Reviews as an AI signal: why an assistant sees none of yours, and where to put them

What can an assistant read of your reviews?

Two things, and only two. Text: a review quoted on a page the assistant fetched, on your site or on a platform's public page. And markup: a Review or an AggregateRating node stating a rating, a count and, for a single review, an author and a body. A star widget that a script draws after the page loads is neither; a fetcher that runs no JavaScript sees the empty container it hangs from.

Which platform pages each assistant fetches is not something any operator documents, and we do not claim to know it. What we can measure is your own site, because that is the page every assistant with retrieval will open when a buyer asks about you by name. On 64.7% of the sites in the audit that page carried no rating markup and quoted no review in its text. Only 8.9% carried rating schema of any kind.

Why the reviews you have are invisible

Three habits produce the gap. The reviews were collected on a platform and left there, so your site links to a profile, or does not, and says nothing about what is on it. Or they were embedded with a widget, which loads them in the visitor's browser from the platform's servers; the HTML your server sends contains a script tag and nothing a crawler can quote, and 3.0% of the audited sites rendered their whole content that way. Or they were pasted in as an image of a screenshot, which is text to a person and pixels to a parser.

None of these is wrong for a visitor with a browser. All three leave the assistant that fetched your page with a business it can locate, perhaps, and cannot vouch for. Between two plumbers it can read, one of whom has six customers on the page saying the invoice matched the quote, the choice is not hard.

What an assistant does with a review, as against a rating

A rating is a number; a review is evidence. When an assistant explains why it named a business, the reasons come from text: someone came out at night, someone explained the estimate, someone was still answering the phone after the job. Those sentences are what a buyer's question is really asking, and an assistant assembling an answer lifts them the way it lifts any other passage. A bare 4.8 tells it that people were satisfied; a quoted review tells it what with.

So the page should carry both, and the text matters more than the stars. Choose reviews that name the job, the place and the thing that went right; a review that says "great service" is true and useless. Keep the customer's words as written, cut for length where you must, and say where and when the review was left.

Where to put them, and what the markup rules allow

Two rules in Google's review snippet documentation settle most of this. Reviews or ratings from other websites are not to be aggregated into your markup, so the reviews you copy from a platform go on the page as quoted text and stay out of the JSON-LD. And when a business controls the reviews about itself, which it does on its own site, a page carrying LocalBusiness or Organization markup is not eligible for review stars, even when a platform's widget supplies the reviews. So no stars will appear under your Google result whichever way you do it. The quoted text is the evidence an assistant lifts, and it never needed the stars.

If customers rate you on your own site, through a form you run, those ratings come directly from users and markup for them states a true fact. Know what it will not do. Google's local business documentation recommends aggregateRating only for sites that collect reviews about other businesses, and the self-serving rule still applies, so that markup earns no stars either. Mark up reviews you collect yourself if you like; never build the rating in your markup out of someone else's platform.

Asking for reviews that say something

The review you want mentions the job and the place, because those are the words a buyer's question contains. You cannot write it, and you should not script it; you can ask at the right moment and ask a useful question. "Would you mind saying what we did and how it went?" gets a sentence a machine can use. "Please leave us five stars" gets a number.

Do not pay for reviews, offer discounts for them, or filter who gets asked by how the job went. Consumer rules in most of the places this site's readers operate treat incentivized or selective solicitation as deceptive, we are not lawyers, and the platforms remove what they detect. Beyond the law, the reviews that persuade an assistant are the ones that read like a person talking, and paid ones do not.

Reply to reviews on the platform, including the bad ones, briefly and factually. The reply is text on a public page: it states your side, it shows that the business answers, and if the review named a wrong fact, the reply is where the right one can be read.

What the audit found, and the order to fix it in

The review gap sits next to the identity gap and is usually fixed in the same afternoon. In the same 9,483 sites, 43.1% were inconsistent about who and where they are across their own three main pages, and 36.9% published more than one phone number. A review that praises a business an assistant cannot pin to one address is praise it cannot use. Settle the name, address and phone first, mark them up on the business node, and then put the quotes on the page.

1. Pick three to six reviews that name the job and the place. Copy the text as written, with the reviewer's name as given, the platform and the date.
2. Put them on the service page they describe and on the homepage, as text, with the platform named. Not in a widget, not as a screenshot.
3. Keep the platform's reviews and its average out of your markup; mark up only reviews customers left on your own site, if you collect any.
4. Check the page's source with scripts disabled: the quotes should be there in the HTML.
5. Ask the next ten customers a question that produces a sentence, and reply to every review on the platform within the week.

How to check what a crawler sees of your reviews

View the source of your homepage and search it for the word "review" and for AggregateRating. If the first appears only inside a script tag and the second not at all, you are one of the 64.7%, and the fix is the quoted text above. Then fetch the page as an assistant's crawler would and read the text that comes back; the crawler access check tells you whether the crawlers are allowed in, and a fetch with a crawler's user agent tells you what they get. Finally, ask the four assistants whether your business is any good, and read what they say and where it came from. Our measurement repeats that last step with one of them, Claude, as part of a fixed panel of twelve buyer questions, asked before the fixes, again once they are live, and every month with monitoring.