In short
- Generative engines favour content written by users, though very unevenly by platform: Reddit is the leading source on Perplexity and Google AI Overviews, while ChatGPT leans first on Wikipedia and, since August 2026, less and less on community content. Brand pages dominate nowhere.
- Your customer reviews are the only user-generated content you host yourself. They describe you in a customer’s words, exactly the register AI assistants pass on.
- Star ratings without text are worthless in this game. What gets quoted is a sentence that answers a question about sizing, fabric or day-to-day use, not a score.
- Markup decides who gets to read them: the crawlers of OpenAI, Anthropic and Perplexity do not execute JavaScript. Reviews loaded after the fact by a third-party script are invisible to them. Gemini is the only exception, because it inherits Google’s rendering engine. The same reviews, marked up in the page source, become a source for all of them.
In our article on citability we set out an uncomfortable truth: most of what decides whether you appear in an assistant’s answer does not sit on your site at all, it sits in what other people say about you. This article covers the one exception to that rule, the only surface where “what people say” belongs to you: your customer reviews.
What changed in August 2026 On 8 August 2026, ChatGPT Search changed how it queries the web: its internal queries targeting a named site, using the “site:” operator, jumped from 0.4% to nearly 17% in a single day, and Reddit’s share of its citations collapsed from 3.8% to 0.5% within a week. ChatGPT now favours official documentation and help centres, while Perplexity and Google AI Overviews still follow the earlier pattern. For the subject of this article the consequence is clear: the register of testimony keeps its full weight on Perplexity and Google, and reviews hosted on your own pages are precisely the content the new ChatGPT comes to read on your site.
Why a review outweighs a product page
When an AI assistant has to answer “does this coat come up big?” or “is this brand any good?”, it is looking for experience, not arguments. A product page asserts; a review testifies. It is the same mechanism that puts forums at the top of the citation lists on Perplexity and Google AI Overviews: lived experience earns the models’ trust because it earns the trust of the humans they learned from.
Your reviews therefore lead a double life. The first, familiar one reassures the visitor on the page. The second, newer one feeds the answers generated about your brand and your products, and that second life comes with demands the first never made.
What makes a review citable, and what kills it
| Criterion | What gets quoted | What does not |
|---|---|---|
| The content | A sentence that answers a buying question: “true to size, go one up”, “washed ten times, no bobbling” | “Very happy, as described”, five star ratings with no text |
| Machine readability | Reviews in the page source, marked up as structured data, aggregate rating declared | Third-party widget loaded after the fact, reviews in an external frame that crawlers do not read |
| Trust | Verified review labels, dates, a consistent volume, some middling scores in the mix | Nothing but recent, anonymous five-star entries: humans distrust that pattern, and the models learned the same distrust |
| Where they live | On your product pages and on the third-party sites the engines already cite | Only on a review provider’s platform, on a domain nobody visits |
The mistake that costs the most Treating a negative review as rubbish to be buried. A review profile with no criticism at all is unreadable: buyers distrust it, and nothing suggests a model trained on their writing sets that aside. A well-argued three-star review with a calm reply from the retailer is an asset: it proves the rest is genuine, and the reply shows how you handle a problem. That is the part that sells.
Collecting reviews that answer real questions
Left to chance, you get “all good, thanks”. Useful reviews have to be prompted, through the question you put to the customer.
- Replace “leave a review” with two concrete questions specific to the category: sizing, fabric, installation, how it holds up after a few weeks. Customers answer what they are asked; ask what future buyers will want to know.
- Pick your moment. A review requested on delivery day is a review about the courier. The right delay depends on the product: a few days for fashion, a few weeks for anything that wears out or has to be fitted.
- Target the products that matter. Your best sellers and your most returned items first: the former concentrate the queries, the latter are the ones future buyers have the most questions about. A review saying “comes up small” on the right page prevents a return, and we have put a figure elsewhere on what an avoided return is worth.
- Reply to criticism, publicly and calmly. The reply is read by everyone, machines included, and it is the only content on the review section that you write yourself.
A ten-minute self-check: take your best seller and read its last ten reviews with one question in mind, would a future buyer find the answer to their main hesitation there? Then view the page source and search for the text of a review. If it is not there, your reviews do not exist for a share of the automated readers.
The options, and what each is worth
| Approach | What you get | Who it suits |
|---|---|---|
| Reviews hosted on your own pages, marked up, prompted by targeted questions | The asset belongs to you, every crawler can read it, and it enriches your product pages with customer vocabulary | Everyone: this is the foundation |
| A third-party review platform on top | Independent proof, useful for trust and sometimes cited in its own right. On top of the foundation, not instead of it | Brands still building recognition |
| A third-party widget on its own, loaded by script | Reassures the human visitor, invisible to a share of the machines: half the value evaporates | Something to fix, not to choose |
| Buying or filtering reviews | Prohibited under article L121-4 of the French Consumer Code, detectable, and counterproductive right down to the generated answers: a profile that clean discounts itself | No case at all |
The tooling
On PrestaShop, we deploy for our clients the module published by Datafirefly Limited, our agency’s sister company: Certified Customer Reviews collects after purchase, publishes reviews inside the page with the structured data markup, displays verified review labels and produces a review summary per product. One-off purchase · 12 months of updates.
Method beats tooling, and here the line is clear: the module collects, marks up and publishes. The questions you put to customers, the choice of which products to target and the replies to criticism remain editorial work, and that is what separates star ratings from a citable source.
The bigger picture
Visibility is shifting towards systems that prefer testimony to argument, or that, like ChatGPT since August 2026, come and look for the answer directly in the pages you own. Either way, reviews published and marked up on your own product pages sit on the right side. The brands that have understood this stop treating reviews as a reassurance box at the bottom of a product page, and start treating them as what they have become: the only community-written content they actually own.
This is one of the most concrete angles of our artificial intelligence expertise, precisely because it requires no new technology: better questions, clean markup, written replies. The rest of the subject, everything that plays out beyond your own site, is covered in our article on citability.
Sources
The technical and legal claims in this article point back to the source that establishes them, accessed on 11 August 2026, the fifth on 24 August 2026. We do not cite a source we have not read.
- Vercel and MERJ, The rise of the AI crawler, December 2024, more than 500 million AI crawler requests analysed. Read
- Google Search Central, official documentation on review snippets in structured data. Read
- French Consumer Code, article L121-4, points 27 and 28, transposing EU directive 2019/2161, in force since 28 May 2022. Read
- Profound, AI Platform Citation Patterns, June 2025, 680 million citations across ChatGPT, Google AI Overviews and Perplexity, August 2024 to June 2025. Read
- Promptwatch, Why Did ChatGPT Stop Citing Reddit, August 2026, daily share of ChatGPT Search citations from 7 July to 17 August 2026. Read
FAQ
Do the star ratings in search results depend on this?
Rich snippets require reviews marked up as structured data on the product page itself, with an aggregate rating, and visible to the visitor: Google rejects markup describing content absent from the page. An external widget with no markup does not qualify. One restriction is often missed: reviews a company publishes about itself, through organisation or local business markup, are explicitly ineligible for star ratings. Reviews of a product remain eligible. So the same work serves both purposes: star ratings in Google and readability for AI assistants.
Can I rewrite or screen reviews before publishing them?
Moderating anything illegal or abusive, yes. Removing negative reviews, no: article L121-4 of the French Consumer Code, which transposes EU directive 2019/2161, lists as misleading commercial practices in all circumstances the act of altering consumer reviews to promote a product, and the act of claiming reviews come from genuine buyers without having verified it. Beyond the legal risk, the spotless profile it produces undermines the very trust it was meant to build. The right answer to a critical review is a reply, not a deletion.
How many reviews do I need for this to count?
Fewer than you think, if the text is good: a page with eight reviews that answer the real questions beats a page with two hundred “perfect, thanks”. Concentrate collection on the products that matter rather than chasing a uniform volume.
Is an AI-generated review summary on my product pages a good idea?
Useful for the visitor in a hurry, on two conditions: that it is clearly identifiable as a summary, and that it does not push aside the reviews themselves, which remain the citable material. The summary helps the human; the full text feeds the machine.
Can Dotsland help?
Yes. An audit of how machine-readable your current reviews are, a rework of your collection questions by category, the markup, and a routine for replying to criticism. It is one of the projects covered by our artificial intelligence expertise. Let’s talk, or start by searching for the text of a review in the page source of your best seller.