Assistants read review text, not just the star average. That single change makes what your reviews say more valuable than how many you have. A review naming the service, the place, the customer type and the outcome gives a model four things it can match to a question. A five star review saying “great service” gives it nothing.
This page is part of the ecosystem validation cluster, which covers the whole corroboration layer outside your website.
What changed, precisely
Under traditional local ranking, reviews were an input with three dimensions: average rating, total count, recency. All three are numbers. Systems weighed them and moved you up or down.
Generative systems retrieve and quote language. A review is now a body of independent text about your business, sitting on a domain you do not control, written by somebody with no commercial interest in your success. That is close to the ideal corroborating source, and it is treated as such.
Two consequences, and they point in the same direction.
Content beats count. Fifty reviews of pure praise contain almost no matchable information. Twelve that describe what was actually done contain dozens of retrievable facts.
Reviews now compete for the same job as your own pages. If a customer’s review explains your service more clearly than your service page does, the review is the more citable source, and it will be used.
Which platforms actually count
Not all of them, and the answer is sector-shaped rather than universal.
Google reviews. First in every sector. The volume is unmatched, they feed Google’s own AI surfaces directly, and they are the most consistently crawled review corpus on the web. If you do one thing, do this one.
Your sector’s platform. One per sector, usually obvious:
| Sector | Platform that carries weight |
|---|---|
| Trades and home services | Checkatrade, Trustatrader |
| Healthcare and private clinics | Doctify, iWantGreatCare, CQC comments |
| Financial advice | VouchedFor, Unbiased |
| Legal | ReviewSolicitors, Trustpilot |
| B2B software | G2, Capterra, TrustRadius |
| B2B services and agencies | Clutch, G2, Trustpilot |
| Hospitality and travel | TripAdvisor, Google, booking platforms |
| General B2B and retail | Trustpilot |
A note on Clutch specifically, because it behaves differently from the rest. It is the dominant review surface for agencies and B2B service providers, its reviews are verified through interviews rather than submitted freely, and its listings carry structured project detail: budget band, service line, client industry and outcome. That structure is unusually well suited to how a model matches a qualified question, so for anybody selling a service rather than a product it is worth more per review than a general platform. It is also where a buyer comparing agencies is likely to be looking, which makes an absent or thin profile conspicuous.
Everything else is optional. A thin presence on eight platforms is worse than a dense presence on two, because thin presence with old data is a contradicting source rather than a corroborating one. The rule from the hub applies here: agreement beats presence, presence beats nothing, and an abandoned profile is a liability.
What a citable review contains
Compare two real-shaped examples.
“Great service from start to finish. Would definitely recommend. Five stars.”
Matchable content: none. It confirms somebody was happy. No system can use it to answer a question.
“We needed an EICR on a three storey HMO in Crookes before the licence renewal. They came out within a week, found two issues with the consumer unit, fixed them the same visit and sent the certificate through the next day. Landlord work seems to be their thing.”
Matchable content: the service (EICR), the property type (three storey HMO), the location (Crookes), the regulatory driver (licence renewal), the turnaround (a week), the outcome (issues found and fixed same visit), and a specialism (landlord work). A model answering “who does HMO electrical safety certificates in Sheffield” now has a source.
You cannot write that review. You can change the question you ask.
How to ask, without crossing the line
The line is real. Google, Trustpilot and the sector platforms all prohibit scripted reviews, incentivised positive reviews and selective solicitation. Breaking those rules risks the whole review profile, which is a serious downside for a marginal gain.
What is allowed, and works:
Ask everybody, not just the happy ones. Selective solicitation is the actual breach. Asking your whole customer base is both compliant and produces a more credible distribution.
Ask for detail, not sentiment. “If you have a minute, it helps other people if you mention what the job was and roughly where you are” requests specificity without suggesting an opinion.
Ask at the point of relief. The moment the problem is solved, not four weeks later in a newsletter. Response rate and detail both drop sharply with time.
Make it one click. A direct link to the review form. Every extra step halves completion.
Never offer anything for it. Not a discount, not a prize draw, not a donation.
Your replies are a signal too
This is the underused half and it is entirely within your control.
A reply is indexable text on a high-authority third-party domain, written by you. There are very few of those. Use them properly:
- Reply to everything. Silence on a two star review reads worse than the review does
- Name the service and the area factually. “Glad the HMO certificate came through in time for the licence renewal” is natural and adds retrievable specifics
- On criticism, be factual and brief. State what happened, state what you did, stop. Do not argue. A calm, specific reply to a bad review is one of the strongest trust signals you can produce, because it demonstrates accountability in public
- Never dispute the customer’s experience. You will lose in front of an audience that includes every future customer and every model that reads the exchange
Why a perfect record is a problem
Say the uncomfortable thing. An unbroken run of five star reviews with no criticism anywhere does not read as excellence. It reads as a filtered or purchased profile, both to a careful human and to systems specifically tuned to detect review manipulation.
A distribution with mostly fours and fives, a few threes, the occasional one, and factual replies to all of it is more credible than perfection. It is also what a real business looks like.
The same logic runs through the whole trust layer, and it is set out properly in AI brand trust signals: fabricated or curated signals are detectable, and the ones that carry real weight are the ones that are hard to fake.
Review schema, carefully
Two rules and they are not negotiable.
Only mark up reviews that are genuinely published and visible on the page. Review or
AggregateRating markup asserting something a visitor cannot see is a fabrication a crawler can
detect and a Google manual action.
Do not expect stars. Self-hosted review markup about your own organisation is not eligible for review rich results in Google. It is still worth emitting for AI consumption and entity clarity, which is the reason to do it, but if somebody is selling you review schema on the promise of star ratings in the SERP, they are describing something that does not happen.
MarGen’s own testimonial markup follows this exactly. It is built, wired to a data file, and emits nothing at all until real named quotes exist. The full standard is in the schema documentation that governs every page on this site.
A 60-day review programme
| Week | Action |
|---|---|
| 1 | Identify your two platforms. Claim both. Correct name, address and description on each |
| 1 | Reply to every unanswered review, oldest first, naming the service and area |
| 2 | Rewrite the review request. Ask for specifics, send at completion, one click |
| 2 | Remove any incentive currently attached to review requests |
| 3 to 8 | Send to every completed customer, no filtering |
| 8 | Read the new reviews and count how many contain a service, a place and an outcome |
That last step is the measurement. Not the star average. The proportion of reviews that contain something a model could use.
Review Platforms and Trust Signals: Common Questions
Do AI models read reviews?
Yes, and they read the text rather than only the score. That is the significant change. A star average is a single number a ranking system can weigh, but review text is language a generative system can match to a query and quote from. It means the content of your reviews now matters more than the count of them.
Which review platforms matter most for AI search?
Google reviews first, because the volume and the integration with Google’s own AI surfaces are both unmatched. Then whichever platform your sector actually uses, which might be Clutch, Trustpilot, Checkatrade, Doctify, VouchedFor, G2 or Capterra. Clutch is the strongest of those for agencies and B2B services, because its reviews are verified and carry structured project detail. A dense presence on the two platforms your buyers consult beats a thin presence on eight.
What makes a review useful to an AI model?
Specificity. A review naming the service, the location, the type of customer and the outcome contains four things a model can match to a query. A review saying great service, highly recommended contains none. Twenty specific reviews are worth more than 200 generic ones for citation purposes, though the count still helps elsewhere.
Can I ask customers to write reviews a certain way?
You can ask for specifics without scripting content. Asking somebody to mention what the job was and roughly where is legitimate and produces better reviews for human readers too. Supplying wording, offering incentives for positive reviews, or filtering who gets asked based on expected sentiment all breach platform rules and are detectable.
Do negative reviews hurt AI visibility?
Far less than people fear, and a perfect record is its own problem. A five star average with no criticism reads as unrepresentative to systems tuned to detect manipulation. What matters is the pattern and the response. A handful of critical reviews answered factually is a stronger trust signal than an unbroken run of five stars.
Should I add review schema to my website?
Only for reviews that are genuinely published and visible on the page, with named authors. Self-hosted review markup on your own organisation is not eligible for review rich results in Google anyway, so do it for AI consumption and entity clarity rather than for stars. Markup claiming ratings the page does not display is a manual action risk.
How many reviews do I need?
There is no threshold, and chasing one misreads the mechanism. Enough recent, specific reviews that a model can find language matching a range of queries is the real target, which in most sectors is a few dozen with genuine detail. Recency matters more than most people assume, because a business whose last review is two years old reads as inactive.
How do I respond to reviews for AI visibility?
Reply to everything, name the service and the area factually in the reply, and never get defensive on a critical one. Your reply is indexable text on a high-authority third-party surface that you control the wording of, which makes it one of the few places you can state your own facts somewhere other than your own website.
Where to Go Next
The rest of the cluster: ecosystem validation hub · social platform consistency · public records and governance · media distribution and entity reinforcement
Related: AI brand trust signals · AI search for local business · schema markup guide
Have it delivered: MarGen is a UK GEO agency with published pricing, a published method and documented results.