Back to blogColorful sticky notes fastened to a board with pushpinsGEO Analytics

    AI review summaries: cohorts, dates and aggregate ratings

    2026-11-03·7 min read

    A summary can reproduce a genuine opinion and still misdescribe the collection. It may treat old comments as recent experience, merge product variants or present an aggregate rating as though it had read every text behind the score.

    This guide is for reputation and product research teams reviewing an AI review summary they have already observed. The deliverable is a record of the collection examined, its dates and the opinions the summary retains or loses. It is not a strategy for acquiring reviews or evidence that the available comments represent all customers.

    The citation support audit checks whether a passage supports a claim. The additional question here is which population that passage allows you to summarize and where the available evidence ends.

    Define the cohort before interpreting the summary

    Use cohort to mean the collection of reviews included under explicit rules. Specify the entity, product or service, variant, market, language and period. Separate an opinion about the seller from one about the item: sharing a page does not make them the same experience.

    Preserve the summary, the question or task that produced it, the observed surface and the capture date. If you do not know which documents the system received, say so. Visible citations do not establish that you have identified all its input material.

    Define the working unit as well. It may be an identifiable published review, not necessarily a person or purchase. Do not infer unique buyers from aliases or join profiles across platforms because their names look similar.

    An unknown variant must remain unknown. You can include it in a broader review with that limitation visible, but do not assign it to your target variant to complete the collection.

    Preserve four dates with different meanings

    Record the experience date if the source provides it, the review publication or edit date, the capture date and the declared aggregate date. They are not interchangeable. If only publication is dated, do not present that as the purchase date.

    A summary captured today may refer to reviews from any period. Preserve expressions such as “recently” or “since the update” and check whether they refer to an identifiable interval. Without it, the summary's recency remains unresolved, even if the prose uses the present tense.

    Keep an authorized copy or reference to the version reviewed. If a review is edited, preserve the change as a version rather than another independent opinion. Its later disappearance does not establish that the original comment was false.

    When comparing rounds, separate new reviews from changes to existing ones. If you cannot reconstruct the earlier collection, state that you are comparing two captures, not a controlled evolution of the same cohort.

    Separate ratings, reviews and accessible texts

    In Schema.org AggregateRating, ratingCount counts ratings and reviewCount counts reviews. Neither establishes how many texts you could read. Record both published values separately from the number of texts accessible in your audit.

    Do not calculate reading coverage by dividing your texts by a counter belonging to another entity, filter or period. If the complete eligible collection is unknown, report the number of texts examined and leave total coverage unknown.

    Preserve the score together with its scale. Google's review documentation explains scale limits and the bestRating and worstRating fields. A number without its endpoints is not enough to compare it with another.

    Do not average scores across platforms without knowing what each scale measures, which observations it includes and which weighting it applies. Multiplying an average by its counter does not reconstruct the original data when rounding, weights or calculation rules are unknown.

    Document which part of the collection you could read

    Record the access route, filters, visible ordering, pagination and capture limits. A “most helpful” selection and a chronological selection follow different rules. Keep them separate even when some reviews appear in both.

    FieldWhat to record
    Entity and variantIdentifier, scope and unresolved cases
    PeriodTime rule and dates actually available
    SelectionSource, filters, ordering and pages consulted
    ObservationsIdentified reviews and versions, without duplicate captures
    AccessFull texts, snippets, blocked material and pending cases
    AggregateValue, scale, counter and date declared by the source

    A search snippet is not a complete review. A summary generated by another platform is not a collection of original opinions either. You can examine these as secondary sources while keeping that distinction visible.

    Deduplicate repeated captures of the same review using reliable identifiers. If you suspect a comment was republished elsewhere but cannot establish it, retain the possible dependence. Do not invent a shared identity or claim independence.

    Check themes, conditions and disagreements

    Within the available texts, record the themes expressed and the passage that identifies each one. Preserve favorable, unfavorable, mixed or ambiguous judgments by theme. One review can support one dimension and criticize another; an overall label should not erase that difference.

    Look for conditions lost in the summary: variant, use, country, date or service circumstances. Check opinions that disagree with the described trend as well. A synthesis need not reproduce every comment, but it should not turn a divided collection into unanimous agreement.

    Decide what “majority” means before counting. Distinguish a majority of all eligible reviews, of accessible texts and of reviewers mentioning a theme. These have different denominators. Failure to mention a theme is not a favorable opinion about it.

    If you count theme presence, one review can appear in several categories. Frequencies need not add up to one hundred percent. Supply counts, denominators and classification rules; if the collection is partial, do not extrapolate its distribution to all buyers.

    Check the summary without attributing hidden inputs

    Link each summary claim to the texts you can inspect. Distinguish support within the reviewed cohort, disagreement under comparable conditions and unavailable evidence. Failing to find a comment in your capture does not establish that it never existed.

    Check statements about scores separately. If the summary cites a published average, verify entity, scale, counter and date. Do not require the average of your partial texts to match that aggregate: they may describe different collections.

    When the system claims to summarize only a collection you supplied, preserve that file and its identifiers. You can then assess fidelity to the supplied material. In a commercial application without that control, limit the verdict to observed sources and state what remains unknown.

    The Perplexity Shopping guide covers preparing product pages and reviews for discovery. This audit does not establish that a platform used a particular review or that correcting a detail will change its recommendation.

    Compare periods without mixing composition changes

    Before interpreting a difference, check whether the product, version, source or selection rule changed. One round of recent texts and another of highlighted opinions cannot attribute the difference to improved service.

    Keep a view of each collection and another of the comparable subset, if one exists. Explain what is excluded and why. Do not remove negative reviews from one round to match its size to the other.

    If new variants or markets enter the collection, show them separately before building a total. An aggregate can move because composition changes even when you do not observe the same movement within each group. Without comparable sizes and rules, report the descriptive difference and avoid a causal explanation.

    The closing date is part of the report. Later corrections and captures should produce a new version with a stated reason rather than silently replacing evidence from the earlier close.

    Deliver a record with limits and verifiable corrections

    Deliver the original summary, the cohort record, permitted sources and a discrepancy log. Each issue should identify the statement, the collection it describes, the evidence reviewed, the lost condition and the person responsible for checking it.

    Distinguish correcting your own data from disagreeing with a legitimate opinion. Do not rewrite other people's comments, invent testimonials or adjust scores to make the summary favorable. When requesting a factual attribution correction, preserve the specific detail and the evidence that contradicts it.

    The schema and structured data guide covers implementation. Consistent markup does not certify that a collection is representative or reveal which texts an assistant read. This guide does not establish that Mentio includes a review-auditing feature either.

    A useful conclusion states what could be checked: which entity, texts and dates, and which discrepancies remain open. If the complete population is not observable, that limitation should accompany the summary rather than remain hidden in a working note.

    Frequently asked questions

    What is the cohort in a review audit? It is the collection of reviews included under entity, variant, source and period rules. It must be distinguished from accessible texts and from all customers of the business.

    Do ratingCount and reviewCount show how many texts the AI read? No. They are published counters of ratings and reviews, not an assistant reading log. The number of texts actually checked must be documented separately.

    Can I compare two scores on different scales? Not directly. Preserve the endpoints of each scale and check the population, period and calculation method before justifying a comparison or transformation.

    Must a summary include every negative opinion? It need not reproduce every opinion. It should preserve material disagreements and conditions without turning a partial selection into consensus among all customers.

    What can I conclude if I can only see some reviews? You can assess the summary against the texts observed. You cannot certify complete coverage or population representativeness without knowing that collection and how it was selected.

    Want to know if AI mentions your brand?

    Discover your visibility in ChatGPT, Claude and Gemini in minutes.

    Related articles