPTThe three-tier signal taxonomy

Not all feedback that reaches you after an MVP session is equally informative. The three-tier taxonomy organizes incoming feedback by the structural properties that determine its evidentiary weight — independently of how enthusiastically it was delivered.

▲

Tier 1 — Behavior changes (highest signal)

A behavior change is anything the customer actually does differently as a result of using the product — sharing it with a colleague, modifying an internal workflow to incorporate it, asking about integration with their existing stack, logging in again without prompting, referencing data from the product in their internal communications. These are observable actions that carry revealed-preference weight. The critical property: the customer bore a real cost (time, social capital, organizational friction) to take the action. That cost is evidence of perceived value independent of what they said in the session. Unprompted referrals — the highest expression of this tier — are the clearest early indicator that the product is creating genuine value, because referring colleagues to a product you are piloting involves real professional risk.

▶

Tier 2 — Usability friction observations (medium signal)

Specific, observable moments where the product failed to support the user's intended action: task abandonment, requests for help on a named feature, navigation errors that reveal a mismatch between the product's information architecture and the user's mental model. These observations are directionally actionable — they tell you exactly where to fix — but they carry medium signal weight because usability friction does not predict whether the product delivers enough value to be worth using once usability is fixed. A highly usable product with no value and a difficult-to-use product with high value are both problems; usability feedback only speaks to one dimension.

▼

Tier 3 — Feature requests and general approval (low signal)

Feature requests are the most frequently collected and most frequently over-weighted form of MVP feedback. The structural problem: feature requests are generated in the context of the demo, not derived from actual workflow pain. A customer who says "it would be great if this integrated with SAP" is responding to what they see in front of them. They may have an SAP integration problem in real life, or they may not — you cannot tell from the request itself. General approval ("this is impressive," "I could see us using this") suffers from the social desirability problem described above and carries almost no predictive weight for actual adoption. The presence of feature requests is weakly positively correlated with engagement (people only suggest features for products they find interesting) — but the specific feature requests themselves should not drive your roadmap.

✗ The feature request trap

The most dangerous outcome of an MVP session is a list of ten feature requests that all sound reasonable. The danger is not that the features are wrong — some of them may be right. The danger is that a list of feature requests can absorb twelve weeks of engineering without addressing any of the fundamental questions the MVP was supposed to answer: does this product deliver enough value in its current form that customers will pay for it and keep using it? Feature requests postpone that question indefinitely, which is why technically-minded founders find them so appealing. The list always grows; the core value question does not get answered by growing it.

Tier 3 — Feature requests & general approval low signal · very common · most often over-weighted Tier 2 — Usability friction medium signal · specific & actionable · fix-points only Tier 1 — Behavior changes & referrals high signal · rare · highest evidentiary weight
The three-tier signal taxonomy. Volume is inversely correlated with evidentiary weight — you will almost always receive more Tier 3 feedback than Tier 1. The mistake is optimizing for the feedback type you receive the most of rather than the feedback type that most accurately predicts whether the product creates real value.

SEE ITFeedback Quality Sorter — Forge's first pilot

⚙ Interactive · Feedback Quality Sorter

Below are 18 feedback items collected from Forge's first three-site pilot session. For each item, classify it as High Signal (Tier 1 — behavior change or revealing behavioral observation), Low Signal (Tier 2 — usability friction or specific but limited observation), or Misleading Signal (Tier 3 — feature requests, general approval, or stated hypothetical). After sorting each item, the widget reveals the expert classification, an explanation, and your accuracy breakdown by category.

Context: Forge ran three one-hour pilot sessions at a single tier-1 auto-parts manufacturing site. The customer contact is the Quality Engineering Manager (QEM). An operations analyst and a line supervisor were also present for parts of the session. Forge's MVP is a dashboard showing real-time defect alerts with a 24-hour alert digest and a basic threshold-calibration interface.

Unclassified feedback items — 18 remaining

Three patterns are worth watching for in your own accuracy results. First, founders almost universally under-classify behavior changes as high signal — items about what the customer actually did get classified as low signal because they seem mundane or ordinary. Second, founders over-classify feature requests as low signal rather than misleading — the logic being "well, at least they told me something specific." This conflates specificity with evidentiary weight. Third, founders are generally good at identifying usability friction (low signal) because it is the category that feels most immediately actionable. None of this means low-signal feedback is worthless — it means it should not drive resource allocation at the expense of the high-signal behavioral observations that are rarer and harder to notice in the moment.

FORGEWhat Forge's first pilot feedback actually showed

After completing the three-site pilot session described in the widget, Forge's raw notes looked like a success: three engaged sessions, multiple feature requests, enthusiastic verbal responses. Underneath that surface, the high-signal behavioral observations told a more specific story.

1

The QEM shared the alert digest with her maintenance lead before the session ended

This is the clearest Tier 1 signal from the entire pilot: an unprompted referral happening during the session itself, with a specific colleague and a specific purpose. The QEM was not asked to share the product; she did so because the alert digest contained information she needed to communicate anyway. This is behavioral evidence that the dashboard is producing output that has value in the customer's real workflow — independent of whether the session was designed to make it look good.

2

Two of three customers could not locate the threshold calibration screen without help

Tier 2 signal — specific, actionable friction on a named feature. The calibration screen is critical to Forge's value proposition (customers need to set alert sensitivity to their specific defect tolerances). If 67% of pilot customers cannot find it, the core workflow is broken at a navigation level. This does not tell you whether customers value the feature — it tells you that even if they value it, they cannot use it. This needs to be fixed before drawing any conclusions about calibration adoption.

3

All three customers asked about Slack integration, SAP export, and mobile alerts

Tier 3 — three feature requests stated in the context of the demo. These are correlated (asking for integrations is a sign of engagement with the product), but the specific requests should not be taken literally. The structural insight: customers asking about integrations are telling you they are already imagining the product in their workflow. The right follow-up question is "tell me about a specific time last month when you needed to communicate a defect alert to your maintenance team" — that question produces real workflow data. The Slack integration request does not.

🔑 Forge's go-signal from the pilot

The single highest-signal observation from Forge's pilot — the QEM sharing the alert digest with her maintenance lead unprompted — is worth more to Forge's product/solution fit assessment than all three feature requests combined. It demonstrates that the product is producing output that has real utility in the customer's operational context, that the customer is willing to act on it immediately, and that the value is clear enough to share without needing an explanation. The correct interpretation of the pilot: the alert digest has demonstrated early value evidence; the calibration workflow is broken and must be fixed before the dashboard can deliver on its core promise; the feature requests are engagement signals, not roadmap items.

Based on this pilot, Forge's next action is not to add Slack integration. It is to rebuild the navigation path to the calibration screen, run a follow-up session with the same QEM to test whether the fixed path resolves the friction, and then instrument the dashboard to track whether the three pilot customers log in again in the following two weeks without being prompted. Those two data points — calibration friction resolved, unprompted re-engagement rate — are the behavioral signal that supports or contradicts the pilot's initial positive impression.

IN PRACTICEThe post-demo survey — why 5-star ratings are almost meaningless

The post-demo survey is the standard closing ritual of most MVP testing sessions. The founder thanks the customer for their time, asks them to fill out a quick feedback form (usually 5–10 questions on a 1–5 scale), and uses the aggregate ratings to evaluate the session. This is a structurally broken feedback mechanism, and the reasons are worth understanding precisely — because it feels rigorous, looks like data, and produces almost no signal.

Three structural problems:

1

Social desirability bias is at its maximum during the survey

The customer just watched you demonstrate something you clearly care about. They are sitting in front of you, or on a video call with you visible. The survey is usually taken immediately after the session, when the social bond created by the hour together is at its strongest. The result: surveys taken immediately after a demo in a social setting systematically overstate satisfaction — a well-documented effect of social desirability bias (see e.g., Podsakoff et al., 2003, on common method bias in self-report data). The magnitude varies by context, but is large enough to make post-demo satisfaction scores unreliable as a primary signal. You cannot know which individual survey is inflated — you only know that the population of post-demo surveys is. The implication: aggregate survey scores from a small pilot cohort (3–5 customers) are nearly uninterpretable.

2

The survey measures the session, not the product

Customers rating "ease of use" on a 1–5 scale after a demo are rating the experience of being walked through a product by a founder who knows every corner of it. That is not the same as their experience using the product alone on a Tuesday morning with a real problem to solve. The usability score from a guided demo is measuring the founder's skill at giving demos, not the product's standalone usability. The only usability data that predicts real-world usage is from unguided task completion — which is what the think-aloud protocol produces.

3

Rating scales collapse information that matters

A customer who rates "overall product value" at 4/5 might mean: "this is valuable for the alert digest but completely useless for calibration," "I'd use this if my team had time to learn it but we don't right now," or "this is more relevant to my QA team than to me personally." A 4/5 cannot communicate any of these distinctions. The survey forces precision onto a response that is inherently multivariate. The thing you actually need to know — does this create enough value that you would pay for it and continue using it? — cannot be answered with a rating scale. It can only be answered with behavior.

💡 Replace the survey with one question and a two-week tracking period

Instead of a post-demo survey, ask one question at the end of the session: "Based on what you've seen today, is there someone else at your company you'd want to show this to?" This question is behaviorally predictive (if yes, it implies enough perceived value to make a referral recommendation — a real cost) and it does not require a rating scale. Then track, over the following two weeks: (1) did the customer log in again, (2) did the referred colleague appear in the system, (3) did the customer reach out unprompted with a question or request. These three data points carry more evidentiary weight than any post-demo survey score.

WEAKNESSWhere this bites

✗ The "everyone loved it" debrief

The post-pilot debrief where the founding team recaps the sessions often produces a false consensus around enthusiasm rather than a structured analysis of signal quality. Verbal recaps naturally favor memorable moments — and the most memorable moments in a customer session are usually the enthusiastic ones, not the hesitation at minute three on the calibration screen. The debrief ritual that actually produces actionable output is a structured review of the behavioral log, not a discussion of how the sessions felt. If your post-pilot decision about what to build next is primarily driven by the feature requests that came up in multiple sessions, you have run the debrief wrong.

✗ Conflating pilot engagement with product/solution fit

Pilot customers are not representative users. They agreed to a structured session — often because they have a relationship with the founder, because they are curious about AI applications in their industry, or because they were offered an incentive. The behavioral observations from a structured pilot session are more reliable than stated feedback, but they are still collected in an artificial context. The critical validation gate is not "did they engage well in the session?" — it is "did they log in again on their own, and did they use the product in a real operational context without support?" Until you have that data, the pilot tells you the product is interesting, not that it creates enough value to be purchased and retained.

✗ Running too many sessions before analyzing any of them

A common founder pattern: run five pilot sessions back-to-back in a week, then try to synthesize all the feedback at once. The problem is that early sessions almost always reveal critical usability issues that would have changed how later sessions were run. If sessions two and three would have revealed that customers cannot find the calibration screen, sessions four and five are collecting data from a broken product state. The better rhythm: run one session, analyze the behavioral log within 48 hours, fix the most critical friction point, then run the next session. Iteration speed within the pilot cohort is more valuable than parallel data collection from a broken baseline.

KEY INSIGHTHow it connects

  • Prototype Strategies — Before You Write Production Code (2.4) — the signal taxonomy in this lesson determines what you were trying to learn from the prototype in the first place. If your prototype was designed to test behavior change signal (will customers use this in their real workflow?), the think-aloud protocol and the post-session tracking period are the right evaluation mechanisms. If it was designed to test usability friction only, structured task completion is sufficient. The feedback methodology should be defined before the prototype is built, not after.
  • Building the MVP — Speed Over Polish (2.5) — the speed-over-polish principle in 2.5 is grounded in this lesson's insight: the most valuable feedback from an MVP is behavioral, and behavioral signal is generated by giving customers a real task with a real product, not by showing them a high-fidelity mockup. An MVP built for behavioral signal collection needs to be functional enough to observe real task completion — and no more polished than that threshold.
  • Next: The Pivot Decision — When and How to Change Course (2.7) — the feedback quality taxonomy from this lesson directly feeds the pivot decision framework. A pattern of high-signal behavioral evidence that the product creates value but low retention is a different pivot signal than a pattern of enthusiastic stated feedback with no behavioral follow-through. The signal quality you are able to recognize here determines the quality of the pivot decision you will be equipped to make there.

The discipline of distinguishing behavioral from stated feedback is not just a pilot-testing technique — it is an operating principle that applies whenever you are gathering evidence about what customers do with your product. In Stage 3, when you are setting prices and designing contracts, the customers who renew are the behavioral signal; the customers who said they would renew are the stated signal. In Stage 4, when you are analyzing the sales funnel, the customers who moved to the next stage are the behavioral signal; the customers who said the proposal looked promising are not. The taxonomy introduced here runs through every evidence-gathering activity for the rest of the course.

GLOSSARYKey terms introduced

Behavioral feedback
A record of what a customer actually did during or after an MVP session — task completion, hesitation points, navigation errors, post-session logins, referrals, and unprompted feature use. Carries higher evidentiary weight than stated feedback because behavior is harder to fabricate or distort under social pressure.
Stated feedback
A record of what a customer said about a product — opinions, ratings, feature requests, and hypothetical commitments. Systematically biased toward positivity in session contexts due to social desirability pressure. Useful as a supplement to behavioral observation; unreliable as a primary signal.
Think-aloud protocol
A usability research technique in which the participant is instructed to verbalize their thoughts continuously while using a product, converting internal cognitive states (confusion, expectation, frustration) into observable, analyzable data. Most informative when combined with a real task rather than a guided tour.
Social desirability bias
The tendency for survey and interview respondents to answer in ways they believe will be viewed favorably by the person asking, rather than in ways that accurately reflect their true opinions or intentions. Particularly severe in post-demo contexts where the founder is present and has an obvious emotional investment in the product.
Unprompted referral
An instance where a customer shares a product with a colleague or contact without being asked to do so. Carries very high signal weight because it involves real professional cost (the referrer's credibility is implicitly attached to the recommendation) and occurs outside the artificial pressure of the testing session.
Feature request
A stated suggestion from a customer for a capability not present in the current product. Low evidentiary weight because requests are generated in the context of the demo (where imagination is stimulated by what is visible) rather than derived from documented, observed workflow pain. Correlated with engagement but not reliably predictive of adoption or willingness to pay.
Usability friction
A specific, observable moment in a testing session where the product failed to support the user's intended action — navigation errors, task abandonment, requests for help on a named feature. Medium signal weight: actionable for fixing the product but does not predict whether the fixed version will be valuable enough to purchase and retain.
Behavioral tracking period
The 14–21 days following an MVP session during which the founder monitors product usage, referrals, and unprompted contact as behavioral signals of perceived value. The most information-dense phase of MVP validation; cannot be replaced by the session itself.

ACTIVE RECALLCheck your understanding

Answer in your own words first, then reveal.

1A founder runs three pilot sessions and collects the following feedback: two customers asked for a mobile app, all three said the product was "intuitive and well-designed," and one customer sent the alert digest to her maintenance lead during the session. Which item carries the most evidentiary weight and why?›

The unprompted referral during the session — the customer sharing the alert digest with her maintenance lead — carries the most evidentiary weight by a significant margin. It is a behavioral observation: an observable action taken by the customer at real cost (professional credibility attached to the referral) outside the explicit structure of the session. The mobile app requests are Tier 3 feature requests — stated in the context of the demo, generated by what is visible, with no demonstrated workflow pain behind them. The general approval ("intuitive and well-designed") is stated feedback under high social desirability pressure, measuring the quality of the demo experience rather than the product's standalone value. The referral is the only datum that would exist whether or not the customer knew they were being observed.

2What is the structural problem with post-demo surveys that makes their scores nearly uninterpretable? Is there any form of post-session survey that carries useful signal?›

Three structural problems make post-demo survey scores nearly uninterpretable: (1) Social desirability bias is at its maximum immediately after a session where the founder is present and emotionally invested — scores systematically overstate satisfaction. (2) The survey measures the session, not the product — guided demos by founders who know the product intimately produce usability impressions that do not generalize to unguided use. (3) Rating scales collapse multivariate information — a 4/5 on "overall value" cannot distinguish "valuable for X but broken for Y" from genuine broad satisfaction. The one question with useful signal: "Is there someone at your company you'd want to show this to?" — it is behaviorally predictive (referral intent is a real cost signal) and does not rely on a rating scale. But the most informative "survey" is the behavioral tracking period in the weeks after the session.

3In the think-aloud protocol, why is giving the customer a task more informative than giving them a tour, and what makes the five-second follow-up question ("what were you expecting to happen there?") more diagnostic than asking "is this confusing?"›

Task vs. tour: A tour sequences the product through the founder's mental model of how it should be understood. The customer learns the product structure that the founder intended. This produces no information about whether that structure matches the customer's actual mental model or workflow needs. A task sequences the product through the customer's mental model — they have to find things, interpret labels, and make navigation decisions using their own prior knowledge. The delta between what the customer does and what the product expected reveals exactly where the product's information architecture diverges from the user's model. This delta is the actionable product data; you cannot generate it with a tour. What were you expecting vs. is this confusing: "Is this confusing?" produces a binary self-assessment that most customers answer "no" (social desirability, and because they eventually figured it out). "What were you expecting to happen?" produces the customer's mental model prediction — a specific diagnostic that reveals why the confusion occurred and what the product would need to change to match the user's intuition. The first question asks for a verdict; the second produces a design specification.

4A founder receives ten feature requests from three pilot sessions. Eight of them seem reasonable and one sounds expensive. The founder's co-founder argues they should build the four most-requested features before running more pilots. What is wrong with this logic, and what should the founder do instead?›

The co-founder's logic contains two errors. First, feature requests do not tell you whether the core product creates enough value to be purchased and retained — they tell you what customers noticed during a demo and could imagine wanting. Building the four most-requested features before running more pilots postpones the fundamental question indefinitely: do customers use this product in their real workflow, log in unprompted, and eventually pay for it? Feature requests cannot answer that question; only behavioral tracking over time can. Second, the logic conflates stated enthusiasm (the requests) with validated need. The four most-requested features might be the four least important ones if the customer's actual workflow pain is somewhere the demo never surfaced. The correct decision: analyze the behavioral observations from the three sessions to identify the usability friction points (Tier 2 signal), fix the most critical one, run a follow-up session with the same cohort to verify the fix, and then track post-session logins and referrals for two weeks. If behavioral re-engagement is low, the product has a value problem that no feature list will solve. If re-engagement is high, the feature roadmap can be prioritized based on what patterns of use you actually observe rather than what customers said during the demo.

Next: the pivot decision — when behavior evidence tells you to change course →