How do I refine my ICP based on early sales conversations?
Compare accounts that progressed, stalled and bought but failed to reach value. Identify one observable ICP criterion that the evidence challenges. Review competing explanations and counterexamples with the founder, then approve a versioned change. Apply it to the next prospect list and record what new evidence would make you reverse the decision.
What should an ICP revision actually change?
An ICP revision should change an observable targeting or qualification rule and explain why. A more detailed persona document is not enough if the next prospect list and sales questions remain identical.
For a post-seed startup, the useful question is which buyer situation can lead from a first customer to another customer who reaches value without exceptional rescue work. Early sales conversations supply part of that answer. Onboarding, product use and delivery supply the rest.
Start with the current rule, even if it is provisional. Write down which accounts it includes, what evidence supports it and what remains unknown. If no usable profile exists, begin with defining an ICP for an early-stage startup. This process starts when you have a rule to challenge.
Which accounts should you compare?
Compare progression, non-progression and post-sale outcomes at account level. A friendly conversation, a signed agreement and a satisfied customer are different observations and should not be combined into one success label.
a16z’s ICP framework recommends learning from successful, lost and churned customers. It also distinguishes customer or product mismatch from other reasons an opportunity is lost. This is practitioner guidance, not a controlled study establishing a particular cohort size.
| Account group | Evidence to inspect | Mistake to avoid |
|---|---|---|
| Progressed to a concrete next step | What the buyer agreed to do and why | Treating praise as commitment |
| Stalled or declined | Recent context, constraints and stated reason | Assuming every loss means poor fit |
| Bought and reached value | Outcome, product use and delivery conditions | Assuming purchase itself proves success |
| Bought but struggled | Missing capability, onboarding and support burden | Hiding delivery problems behind a won stage |
| Useful counterexample | Why the account differs from the current pattern | Removing it because it complicates the story |
Count accounts separately from conversations and quotations. Several contacts repeating one internal concern do not create several independent customer examples. Preserve which contacts and events belong to the same account.
How do you inspect the conversation evidence?
Use specific past events and observable actions to support an ICP change. General opinions about an industry or predictions about future buying are weaker grounds for an operational rule.
Product Talk’s story-based interview method recommends asking about concrete experiences instead of what people think they will do. Applied here, ask what happened, what the buyer tried and what changed next. Do not ask the buyer to validate your preferred segment description.
- Retrieve the original note, recording or reply behind the observation.
- Identify the buyer’s situation, current workaround and recent trigger.
- Separate the buyer’s words from the seller’s interpretation.
- Record the actual next action and whether it happened.
- Add the product and delivery conditions that could explain the outcome.
- Mark missing or disputed evidence explicitly.
Keep a warm introduction, unusually intensive founder help, a product change or exceptional commercial concession visible. These may explain progression that would not recur in a normal acquisition motion. You do not need to discard the account, but you do need to avoid attributing everything to its company characteristics.
The supporting capture fields belong in a startup ICP evidence record. Do not turn the revision meeting into a search through an unstructured transcript archive.
Which criterion is the evidence challenging?
Change one criterion whose meaning another person can apply. Prefer an observable workflow or constraint over an attractive description that cannot be checked before or during a conversation.
The following are specification patterns, not findings about a particular market.
| Vague criterion | More inspectable formulation | Evidence still needed |
|---|---|---|
| Innovative company | A named team is actively replacing a specified process | Recent project or buyer confirmation |
| Large enough to need us | The relevant workflow occurs and has an owner | Workflow description and responsibility |
| Loves automation | A repeated manual handoff is creating a recognized problem | Specific incident and current workaround |
| Has buying intent | A relevant project has a stated next decision | Buyer-confirmed action and participants |
| Good customer | Reaches the defined outcome under supportable conditions | Post-sale outcome and delivery record |
Do not infer the presence of a workflow from company size alone. Size can remain a useful search proxy, but label it as a proxy. Decide whether a condition must be observable before outreach or can only be confirmed in discovery.
Keep stable fit, current readiness and ability to implement separate. A suitable company may have no active project. A motivated buyer may need a capability you cannot deliver. Those situations call for different actions rather than a single good-or-bad score.
How do you challenge the proposed change?
Before approving a new criterion, ask what evidence would support the opposite interpretation. A revision is more defensible when the founder can explain both its supporting accounts and its strongest counterexample.
Product Talk’s discussion of confirmation bias recommends seeking contrary information, making disagreement easier and considering the opposite. Use those practices in the review rather than asking colleagues to endorse a polished summary.
Ask whether the same outcome occurred without the proposed characteristic. Ask whether apparently suitable accounts failed because of the product or delivery. Ask who was never contacted under the old rule and is therefore absent from the evidence. A list created by the original assumption cannot independently prove that assumption.
Invite the person responsible for onboarding or support. They may see that easy-to-close accounts are difficult to serve. Let sales explain genuine timing or execution problems without forcing every loss into an ICP conclusion. The decision should serve both a viable business and customers the product can help.
What should the versioned decision contain?
A versioned ICP decision preserves the old rule, the proposed replacement, the supporting evidence and the reason to revisit it. It should be possible to reconstruct why the next campaign targeted different accounts.
Use a short decision record:
- previous criterion and profile version;
- observations that challenge it, linked to accounts and sources;
- proposed observable replacement or narrower interpretation;
- counterexamples and competing explanations;
- founder or decision owner and decision date;
- accounts entering or leaving the next test;
- expected observation and a condition that would reverse the choice.
This is a proposed working method. The W3C PROV Primer supplies the broader provenance principle: a revision can be represented as a new entity related to the previous version, with responsibility and time recorded. A startup can keep that trail in a shared document or CRM without implementing the full standard.
Do not overwrite the previous profile and then reinterpret old deals as though the new criterion existed all along. Preserve what was known at the time. Separate a correction to bad data from a genuine change in the team’s decision.
How does the revision reach the next prospect list?
Apply the changed rule to a reviewable next batch and inspect the accounts it includes and excludes. The revision becomes useful when it changes action while preserving room for informed exceptions.
Have the list owner attach this change record to the next batch: previous and approved profile versions, changed criterion, decision owner and date. For each affected account, record the status below with the observation and source behind it.
| Account status | Evidence to record | Action on the next list |
|---|---|---|
| Include | Observation supporting the revised criterion, source and relevant conditions | Include in the test; retain the reason and profile version |
| Exclude | Observation showing the criterion is not met, source and any counterexample | Remove from this test; keep the exclusion reason reviewable |
| Unknown | Missing or disputed fact and the question needed to resolve it | Research or ask in discovery; do not treat absence as a failed criterion |
Keep the expected observation and reversal condition from the decision record with the batch. Record any informed exception explicitly rather than silently changing the rule.
Update the relevant qualification prompt, CRM definition and call question together. Keep the profile version on the batch so later replies can be interpreted against the correct rule. Do not simultaneously change every message, offer and channel if you want to understand what the profile revision contributed.
If the evidence points toward a different urgent problem, make that a separate campaign decision. The comparison method is described in choosing the buyer pain for your first GTM campaign.
When should you revisit or refuse a revision?
Revisit the criterion when new evidence contradicts its explanation, delivery conditions change or the next batch fails to show the expected buyer situation. A calendar review can help, but it is not proof that the evidence is ready.
Do not revise solely because one large opportunity requests a roadmap change. Record that request and the trade-off. It may justify a separate product decision, but it does not automatically define the next repeatable market.
With very limited evidence, preserve uncertainty and choose a reversible test. There is no universal interview count or survey threshold that converts early B2B conversations into a proven ICP. If the notes are unreliable or no customer outcome is known, repair the evidence supply first. See also how narrow a seed-stage ICP should be for the tests that show when a revision has made the profile too wide or too tight.
AI can retrieve, cluster and compare material, but the founder should inspect sources and approve the operational rule. Before expanding acquisition, use the broader checks for turning traction into a repeatable GTM motion.
Turning the decision into a working GTM process
Piet Baudoin of Poldermarketing works remotely in Dutch and English with post-seed startups building go-to-market: a first customer, a second, then a repeatable process that reaches customers who are genuinely happy with the product. He can connect the founder’s ICP decision to positioning and visibility in Google and AI answers.
He also builds and runs AI automations with care, including lead follow-up, enrichment and qualification, CRM setup, reporting, content distribution and tool integrations. He can set up the CRM and read Search Console queries. Explore positioning and messaging or start with the free growth scan.
Questions people ask
How many sales calls are enough to change an ICP?
There is no universal number that makes an early ICP revision reliable. Inspect the independence, relevance and quality of the evidence, including accounts that contradict the proposed rule. With limited evidence, make a reversible targeting hypothesis and test it in a small next batch. Do not present a directional choice as proven market fit.
Should every lost deal be excluded from our ICP?
No. A suitable account can be lost because of timing, an implementation constraint, a missing product capability or your sales execution. Record the reason and its source before changing the profile. A loss can expose a product or process problem without showing that the buyer's underlying situation was wrong for the offer.
Can AI decide which customer characteristics matter?
AI can organize account evidence and suggest patterns, but the founder should review the sources, alternative explanations and excluded accounts. A repeated phrase is not necessarily a buying condition. Keep unknown values separate from failed criteria and require an explicit decision before a proposed pattern changes qualification or targeting.
What if a valuable customer falls outside the revised profile?
Keep the customer and inspect the counterexample. The profile guides the next acquisition test; it does not cancel existing commitments or invalidate delivered value. Record whether the account reveals a second use case, an exception or a flaw in the new criterion. Reopen the rule when the evidence warrants it.