Check an AI theme summary by linking each theme to response IDs, recording counterexamples and preserving minority views. Judge the summary against the underlying responses, not its fluency. A plausible label can still combine unrelated issues or exaggerate how widely an opinion was expressed.
Define the unit you are counting
Decide whether you are counting respondents, comments or mentions. One comment may contain two themes, but that does not make it two respondents. Avoid describing a theme as “most common” unless the counting method and denominator support that wording.
Write a short code definition for each theme. For example, “schedule access” concerns timing, while “connection quality” concerns technical interruption. Keeping them separate prevents a broad “access problems” label from concealing different next steps.
Build a synthetic evidence matrix
Use this explicitly hypothetical dataset about an online workshop:
- R1: “The evening session suited me.”
- R2: “The evening session suited me, but audio kept breaking.”
- R3: “Morning sessions would be easier.”
- R4: “The explanations were clear.”
- R5: “The explanations were clear, but I wanted a Marathi glossary.”
- R6: “I could not follow two technical terms.”
Represent the theme-to-evidence matrix as records:
- Evening timing works: support R1 and R2; counterexample R3; scope two of six responses, not a majority.
- Explanation clarity: support R4 and R5; counterexample R6; boundary R6 concerns two terms, not necessarily the entire workshop.
- Language support: direct evidence R5; related but not equivalent evidence R6; minority view worth retaining.
- Connection quality: evidence R2 only; no basis for claiming widespread technical failure.
Test the summary against exclusions
A hypothetical summary saying “Participants preferred evenings and found the content accessible” overstates the timing evidence and suppresses qualification.
A more faithful version is: “Two responses favoured evening timing, while one preferred mornings. Two praised clarity; requests concerning a Marathi glossary and technical terms indicate possible language-support needs. One response reported audio disruption.”
Ask a second reviewer to assign the six comments independently. Investigate disagreements about theme boundaries before turning counts into recommendations. Do not infer reasons that respondents did not provide.
Choose the next analytical step
Perplexity’s supplied upload documentation describes document-based tasks including summarising, analysing code and translating. It states uploaded files are retained for 30 days, or seven days for Enterprise Pro, then automatically removed. That does not establish thematic accuracy.
Stop acceptance if a central theme lacks traceable evidence or omits a material counterexample. Revise the code definitions and recount. For access comparisons, check the AI tools catalog for access format and saved terms before buying; marketplace account, code or token offers are not proof of official subscription equivalence.
Sources and scope
Source material checked 29 October 2026; features/policies may change. Perplexity question and upload guidance supports the document-task and retention statements. The synthetic matrix and review method are editorial advice, not a tested accuracy claim.