Article contents
Reviews are the corroboration generative engines read most
For a recommendation question, a generative engine looks for a consensus external to the company, and customer reviews are the most accessible form of consensus: they are numerous, dated, signed, structured as a rating and a text, and published on sites that AI search crawlers explore. In our panels of questions of the “which one should I choose” or “is it any good” type, review platforms are among the most cited sources, often ahead of the sites of the companies concerned.
No publisher has documented the weighting it applies to reviews; what follows is a matter of observation and of how the systems work. A model that has to recommend extracts facts: a rating, a number of reviews, recurring strengths and weaknesses, a recent trend. It finds them in detailed reviews, not in stars alone. And it readily reports them with their source: “rated 4.7 out of 5 from more than three hundred Google reviews, customers mention punctuality”. This is the “corroborated” condition of our complete method, applied to what your customers write about you.
- 1VolumeEnough reviews for a rating to mean something; an average from six reviews is not a fact.Otherwise: the engine prefers a better-documented competitor.
- 2RecencyRecent, regular reviews; a flow that stopped two years ago says the business has changed.Otherwise: the rating is treated as stale.
- 3ContentReviews that name what was bought, what pleased, what disappointed; these are what the model cites.Otherwise: nothing to extract beyond a number.
- 4ReplyFactual replies from the business, especially to negative reviews, read together with the review.Otherwise: the weakness stays without context.
These four dimensions are read together. A high volume of old, empty reviews weighs less than a regular flow of recent, detailed reviews, even fewer of them. And an average rating of 4.3 with reasoned replies to criticism can be recommended ahead of a 4.9 with no text and no replies, because the first profile provides facts and nuance, the second only a number that nothing explains.
Which platforms engines read, depending on your business
The platforms cited depend on the sector and the engine. For a local business, the Google Business Profile dominates on Google, AI Mode and Gemini, and the other engines cite it too, with Tripadvisor for restaurants and hotels and specialist platforms depending on the trade. For software, it is the B2B review sites: G2, Capterra, Gartner Peer Insights. For a product, marketplace reviews and testing sites. For an online service, Trustpilot and verified review platforms. Forums and Reddit form a category of their own, with no rating but rich in accounts, covered in the article on communities.
| Type of business | Platforms most often cited | What matters first |
|---|---|---|
| Local shop or service | Google Business Profile, sector platforms (health, trades, motoring) | Recency, replies, exact category |
| Hotels, restaurants | Google, Tripadvisor, booking platforms | Recency, replies, consistency across platforms |
| Software and SaaS | G2, Capterra, Gartner Peer Insights, specialist forums | Detailed reviews by use case and company size, weaknesses |
| Consumer product | Marketplaces, testing sites, specialist press | Volume, reviews per variant, reasoned negative reviews |
| Online service, e-commerce | Trustpilot, verified review platforms, Google | Volume, replies, transparency about collection |
Table: scroll horizontally.
One check comes before everything else: can the platform be read by AI search crawlers? Some block all or part of the crawlers, or only render reviews after JavaScript execution. A platform the engine cannot read contributes nothing to your generative visibility, whatever your rating. The test is simple: ask the engine “what do the reviews of [company] on [platform] say” and see whether it cites the page, or answers beside the point.
Illustrative data. How to read it: at equal ratings, the brand whose reviews are old looks less active; the recency gap often explains a lost recommendation better than the rating gap.
What a model extracts from a review, and how to get usable reviews
A model extracts from a review whatever can be named: the service or product concerned, a precise fact (lead time, price, welcome, result), a comparison, a reservation. “Very good” cannot be extracted; “came out the day after the call, quote honoured, a seal to redo two weeks later, fixed free of charge” can, and it contains both a strength and a weakness that was dealt with, which makes it credible for a reader as for a model.
You do not write your customers' reviews, but you can steer the moment and the question. Asking for the review right after the job, with a message that recalls what was done, produces more precise reviews. Ask an open question rather than request a rating. And ask every customer, not only those you know are satisfied: pre-selection is contrary to Google's rules for business profiles, and since 2022 European law has required traders to state whether and how they verify that published reviews come from consumers who have actually used the product. The ISO 20488 standard describes good practice for collecting, moderating and publishing reviews.
Regularity matters more than campaigns. Ten reviews a month for a year are worth more than a hundred and twenty reviews in one week, which signal an operation rather than an activity. A regular flow shows the engine that the business is trading today, which no “about” page can prove.
Replying to reviews: the only part you write
The business's reply is read together with the review, and it is the only text on the page that you write. Its function is not to thank or to defend, but to add facts: what was done, what was corrected, what is planned. A reply to a negative review that acknowledges the problem, gives the cause and the correction, turns a criticism into proof of seriousness, and gives the model a nuance it will repeat: “customers report delays in high season; the company says it has strengthened its team”.
Three writing rules. Reply to all negative reviews, and to some of the positive ones, with different replies: copy-pasted replies are spotted by readers and bring no facts. Never question the customer's good faith in public; dispute the facts, with evidence, if necessary. And sign with a name and a role, which ties the reply to an identifiable person, in the logic of author pages.
Review markup: what is permitted, what is excluded
The schema.org vocabulary describes reviews with two types: AggregateRating for the average rating, the number of reviews and the scale, and Review for an individual review, with its author, date, text and rating. Google documents precisely what it accepts: reviews must have been collected on your site from your customers, relate to the item on the page, and not be copied from other sites. For the Organization and LocalBusiness types, Google does not display a review snippet when the entity marks up its own reviews, which it considers self-serving. For Product, Course, Event or SoftwareApplication, display remains possible.
{
"@context": "https://schema.org",
"@type": "LocalBusiness",
"@id": "https://www.plomberie-ferrer.example/#organization",
"name": "Plomberie Ferrer",
"url": "https://www.plomberie-ferrer.example",
"telephone": "+34…",
"address": {
"@type": "PostalAddress",
"addressLocality": "Valencia",
"addressCountry": "ES"
},
"aggregateRating": {
"@type": "AggregateRating",
"ratingValue": 4.6,
"reviewCount": 148,
"bestRating": 5
},
"review": [
{
"@type": "Review",
"author": { "@type": "Person", "name": "Mr Navarro" },
"datePublished": "2026-06-18",
"reviewBody": "Water heater replaced the day after the call, quote honoured. A seal to redo two weeks later, fixed free of charge.",
"reviewRating": { "@type": "Rating", "ratingValue": 4, "bestRating": 5 }
}
]
}What should we conclude? That review markup on your own site has descriptive value, not display value for a service business, and that it in no way replaces third-party platforms. The weight of reviews in recommendations comes from their presence on independent sites; markup serves above all, on a product page, to make the rating and the number of reviews unambiguously readable, as explained in the article on product structured data. Marking up a rating without displaying the corresponding reviews, or with a number of reviews that matches nothing visible, is a mistake that Google penalises and that a model has no reason to believe.
- 01JobThe customer has an experience, with dated facts: lead time, result, welcome.
- 02RequestEvery customer is invited, at the right moment, with an open question.
- 03Readable publicationThe review appears, dated, on a platform that AI search crawlers can read.
- 04ReplyThe business adds facts, corrects if necessary, signs.
- 05RecommendationThe engine cites the rating, the number of reviews, the strengths and the nuance.
The mistakes that cancel out the effect of reviews
The review profiles we audit rarely lose recommendations because of a poor rating. They lose them because of practices that make the reviews unusable, or suspect, for a reader as for an engine.
- Pre-selecting the customers asked, which produces a profile with no criticism and no nuance to cite.
- Reviews incentivised by an undeclared benefit, or written by the company itself.
- Identical replies to every review, or no reply to negative reviews.
- A rating displayed without a number of reviews, or a marked-up number of reviews that matches nothing visible.
- Concentrating on a platform that AI search crawlers cannot read.
- Several business profiles for the same location, which scatter the reviews and blur the entity.
- A removal request for every negative review, instead of a factual reply.
- A massive collection campaign followed by eighteen months of silence.
Measurement follows the same logic as citation share. On the question panel, you isolate the recommendation questions, record on each run whether the company is cited and which review platform serves as the source, then track every month, platform by platform, four indicators: rating, volume, share of reviews under twelve months old, reply rate. The third-party pages that aggregate these reviews, comparison sites and directories, are the other half of the subject, covered in the article on third-party pages.
What to remember
- To recommend, a generative engine reads reviews on readable third-party platforms; it is the most accessible corroboration.
- Four dimensions count together: volume, recency, review content, business replies. Recency often explains a lost recommendation better than the rating.
- Ask every customer, at the right moment, with an open question; pre-selection is prohibited by Google and regulated by European law.
- Reply to negative reviews with facts: acknowledgement, cause, correction, opening, signature.
AggregateRatingandReviewmarkup describes reviews collected on your site; for a service business, it has no display value and does not replace the platforms.
Frequently asked questions
How many reviews do you need to be recommended by a generative engine?
There is no documented threshold. What counts is that a rating rests on enough reviews to be a fact, that the reviews are recent and regular, and that they contain citable elements. In our audits, a regular flow of detailed reviews, even a modest one, weighs more than a large volume of old, empty reviews.
Can a 4.3 rating be preferred to a 4.9 rating?
Yes. A 4.3 profile with detailed, recent reviews and reasoned replies to criticism gives the model facts and nuance; a 4.9 profile with no text and no replies gives only a number. The engine cites what it can explain, and it is wary, like readers, of profiles without a single reservation.
Should you reply to every review?
To every negative review, without exception, and to some of the positive ones, with different replies that add facts. The reply is read together with the review, by readers and by engines; a reply that acknowledges, explains and corrects turns a criticism into proof of seriousness. Copy-pasted replies bring nothing.
Does review markup on my site improve my visibility in AI?
Marginally. It makes the rating and the number of reviews unambiguously readable, which is useful on a product page. But for a service business, Google does not display self-serving reviews, and no engine has any reason to believe a rating that only your site declares. The weight comes from independent platforms that engines can read.
Can you ask for reviews only from satisfied customers?
No. Google's rules for business profiles prohibit selectively soliciting positive reviews, and since 2022 European law has required traders to state how reviews are verified. Beyond the law, a profile with no criticism at all offers the engine no nuance to cite, and looks suspect. Ask every customer and reply to criticism.
An SEO practitioner since 2001, Kamel Malek runs SEO360 (Alicante, Valencia, Madrid, Paris). He has published three books on visibility in generative engines, including Generative Engine Optimization and Rétablir les faits.
