UX Research · Step 6 · 7 min read
Writing insights that change decisions
A finding says what you saw; an insight says why it matters to a decision. The difference between data, findings, insights and recommendations, a template for one insight with a worked example, how to show evidence and confidence, and how to write recommendations someone will own.
"Users find the reminders annoying" is true, and it will change nothing. Everyone nods, someone says "good to know", and the meeting moves on. What changes a decision is a sentence that says why it happens, why it matters to something the team is trying to achieve, and where it points. That sentence is an insight, and writing it is the last, hardest step of synthesis.
What you'll be able to do after reading:
- Tell data, findings, insights and recommendations apart
- Write an insight statement with the five parts it needs
- Show the evidence and the confidence behind it
- Turn an insight into a recommendation with an owner
1. Data, findings, insights
- 1Data
One observation, not yet analysed.
P4: "I turned the reminders off." - 2Finding
A pattern across the data.
4 of 6 wait to ask; 2 turned reminders off - 3Insight
The pattern explained, in the context of a goal.
Asking feels rude, and the reminder speaks in their voice - 4Recommendation
What to do, and who decides.
Test an app-voiced, monthly reminder
NN/g separates three levels of analysis (NN/g, Data vs. Findings vs. Insights, 2023):
- Data is the unanalysed collection of observations: transcripts, notes, metrics, survey output. On its own it supports no conclusion.
- Findings describe patterns across the data. They're what Thematic analysis and affinity mapping give you.
- Insights are focused explanations of opportunities, based on research and business context. Interpreting findings in context yields insights.
NN/g's example is worth retelling. A redesigned booking site has an NPS of 40, ±10: that's a finding. Add the context: the old design scored 35, ±15, so the difference isn't significant; the redesign's goal was to cut support calls; usability testing showed people struggled with jargon and weak information scent. Together they make an insight: the redesign hasn't solved the problem it was for, and the reason is visible.
A recommendation is the step after: what to do about the insight, and who decides. Keep it separate. NN/g notes that insights aren't meant to be prescriptive; they narrow the design possibilities, which can then be tested.
2. The five parts of an insight
who do or struggle with what, because the reason, in their terms. This matters because the effect on a user or business goal, so the direction it points to.
Insight 1 of 3
People who pay shared bills wait weeks before asking to be paid back, because asking a friend for money feels rude, and our reminder speaks in their voice. This matters because late repayments are why groups stop using the app, so reminders should come from the app, on a date the group agreed, not from a person.
- 4 of 6 interviews: waiting to ask (P2, P3, P4, P5)
- 2 of 6 turned reminders off and said why (P4, P5)
- "It felt like the app was me, nagging Ana." P4
Six flat-sharers in one city; no couples. The link to groups leaving is from analytics, not from these interviews.
The reminder redesign. Next: the A/B test of the app-voiced, monthly reminder.
…make settling up happen without anyone having to ask?
Payments team, review on 14 April
An insight statement answers five questions, in one or two sentences:
- Who. A specific group, ideally a segment from your synthesis, not "users".
- What. What they do or struggle with, observed, not assumed.
- Why. The reason, in their terms. This is the part findings don't have, and the part that makes the insight useful: you can design for a reason.
- So what. The effect on a goal the team cares about: a user outcome, a business metric, a risk.
- Where it points. A direction, not a solution. A "How might we" question is a good way to say it; Problem statements and "How might we" explains how to write one.
Then, around the sentence, the rows that make it credible: the evidence (how many participants, which quotes, which other sources), the limits (who wasn't in the study), the decision it informs and an owner.
Test each insight with three questions. Would anyone disagree with it? If not, it's probably too obvious to matter. Does it explain something? If it only describes, it's still a finding. Could a team act on it next sprint? If not, it's too abstract, or it's missing the "so what".
The Insight statement card has the five parts as steps, a second worked example and a template you can copy.
3. Evidence and confidence
Every insight carries some of the researcher in it. NN/g notes that, because the researcher designs, runs, analyses and interprets the study, insights inevitably contain some bias, and the remedy is triangulation: several data sources, several analysis approaches, several people analysing (NN/g, 2023).
A simple confidence scale helps readers weigh what they read:
| Confidence | Typical evidence |
|---|---|
| High | Seen across most participants, said and observed, and backed by another source such as analytics or a survey |
| Medium | A clear pattern in several participants, from one method |
| Low | One or two participants, or told but not observed; worth following up |
Write the level next to every insight. It lets you share early, low-confidence insights without anyone mistaking them for proof, and it tells the team where the next study should go.
When methods disagree, don't pick the one you like. NN/g's advice for contradictory findings is to first check each study: who took part, what tasks they did, how realistic the setting was, and whether the numbers were significant. Often the contradiction explains itself, for example novices in one study and experts in the other (NN/g, Interpreting Contradictory UX Research Findings, 2019).
4. From insight to recommendation
A recommendation turns the direction into an action someone owns. A good one is:
- Specific enough to start: "Test a reminder that speaks as the app, on a monthly date the group agrees", not "improve reminders".
- Proportionate to the confidence: high confidence can justify building; low confidence justifies a cheap test. The A/B test planned in Analytics and A/B tests for designers is a medium-confidence insight turned into an experiment.
- Owned and dated: a team, and a date when you'll look at it again.
NN/g makes a practical point about certainty: concerns about statistical significance and validity are common, but it's wiser to make some recommendation that could help than none at all (NN/g, 2023). Say how sure you are, and recommend anyway.
5. Fewer, sharper insights
A study rarely needs more than three to five insights. Thirty "insights" are usually thirty findings, and a team can act on none of them. To cut down:
- Merge findings that share a reason. "People wait to ask" and "people turn reminders off" are one insight with one why.
- Rank by the size of the effect on the goal, and by confidence.
- Move the rest to an appendix or the research repository, still tagged and searchable. Nothing is lost; it just isn't in the way.
Then write each insight as if it were the only thing a busy stakeholder will read. Often it is. The next step, Research reports people read, is about packaging them.
6. Common mistakes
- Findings presented as insights: "62% of users skip onboarding".
- Insights with no why.
- Insights that are really solutions: "Users need a dashboard".
- No evidence or confidence, so a single quote looks like a law.
- "Users" instead of a specific group.
- Thirty insights, all equally important.
- Recommendations with no owner, so everyone agrees and nobody acts.
7. Vocabulary
| Term | Meaning |
|---|---|
| Data | Unanalysed observations: notes, transcripts, metrics, answers |
| Finding | A pattern across the data, without context |
| Insight | A finding explained, in the context of a goal, pointing somewhere |
| Recommendation | What to do about an insight, and who decides |
| Triangulation | Checking an insight with several sources, methods or analysts |
| Confidence | How strongly the evidence supports an insight |
| How might we | A question that turns an insight into a design direction |
8. Try it (no tools needed)
- Take this finding: "Half the participants printed the confirmation page."
- Invent two different whys for it. Notice how each points to a completely different design.
- Write the insight statement for one of them, with the five parts, and add what evidence you'd need to choose between the two.
Reflection questions: In the last research presentation you saw, which slide changed a decision? Was it a finding, an insight or a recommendation?
9. If you already practise
- Put the confidence level and the participant count next to every insight in every report.
- Run a short "insight review" with one other person before sharing: for each insight, would anyone disagree, does it explain, could a team act on it?
- Write insights in the research repository as their own entries, linked to their evidence, so later studies can strengthen or weaken them.
- Track what happened to each recommendation a quarter later. It's the start of Showing the impact of research.
Checklist
Learn more
- Sara Ramaswamy, Data vs. Findings vs. Insights: The Differences Explained (NN/g, 2023)
- Kate Moran, Interpreting Contradictory UX Research Findings (NN/g, 2019)
- NN/g, Triangulation: Get Better Research Results by Using Multiple UX Methods
Books, courses and videos to go deeper
- It's Our ResearchTomer SharonBook · Morgan Kaufmann, 2012 (Amazon)
- Just Enough ResearchErika HallBook · A Book Apart, 2nd edition, 2019 (Amazon)
Book links go to Amazon. As an Amazon Associate I earn from qualifying purchases, at no extra cost to you.
Related
Problem statements and "How might we" · Thematic analysis · Personas and journeys from data · Research reports people read · Present findings in ten minutes