ReviewProof ⚖

How does the AI summary of your reviews work?

By the Ingenious AI team · Published 14 November 2026 · Reviewed 14 November 2026

Run the free review audit See pricing · See a sample provenance report

Short answer

The summary is computed only from your verified reviews, by a deterministic extraction rather than a creative generator: it reports how many verified-experience reviews exist, their average rating, and the words those customers most often use (words longer than five letters, counted). It produces nothing when there are no verified reviews. And it is labelled as AI-generated wherever it is displayed — because a summary that hid its nature would break, inside your own product, the same honesty rule the FTC applies to AI-generated review text.

What is the AI review summary?

The AI review summary is a machine-generated digest of a business's verified review set: an extractive statistic (verified count, mean rating, most frequent substantive words) rather than a written testimonial. Its two design constraints are its meaning: the input set is verified reviews only, so UNLINKED and flagged reviews cannot leak their sentiment into your public summary; and its presentation is labelled, so a reader always knows they are reading a generated digest of evidence, not the evidence itself — which sits one click away in the widget and certificates.

Why extraction, not generation

A generative summary could write beautiful, plausible, untrue sentences about your customers ('guests loved the breakfast') assembled from patterns rather than from your file. An extractive summary cannot: it can only count what verified customers actually wrote, which bounds its claims to arithmetic the reader can audit against the displayed set. The top-words list is the signature of that honesty — sometimes slightly unglamorous ('parking' among a hotel's frequent words), always checkable. When a summary's vocabulary surprises you, the surprise is information: it is what your provable customers actually talk about, as opposed to what your marketing wishes they talked about, and more than one business has fixed a real operational theme it first noticed as an awkward word in its own summary.

A detail worth knowing before you start: ReviewProof analyses what you give it. The CSV importer reads reviewer, rating, title, body, date and source columns, plus a verified-transaction flag and a transaction reference when you have them. Reviews with a transaction link are marked verified; reviews without one are marked UNLINKED — kept in the audit, visible in the report, but excluded from the verified-experience widget, which only displays reviews that are both transaction-linked and in a kept or published state. That single rule is the difference between a widget that markets and a widget that proves.

Where the label does its work

The label 'AI-generated' beside the summary performs the same function as the disclosure label on reply drafts: it keeps the business honest at exactly the point where honesty is cheapest to skip. A visitor reading the widget sees the layers correctly sorted — reviews (human, verified, certificated), summary (machine, derived, labelled) — and can trust each layer at its proper weight. Businesses sometimes ask whether the label weakens the summary's marketing effect. The evidence of this whole product category runs the other way: readers burned by a decade of dressed-up content extend more trust to the surface that sorts its own claims than to one that blends them — and regulators extend more forbearance, which at the scale of an enforcement letter matters more.

After the first audit, the habit that matters is cadence. Re-import on a regular rhythm — monthly suits most locations — so new reviews enter the provenance record while the underlying transactions are still easy to match. Each import logs provenance events, moderation starts from kept (no suppression), and the certificate for any review shows its chain: imported, transaction-linked or not, analysed with its signals, moderated with its decision. A record built gradually like this is worth far more in a dispute than one assembled in a panic the week the dispute arrives.

Mistakes that undo the work

Editing the summary's wording to remove an awkward theme is falsifying a labelled statistic — fix the theme or let the summary stand. Showing the summary without its generated-content label, because the design looks cleaner, converts an honest digest into exactly the kind of undisclosed generated content the FTC rule targets. And computing anything public-facing from unverified rows to get a larger count breaks the input rule that gives the number its meaning.

The workflow is the same on every plan, and the free tier does the complete job for one location: import by CSV upload or paste (up to 500 reviews per import on Free), read the full risk analysis and report, open a provenance certificate for any review, and publish the verified-experience widget. Pro ($29/month, 3 locations, 5,000 reviews per import) adds the full provenance chain detail, the evidence export, AI reply drafts and the reviewer-fraud graph. Agency ($79/month, 50,000 reviews per import) is the multi-location tier. Nothing about the free audit is a sample or a teaser — the report you get free is the real analysis of your real rows.

Worked example: the word that told the truth

A fictional worked example, written to show the numbers and the decisions in practice.

A guesthouse's summary, computed over 143 verified reviews, listed its most-mentioned themes as expected — 'breakfast' and 'quiet' among them — plus one unwelcome guest: 'stairs'. The owners checked the verified set and found eleven reviews mentioning the steep staircase, six of them from older guests, all transaction-linked and kept. No one had complained directly in months; the pattern was visible only in aggregate, only because the summary counted verified text rather than impressions. The response was a handrail, a ground-floor room reclassification and honest wording on the booking page. The next quarter's summary dropped 'stairs' from the top words. The owners now read the summary as an operations report that happens to be public, rather than as marketing that happens to be accurate.

Your five-action checklist

  1. Confirm the summary is generated from verified reviews only
  2. Leave the AI-generated label attached wherever the summary appears
  3. Read the top words quarterly as an operations signal, not a slogan
  4. Never hand-edit the summary's text — change operations, then let it recompute
  5. Check that the summary vanishes, rather than guessing, when no verified reviews exist

Frequently asked questions

What does the summary actually compute?

Over verified reviews only: their count, their average rating, and the most frequently used substantive words (longer than five letters). It is deterministic extraction, not free generation.

Why does it use only verified reviews?

So the public summary can never borrow credibility from unverifiable sentiment. UNLINKED and flagged reviews inform your audit; they do not speak in your summary.

Is the summary displayed on Free?

The widget and certificates are part of the free tier for one location; the summary accompanies the verified display and follows the same verified-only, labelled rules.

Can we write our own summary instead?

You can write marketing copy anywhere marketing copy belongs — but not in the summary's place and dress. The labelled, computed digest is a different instrument with a different trust basis.

Keep reading

Related guides from the ReviewProof library:

Start with the tool itself: open the free audit desk, see how ReviewProof works, compare Free, Pro and Agency plans, inspect the sample provenance report, or browse all ReviewProof guides.