How to build an empathy map from real participants

A typical empathy map is built in a workshop. The team divides into four corners labeled Says, Thinks, Does, and Feels, fills sticky notes with what they imagine the customer would do in each, photographs the wall, and treats the result as research. Most of the notes were written by people who have not spoken to a customer this quarter. The map gets hung in a Notion page next to the persona deck and gradually becomes wallpaper, referenced in arguments by people who no longer remember whether any specific cell came from a real participant or from the senior designer's intuition about how a marketing manager probably feels on a Tuesday.

This is a working guide on how to build an empathy map that informs a real decision: scope it to one persona doing one job, fill the four quadrants with verbatim from recorded participants, mark the contradictions between what the customer says and what the customer does, and keep the artifact standing rather than treating it as a workshop deliverable. The piece sits inside the wider voice user research guide and pairs naturally with the playbooks on user personas and customer journey maps.

What an empathy map is

An empathy map is a four-quadrant artifact that captures what a specific user segment says, thinks, does, and feels in the context of a specific moment in their experience. The four quadrants force the team to separate what is observable (what the user says out loud, what the user does with their hands) from what is inferred (what the user appears to be thinking, what the user appears to be feeling). The separation is the point: a useful empathy map exposes the gap between the polite stated answer and the underlying behavior or emotion that contradicts it.

The method was introduced by Dave Gray in the 2010 book Gamestorming and refined in his 2017 Updated Empathy Map Canvas, which added a "Pain" and "Gain" footer and a "Goal" header to anchor the quadrants to a specific job the user is trying to accomplish. The cleanest modern primer is Nielsen Norman Group's empathy mapping article, which distinguishes the one-user map (built from a single interview, scoped to one moment) from the aggregated map (synthesized across a segment, scoped to a job) and notes that the aggregated version is the one most product teams actually need.

The artifact is small on purpose. A persona answers "who is this segment over time?" A journey map answers "what arc do they move through?" An empathy map answers a narrower question: "in this specific moment, what is the user saying, thinking, doing, and feeling, and where do those four things contradict each other?" The contradiction is where the research signal lives.

Why most empathy maps fail

Three failure modes recur across teams that drew an empathy map and then quietly stopped using it. All three are structural. Working harder inside the wrong shape does not fix any of them.

Filled from team imagination, not participant talk

The most common failure. The workshop runs, the wall fills with sticky notes, and the notes describe what the team thinks the user is probably saying, thinking, doing, and feeling. The Says quadrant collects paraphrased imagined quotes. The Thinks quadrant collects the team's hypotheses about cognition. The Feels quadrant collects the senior designer's projection of what the user emotion arc probably looks like. None of those cells trace back to a participant, a transcript, or a timestamp. The map looks like research output because it has the shape of one. It is documentation of organizational consensus about a user the team has not recently met.

Drawn at the wrong altitude

The map is scoped to "the customer", which is to say nobody. Or it is scoped to "the persona during onboarding", which is still broad enough that the cells flatten into generic statements that would describe any user in any product. An empathy map that tries to capture the entire arc of a customer's relationship with a product produces cells too abstract to argue from. The job of the map is to capture one moment in high resolution; treating it as a summary of the whole journey produces a worse version of a journey map.

Treated as a workshop output, not a standing instrument

The map gets drawn once, off a single round of research, often during a quarterly planning offsite. New evidence shows up over the following months (a churn interview, a support ticket cluster, a new feature that shifts the activation behavior), and none of it lands on the map. By the time the next planning cycle reaches for the artifact, the team is making decisions against a snapshot of a user from two quarters ago. The snapshot was honest the day it was taken. It is not honest anymore.

How to build an empathy map, step by step

Six steps. The order is opinionated. Most teams start at step three (filling the quadrants) without doing steps one and two, which is the structural reason the quadrants end up filled with guesses. Done in the order below, the quadrants are an output of the evidence, not a workshop activity.

01 · Pick one persona, one moment, one job

Scope is everything. The empathy map that tries to describe a persona across their whole relationship with the product collapses into platitudes. The empathy map that describes the persona in one specific moment, attempting one specific job, holds its shape.

The persona side is straightforward if the team has done persona work grounded in real interviews. Pick the persona closest to the decision the team is making this quarter. The moment side is where most teams under-scope. "During onboarding" is not a moment. "The first time the user tries to invite a teammate and the invite fails silently" is a moment. The narrower the moment, the more useful the four quadrants. The job statement reuses the framing from jobs to be done interviews when the team has done switch-interview work: it names the progress the user is trying to make, not the feature they are trying to use.

One persona, one moment, one job, one map. If the team needs more than one map, build more.

02 · Recruit on current behavior, not stated identity

Recruiting is where empathy maps get distorted before the first quadrant is filled. The wrong recruit is the friendly customer, the internal champion, and the LinkedIn-network respondent. The right recruit is the user whose current behavior demonstrates the moment the map is trying to capture, whether or not they already use the product.

The screener filters on observable behavior. "Tell me about the last time you tried to add a teammate to a tool you were already using" is a screener. "Are you interested in collaboration features" is not. The first returns participants who have a specific recent episode to recall. The second returns participants willing to be polite about the category. Eight to twelve participants per moment is enough to see the quadrants stabilize; the full operational treatment is in how to recruit user research participants.

For the empathy map specifically, recruit a few participants who completed the moment successfully and a few who abandoned it. The contrast is what makes the contradictions in step five visible.

03 · Capture the four quadrants in the participant's own words

Once participants are recruited, the prompts for each quadrant are built to elicit the kind of evidence that quadrant needs. Generic prompts produce generic answers. Quadrant-specific prompts produce evidence the team can argue from.

• Says is captured in voice. Voice transcripts preserve the exact phrasing, the hedge, the qualifier, and the half-sentence the participant abandons mid-thought. A typed answer to "what did you say to your colleague after the invite failed" gets the cleaned-up paraphrase. A voice answer gets the actual phrasing: the hesitation, the "I, uh, I think I told her to just try a different tool", the tone. The full case for the modality difference is in what we hear when we stop asking people to write.
• Thinks is captured indirectly, through prompts that ask the participant to narrate their reasoning at the moment. "What were you trying to figure out right then?" gets closer to the actual cognitive state than "what did you think." The prompt asks for reconstruction, not introspection.
• Does is captured through walkthrough prompts. "Walk me through exactly what you did, step by step, after the invite failed" returns the observable behavior. Resist the temptation to ask "what did you do" without the "walk me through" frame; the latter returns the summarized version that papers over the actual sequence.
• Feels is captured through prompt + sentiment signal. The participant's stated emotion is one input; the sentiment of the transcript (tone, pace, the words used) is a second. The two inputs sometimes agree and sometimes diverge, and the divergence is itself a finding.

Participants pick the input mode per question (voice, text, choice, rating) and the system fills the quadrant from whatever modality each participant chose. Voice carries more weight on Says and Feels where tone and verbatim matter; text covers Thinks and Does where the participant may want to write a longer, more structured narrative. The case for letting the participant pick is in voice vs text surveys.

Probing depth is set per quadrant, not globally. Shallow depth on the closed-ended prompts where one clarifier is enough. Medium depth on Says and Does, where a small chain of probes turns a vague first answer into a specific sequence the team can mark up. Expert depth on Thinks and Feels, where the AI keeps probing until it has the same context a senior researcher would dig out: contradiction, scope, who, when, how, the alternative considered, the reason the participant chose not to do something. The participant retains the right to skip on every probe. The full pattern for adaptive follow-ups is in AI follow-up questions for user research.

04 · Fill the quadrants with verbatim, not paraphrase

Synthesis is where the empathy map either stays honest or becomes another consensus document. The discipline is verbatim before paraphrase.

Each cell in each quadrant carries at least one direct quote from a recorded participant, with attribution to a participant ID and a timestamp. The team's gloss sits next to the quote, not in place of it. "Participant felt confused" is a paraphrase. "I genuinely thought I had broken something" is the verbatim. The first describes the team's interpretation; the second is the evidence behind it. A future team member should be able to click any cell on the map, read the source transcript, and verify the gloss against the participant's actual words.

Recurring quotes cluster into themes. Two participants describing the same friction become one theme with two pieces of evidence and the team's gloss attached. The synthesis pass is the same procedure we describe in how to synthesize user research, applied per quadrant rather than across the whole transcript: code Says into verbal patterns, Thinks into reasoning patterns, Does into action patterns, Feels into emotion patterns. The qualitative pass surfaces the themes; the team's job is to keep the verbatim attached as the cell summarizes.

05 · Mark the contradictions between quadrants

The whole point of the four-quadrant shape is to surface the places where what the user says, thinks, does, and feels do not line up. A user who says the onboarding was "fine" but spent eleven minutes on a single screen is not telling the team that onboarding was fine. A user who thinks they understand the export feature but does not actually export anything is not understanding the export feature. A user who feels frustrated during a moment they describe as smooth is reporting frustration about something else nearby.

"It was fine, I just figured I'd add her another way later. I mean, I never did, but at the time it felt fine."

The quote above is the kind of evidence the contradiction pass is built to surface. The Says quadrant gets "it was fine." The Does quadrant gets "never added the teammate." The Feels quadrant gets "frustrated enough to silently give up." All three came from the same participant in the same answer, and the gap between them is the finding. An empathy map that does not mark contradictions hides this kind of signal in the smooth surface of the cells.

Three contradictions to look for explicitly:

• Says vs Does. What the user reports out loud against what their behavior demonstrates. The most common source of false-positive feature validation.
• Thinks vs Feels. What the user appears to be reasoning about against the emotion that shows up in tone. The most common source of dropped retention warnings.
• Does vs Feels. Effortful behavior described in flat emotional terms. The most common source of underestimated friction.

Each contradiction earns a callout on the map, anchored to the verbatim that produced it. The contradictions are the cells that drive the next research question, not the cells that already feel resolved.

06 · Share internally and keep the map standing

An empathy map that lives in one product manager's Figma file does not change a decision. An empathy map that has been read by engineering, design, support, customer success, and the executive sponsor is one that informs the next sprint.

The async version of the internal pass is a study link shared inside the company, the draft map attached. Engineering reads the Does quadrant and flags the behavior that breaks the build assumption. Support reads Says and surfaces the verbatim pattern that matches historical tickets. Sales reads Thinks and Feels and adds the objection the customer-facing call missed. Each stakeholder answers in their own modality on their own time, voice for design and support where tone matters, text for legal and finance where structure matters, rating for the executive pulse, choice for the tradeoff prompts. The synthesized return is a cross-functional pressure test of the map before it informs a decision. The stakeholder-interview playbook covers the internal-feedback pattern in detail.

After the internal pass, the same study links that produced the map stay open as continuous feedback placements: an in-product link on the surface the moment lives on, a churn-flow link that captures the abandonment version of the same moment, a post-activation prompt that captures the success version. Every new response is tagged against the existing quadrant clusters and routed back to the team. The map updates as the evidence shifts. The customer-feedback-loop playbook covers the continuous-capture pattern at the level of the whole product, not just one moment.

Where the empathy map pairs with other artifacts

The empathy map is rarely the whole thinking. Three artifacts it interlocks with cleanly:

Personas set the "who" the map is about. One map per persona per moment. A map without a persona drifts toward an aggregate that describes no one. A persona without an empathy map is a behavioral profile without a moment of high-resolution evidence to anchor it to.

Customer journey maps set the "where in time" the moment lives. The journey map is read horizontally across stages. The empathy map zooms into one stage and reads vertically through the four quadrants. The two artifacts are designed to interlock: the journey map flags the moments of truth and pain points, and each one is a candidate scope for an empathy map.

Opportunity solution trees sit downstream. A contradiction surfaced on an empathy map (the user says onboarding is fine but never invites a teammate) becomes an opportunity candidate on the tree, then connects to candidate solutions and assumption tests. The empathy map answers "where is the friction at this moment." The tree answers "what to do about it."

When the empathy map is the wrong tool

Three cases where the empathy map is the wrong artifact, and reaching for it anyway produces a polished version of the wrong answer.

The team has no participants for the moment. Empathy maps are evidence-driven. Without recruited participants who experienced the moment recently, the map regresses to a workshop output and the four quadrants fill from imagination. A team in that position should run the recruiting and the interviews first, or scope the map to a moment they do have participants for.

The question is comparative across features, not exploratory within a moment. A team prioritizing a backlog of feature candidates against a known segment does not need an empathy map; it needs usage data and a weighted scoring exercise. The empathy map answers questions about one moment in high resolution. It does not answer "which of these six features should ship next."

The team needs cross-functional alignment without external research. Sometimes a team reaches for an empathy map as a way to align internally on a shared mental model of a user. That is a meeting, not a research artifact. Run the meeting; do not call the output an empathy map. Better: spin up a small internal study, share the link in engineering, design, support, finance, and exec channels, and collect a synthesized view of each function's current mental model of the user before the artifact gets locked. Name it correctly. A stakeholder-aligned hypothesis is a useful thing; mis-labeling it as an empathy map sets up the team to confuse internal consensus for external evidence.

FAQ

What is an empathy map?

An empathy map is a four-quadrant artifact that captures what a specific user segment says, thinks, does, and feels at one specific moment in their experience. The method was introduced by Dave Gray in the 2010 book Gamestorming and refined in his 2017 Updated Empathy Map Canvas, which added Pain and Gain rows and a Goal header to tie the quadrants to a specific job the user is trying to accomplish. The useful version of the map traces every cell back to a verbatim quote from a recorded participant, with attribution to a participant ID. The non-useful version is a wall of sticky notes filled with the team's intuitions about a user the team has not recently met.

What are the four quadrants of an empathy map?

The classic empathy map has four quadrants: Says (what the user says out loud, captured as verbatim), Thinks (what the user appears to be reasoning about, captured through narration prompts), Does (the observable behavior, captured through walkthrough prompts), and Feels (the emotional state, captured through stated emotion plus sentiment signal in tone). Dave Gray's Updated Empathy Map Canvas adds a Pain row and a Gain row below the quadrants and a Goal at the top, anchoring the four quadrants to a specific job. The quadrants are useful because they force the team to separate observable evidence from inferred state and surface the contradictions between the two.

What is the difference between an empathy map and a persona?

A persona is a behavioral archetype of a user segment over time; an empathy map is a high-resolution view of one moment in that segment's experience. A persona answers "who is this segment, and what do they do consistently?" An empathy map answers "in this specific moment, what is the user saying, thinking, doing, and feeling, and where do those things contradict?" The two are designed to interlock: one persona, multiple empathy maps, each scoped to a moment the persona moves through. Both should be evidence-driven, both should anchor each claim to a participant ID, and both should be kept current rather than treated as a one-time deliverable.

When should I use an empathy map vs a journey map?

A customer journey map is read horizontally across stages: it shows the arc of an experience over time. An empathy map is read vertically through four quadrants: it zooms into one stage and shows what is happening for the user at high resolution. The two artifacts pair: build the journey map to surface the moments of truth and the pain points, then build an empathy map for each moment that earned a callout. A team that needs both will usually build the journey map first to scope which moments deserve the high-resolution treatment.

How often should you update an empathy map?

Treat the map as a standing instrument fed by continuous capture, not a workshop deliverable. The cells refresh as new evidence lands from the in-product feedback link on the moment, the churn flow capturing the abandonment version, and the post-activation prompt capturing the success version. A standing weekly review by the team that owns the map is enough to keep the quadrants current. A map that goes more than four weeks without an update is no longer a working artifact; it is a snapshot. The continuous-capture pattern is covered in detail in the customer feedback loop playbook and the cadence problem in continuous discovery interviews.

Who should build an empathy map?

Whoever is going to maintain it. A map built by a consultant or a lone researcher and handed to the team usually dies on day one because the team has no ownership of it. A map built collaboratively by the product trio (product, design, engineering) plus the research function, anchored to real participant transcripts, is the version that survives. For the internal pressure test before the map is locked, share the in-progress map inside the company on a Talkful link and collect a synthesized view of each function's mental model of the moment. Engineering flags the behavior assumption, support adds the verbatim from historical tickets, sales surfaces the objection, exec calibrates the priority. The map that ships after that pass is one the rest of the team will recognize.

---

An empathy map built well is a working instrument: four quadrants scoped to one moment, each cell anchored to verbatim from a recorded participant, contradictions surfaced explicitly, the artifact kept current by continuous capture across the surfaces the moment lives on. An empathy map built badly is a wall of sticky notes that becomes Notion wallpaper. Talkful is built for the first shape: one study link per moment, voice or text or choice or rating from the participant's pick, smart follow-ups that probe the polite first answers into the honest second ones, and a synthesis engine that streams quadrant-tagged themes and quotes back to the team as the responses land, ready for product teams to ship from or for the agents you build with to act on. The wider voice user research guide covers where the practice sits inside a continuous product-research rhythm.