At eight in the evening someone picks up their phone and types this to an AI assistant: "Can you recommend a dental clinic in Kadıköy that is open at the weekend and sees children?" The answer lands in a few seconds: three names, each with a sentence or two of justification underneath. One of them reads "reviewers say appointment times are kept and the wait is short."
That sentence did not come from an ad. It does not appear on the clinic's own website either. The model pulled the trait out of a theme running through the reviews people had left for the business. The user got a reasoned recommendation rather than a list, and the raw material for the reasoning was text written by other people.
For years reviews were reduced to a single number: the star average. In local search results that number moved rankings, so businesses worked to pull the average up. Language models changed the equation. To a model the figure 4.7 is close to meaningless; the meaning sits in the sentences that produced it. This article sets out the routes by which reviews reach AI answers, and the legitimate tools a business has for managing that signal.
Where does AI read reviews from?
A language model can reach the reviews written about you through three separate routes, and confusing them is how effort ends up in the wrong place.
Source one: Google Business Profile. This record behind Maps and local search results is where the densest pile of reviews sits for most local businesses in Turkey. Google's AI Overviews and Gemini's answers to local queries are expected to draw on the local data of the same ecosystem; business name, category, opening hours, location and review bodies all sit together there. A missing category or a blank service area on this profile can keep you out of relevant queries no matter how good your reviews are.
Source two: third-party platforms. Sector directories, booking and appointment sites, e-commerce marketplaces, forums, community sites in the vein of Ekşi Sözlük (a large Turkish user-written entry site), and social media. Tools such as ChatGPT's search mode and Perplexity send a query to a search index before composing an answer and read the pages that come back. In that architecture it matters which platform your review sits on, and it matters just as much whether that platform's page can be crawled. Review sections that stay hidden behind a login, or that load entirely through JavaScript afterwards, tend to fall outside this reading chain.
Source three: structured data on your own site. Review and AggregateRating markup puts the reviews on your page into a form a machine can parse directly. This source carries a trap of its own, which gets its own section below.
That three-way split also shows where classic local SEO and work aimed at generative engines part company; we compared the boundaries of the two disciplines in more detail in our piece on the difference between local SEO and GEO.
Star average or review content?
Hand a language model the fact "4.8 out of 5" and it holds one scalar. Hand the same model the sentence "I went twice, both times I was taken in at my appointment time, and the price was given to me in writing before treatment" and it holds at least three attributes: punctuality, price transparency, written information.
User queries are rarely bare either. Instead of "a good lawyer" people write "a lawyer who keeps me informed through a divorce case"; instead of "beauty salon", "a salon that does not push you on the first session". The model tries to match those qualifiers against the attributes it extracted from review text. The match happens at the level of words, not numbers.
In practice this has three consequences. First, one-word reviews ("great", "recommended") add little meaning; they lift the score and feed no reasoning. Second, it is entirely possible that the areas where your business is genuinely strong never appear in the reviews at all, and in that case the attribute does not exist as far as the model is concerned. Third, freshness matters: a review written three years ago does not represent a team or a pricing policy that has changed since, and models generally weight current content more heavily.
The job here is not to dictate the text of a review; it is to ask a question that reminds the customer of the point where the service actually differed. "Were you satisfied?" produces a one-word answer. "Which treatment did you have, and how did the process go?" produces a paragraph.
Why the reply to a negative review is a signal
A review page is not a one-sided text, it is a two-turn dialogue. When the model reads a negative review it reads the business reply directly beneath it as well, and the last word on the page is your sentence.
A complaint left unanswered means the claim stands unverified but also uncontested. A reply that is defensive, accusatory, or that exposes the customer's identity does more damage than the complaint itself. A reply that works usually carries three elements: acknowledgement of what happened or a courteous correction, a concrete explanation where one exists, and a channel that moves the matter out of the public area ("call us and we will go through your record together").
In fields such as health, law and finance there is an extra boundary here. Confirming that the person received a service from you, or going into the details of a treatment or a case file, is a serious risk in terms of personal data processing. Sharing health information — a special category of data under KVKK, Turkey's personal data protection law — in a review reply is not something to take on merely to be proved right. A safe reply explains the principle without confirming the event.
A flawless profile does not always inspire confidence either. A page made up only of five stars, holding no criticism at all, can look like "thin data" or "filtered data" to a user and to a model alike. A handful of negative reviews answered in measured terms makes the picture more credible.
Legitimate ways to collect reviews
There are shortcuts to raising review volume and all of them are bad ideas. Getting reviews written in exchange for payment or a discount is a plain breach of platform rules; it can lead to the profile's score being reset or the record suspended. There is a misleading-commercial-practice dimension on top of that. Bought review piles usually produce texts that resemble one another, carry no detail, and spike within the same period; that pattern is visible both to platform filters and to a model reading the text.
The legitimate route is slower, but what accumulates there stays:
- Timing. Send the request right after the service, while the experience is fresh. A single reminder is enough; insisting annoys.
- Do not filter. Sending the review link only to the people you know were happy, and routing the unhappy ones to a separate form, is a practice platform rules forbid. The request goes to everyone in the same form.
- Consent and notice. Messages sent by SMS or email can fall within the scope of commercial electronic messages; in that case obligations such as consent, registration with the İleti Yönetim Sistemi (Turkey's central commercial message management system) and the right to opt out come into play. If you are going to use customer contact data for this purpose, that purpose has to appear in your privacy notice. Confirm exactly where the line runs with your legal adviser.
- Reduce friction. A QR code at the till or in the waiting area, a short link at the foot of the invoice, a one-tap link in the appointment closing message. If the user has to navigate three screens, they give up.
- You set the question, not the answer. An open-ended reminder such as "which service did you receive, and what made the biggest difference for you?" produces substantive text without steering it.
- Continuity. A steady flow of a few new reviews a month builds a healthier profile than thirty reviews arriving once a year.
If you are wondering how this flow gets built sector by sector, our GEO scenario for dental clinics shows how the contact points after an appointment are used.
How to show reviews on your site
There are two separate goals here, and mixing them up is a common mistake.
The first goal is stars appearing in search results. Google's rich result rules do not accept business-level rating markup that a business places on its own site about itself; for LocalBusiness and Organization, this kind of "self-serving" markup produces no rich result. So embedding AggregateRating on your home page to say "our customers rated us 4.9 out of 5" earns you no stars in the search result. For separate entities such as a product, a recipe or a course the situation is different; there the rating belongs to the object, not to the business.
The second goal is what the page says when a machine reads it. That goal still holds, and it is usually the more valuable one. When you show reviews on the page as real text, with the source named and not hidden behind JavaScript, an AI crawler reading the page as raw HTML sees both the praise and the reason for it. If you shorten a quotation, do not change its meaning, do not write fake reviews, and do not publish the customer's full name without permission.
On the markup side the healthy approach is to describe your business as LocalBusiness, separate your services out with Service, and support the expertise signals on the page with Person and Organization links. We took this whole structure and its relationship with E-E-A-T apart piece by piece in our article on schema markup.
Measurement: which question, how often
The weakest link in reputation work is usually the measurement side. Review counts go up, but whether the business appears in AI answers stays unknown. The plainest arrangement you can set up is this:
Query set. Write between ten and twenty natural-language questions that do not contain your brand name. Each one should carry a location and a qualifier: "open at night", "somewhere you can take a child", "tells you the price in advance", "free first consultation". They should resemble the sentences your customer actually builds.
Engine list. Run the same sets on ChatGPT (search on), Gemini, Perplexity and Google AI Overviews. Ask while logged out, or from a separate browser profile, so personalization does not distort the result.
Fields to record. Note four things for each query: did your name appear, in which position it appeared, which attribute was ascribed to you, and which sources the answer cited. The fourth column is the most valuable, because it tells you directly which platform's reviews are being read.
Frequency. One full measurement a month is enough. Take an interim measurement after a marked change in review flow, a wave of negatives, or a profile update.
Keep a few internal indicators on the review side as well: new reviews per month, reply rate, average reply time, and the distribution of the themes standing out in the reviews. The last indicator teaches the most; it shows which words your customers describe you with, and those words are exactly the attributes a model will ascribe to you. We set out which metrics are worth tracking on the visibility side in our article on how to measure AI visibility; and if you would like to build the measurement for your own business together, we can agree the scope in a free business analysis you can request from the contact page.
How much a review weighs varies by sector; we treated that difference separately on the solution pages we prepared for beauty centres and for hotels, restaurants and cafés.
Reviews are the only source that says the things a business cannot say about itself. AI answers use that source by turning it into a narrative rather than a score. So what has to be managed is not the score, but whether that narrative is accurate, current and verifiable.
Frequently Asked Questions
Do AI assistants really read business reviews?
They read them directly. Assistants working with search mode send a query to a search index before producing an answer and process the returned pages as text; map profiles, directories and review pages are among those pages. The model uses the recurring attributes in the review text rather than the star average. That is why a business can end up described as "short waiting time" or "tells you the price in advance"; those phrasings were extracted from the reviews.
How many reviews do I need in order to appear in AI answers?
There is no valid threshold number; the right question is not "how many" but "how detailed and how current". Dozens of one-word reviews carry less information than a few detailed ones explaining what the service is and how it runs. Not falling markedly below the band your competitors sit in, sustaining a regular flow, and having the reviews spread across recent months are more decisive in practice.
Is it possible to get a negative review removed?
Reviews that breach platform rules (insults, irrelevant content, disclosure of personal data, plain spam) can be reported and taken down. But a valid review describing dissatisfaction should not be expected to be deleted, and the effort to delete it usually achieves less than replying does. A measured, non-defensive reply that moves the matter to a private channel leaves the last word on the page to you; both the person reading and the model processing the text see that reply.
Is it legal to ask customers for reviews?
Asking is free; how you ask is what sets the boundary. Getting reviews written in exchange for payment, a discount or a gift is contrary to platform rules and carries the risk of a misleading practice towards consumers. Sending the request only to satisfied customers is a forbidden method as well. Since messages sent by SMS or email can fall within the scope of commercial electronic messages, confirm the consent, İleti Yönetim Sistemi registration and privacy notice obligations with your legal adviser.
Should I put AggregateRating markup on my own site?
Business-level rating markup that a business places on its own site about itself is not shown by Google in rich results; for LocalBusiness and Organization this kind of self-rating is not accepted. Publishing the reviews on the page as real, crawlable text with the source named is valuable by contrast, because the AI crawlers reading the page see both the assessment and the reason behind it.