Synthesize research with evidence

Begin with consent, permitted use, and data minimization. Interview notes may contain names, contact details, health information, commercial secrets, or candid comments about colleagues. Remove identifiers that are not required for analysis. Use only approved tools and storage, preserve access controls, and respect retention or deletion commitments made to participants.

Create a source register. Give each item a stable identifier such as INT-04, SURVEY-Q7, or TICKET-1821. Record method, date range, participant segment, sample size, researcher, and material limitations. If the source is incomplete or auto-transcribed, label it.

Do not combine sources as if they were equivalent. A support ticket documents an incident, an interview reveals reported experience, behavioral analytics records an event, and a survey estimates self-reported patterns under a sampling method. Each can inform a decision, but each supports different claims.

Ask the assistant to create evidence units before writing themes. Each unit should include:

• source identifier;
• exact quote or faithful observation;
• participant or segment label when permitted;
• context and task;
• interpreted need or friction;
• confidence in the interpretation;
• alternative interpretation;
• researcher note or follow-up question.

This two-stage process makes the transformation inspectable. Review the extraction against the source, especially negations, numbers, comparisons, and emotionally charged statements. A fluent theme built on faulty extraction remains faulty.

Use quotations carefully. Models may smooth grammar or merge nearby passages while presenting the result as a quote. Require exact text for quotation marks. Use paraphrases without quotation marks and retain a source pointer.

A theme is an analytical grouping, not a vote count generated by intuition. State the inclusion rule: what evidence belongs, what does not, and why. Name the underlying task or mechanism rather than repeating a broad sentiment.

For example, “Import setup lacks recoverable error guidance” is more useful than “Onboarding is confusing.” It can include unsupported format failures and unclear remediation while excluding difficulty locating the import button if that has a different cause.

For each theme, capture:

• supporting source IDs;
• contradicting or disconfirming evidence;
• segments and contexts represented;
• frequency within this dataset, not the population;
• severity or consequence;
• confidence and limitations;
• open research questions.

Ask AI specifically to search for counterexamples and minority views. Compression naturally favors repeated language and can bury an important but rare accessibility or safety issue.

Research synthesis often becomes accidental prioritization. A frequently mentioned inconvenience may receive more attention than a rare blocker. Keep at least three dimensions separate:

Frequency in the sample: How often did this evidence appear in the collected material?
Severity: How strongly did it prevent or harm the user’s goal?
Potential reach: How many relevant users may encounter it, based on separate evidence?

Five interviews cannot establish population reach by themselves. Product analytics may estimate affected sessions; operational data may show escalations; additional research may test the pattern. Ask the model to label unsupported extrapolations rather than turn sample counts into market facts.

Stronger product understanding often comes from comparing methods. Interview participants may report that setup is easy while event data shows repeated retries. Tickets may overrepresent users who contact support. Analytics may show abandonment without explaining why.

Create a triangulation matrix with a row for each candidate finding and columns for interviews, surveys, analytics, support data, and domain review. Record support, contradiction, absence, and method limitations. AI can populate a draft from labeled sources, but a researcher must verify that each cell reflects the evidence.

Contradiction is productive. It may reveal segments, contexts, measurement errors, changing behavior, or a poorly framed theme. Do not instruct the assistant to “resolve” disagreement by selecting the neatest explanation. Ask for competing hypotheses and the next evidence that would distinguish them.

A synthesis should help a decision without pretending to make it. A useful finding card includes:

Finding: A precise statement bounded to observed users and context.
Evidence: Source IDs and representative exact excerpts.
Variation: Contradictions, segments, and exceptions.
Implication: Product questions or opportunities, clearly labeled as interpretation.
Confidence: Strength and limits of the evidence.
Next step: Decision, experiment, measurement, or further research.

Keep findings separate from recommendations. “Four of six observed administrators could not recover from an unsupported file error” is a finding. “Build an AI import guide” is one possible response. The team should compare it with better error messaging, format detection, documentation, and format support.

Run two directions of review. From output to source, verify every material claim. From source to output, check whether severe issues, contradictions, and minority experiences disappeared. Look for fabricated quotes, mixed participants, wrong quantities, inflated certainty, and unsupported causal language.

Invite the researcher who collected the data to review contextual interpretation. Where possible, have another reviewer inspect a sample independently. Record model, prompt, source set, and revision date so future readers know what produced the artifact.