Your Reviews Are Becoming the Answer: The Review Corpus as an AEO Asset

The short version: when someone in Toronto asks ChatGPT for a "good emergency plumber in Etobicoke" or asks Gemini "which physio clinic near me takes same-week bookings," the engine no longer just counts your stars. It reads your reviews as text, and increasingly it paraphrases that text straight into the answer. Licensing deals are feeding review corpora into the models, and in the Toronto field tests we run the review count is already the stated reason an engine picks one business over another. Your review corpus is raw material now. An unanswered one-star complaint can be folded into the first sentence a buyer reads about you. So can your reply. Here is how that works, what our audits flag, and what we fix first.

Reviews used to be a signal. Now they are copy

For a decade the local-SEO view of reviews was arithmetic. More reviews, higher average, better map-pack position. That still holds, and our guide to ranking a service business on Google Maps covers it. What changed in 2026 is what happens next.

In August 2026 Yelp licensed its reviews, ratings and photos to OpenAI, as Suganthan Mohanadasan reported at the time. ChatGPT's local-business answers now carry a review-highlights field. Google has had a public licensing deal with Reddit for longer, which is why community threads sit so high in AI Overviews for decision questions. Engines are paying for text written by customers, because that text answers the question a buyer is really asking: what is it like to hire these people.

So the sentence a Scarborough homeowner reads about your roofing company may be a machine's summary of what eleven strangers wrote. Not your homepage. Them.

What the engines do with review text today

The pattern differs by engine, and we laid out the full split in why AI engines cite different sources for the same query. The review-specific part looks like this.

ChatGPT grounds local picks in a Google Maps style panel and quotes the star rating and review count as its reasoning. In every Toronto category we have published a study on, the top pick had a deep review base, and the assistant said so. The review-highlights field adds the text layer on top of the count.

Gemini returned place cards with ratings and hours in our consumer-category tests, and treats reviews as trust evidence alongside licensing and credentials.

Perplexity is mostly directory and listicle grounded, but in the review-heavy med spa test it rendered a live places map. Consumer categories pull every engine toward Maps data.

Google AI Overviews lean on Reddit for "who should I use" questions, and a Reddit thread about a GTA contractor is itself a review corpus, just one you do not control.

Claude we measure through its Brave-heavy citation pattern. Whether it paraphrases review text in local answers we have not isolated cleanly yet, so we mark it unknown.

One thing is constant across all five. The engine is reading words customers wrote, and it is choosing which of those words to repeat.

The unanswered negative is the sentence they read first

Here is the part that should change how a clinic manager spends Friday afternoon.

A model summarising reviews is looking for the most informative sentences. Praise is repetitive. "Great service, would recommend" tells the model nothing it can use. A detailed complaint is the opposite. "Quoted $450 on the phone, charged $900, would not explain the difference" is exactly the kind of sentence a summariser keeps. If nobody from the business responded, it stands alone as the record.

Now picture the query. A Mississauga parent types "dentist in Port Credit that is good with anxious kids" into ChatGPT. The answer names three clinics and adds a line under each. For one, the line paraphrases the only detailed review in the corpus, the billing complaint. The parent never opens the Google listing to see the other forty happy patients. The answer was the visit.

We see the shape of this in our tracking. Where an engine adds commentary under a local pick, the commentary tracks the most specific review language available, good or bad. We do not have a clean count of how often that language is negative, and we will not invent one.

Your reply is part of the corpus

The good news sits in the same mechanism. On Google Business Profile an owner reply is public text attached to the review, so a crawler reading one reads the other. A reply is you adding a sentence to the document the engine will summarise.

What a useful reply does:

• States the fact that was missing. If the quote changed because the job scope changed, say what the scope change was, in one sentence. The model now has two sides and a reason.
• Names the service and the place, once, naturally. "Our Etobicoke drain team" or "our North York clinic" anchors the reply to the geo-bound queries that are actually winnable, a point we made in what actually gets you cited by AI. One mention reads as context. Three reads as spam.
• Offers a concrete next step. A phone number, a named person, a refund policy. Specifics survive summarisation. Apologies do not.
• Never argues. A defensive reply becomes the thing that gets paraphrased. "Owner disputes customer accounts" is a sentence you do not want written about you.

Replies to positive reviews matter less. A short one that restates the service ("glad the furnace install in Vaughan went smoothly") still adds the words you want associated with you.

Recency and substance beat the raw count

Review count is the number ChatGPT quotes, so it is not nothing. Still, across the sites we monitor, a steady flow of recent, detailed reviews holds up better in AI answers than a large but stale pile. Two hundred reviews from 2022 and nothing since reads like a business that stopped caring.

Short reviews are nearly invisible to a summariser. Reviews that name the job, the neighbourhood and the outcome get lifted. "Would you mind mentioning what we fixed and which area you are in?" produces the kind of review an engine can use. "Please leave us five stars" produces noise.

What an AEO audit flags in the review corpus

When we audit a Toronto or GTA local-service business for AI visibility, the review layer is its own section of the report. The checks, in the order we run them:

1. Reply rate on negative reviews, Google first. Any one- or two-star review with no owner response is flagged, with the date. Older unanswered negatives are flagged harder, because they have had longer to become the canonical text.
2. Reply quality. Templated replies that say nothing ("we are sorry to hear this, please contact us") are flagged as missed text.
3. Recency cadence. How many reviews in the last ninety days, and whether there is a gap.
4. Substance ratio. What share of recent reviews name a specific service or location. Low substance means the corpus cannot carry a query like "sump pump replacement in Richmond Hill" even if the count looks healthy.
5. Platform spread by category. For trades, HomeStars is part of the surface ChatGPT reads, so the same reply discipline applies there. For clinics, the medical rating sites. Yelp now sits inside a licensed corpus, so even a thin Yelp profile with one unanswered complaint is worth clearing.
6. Consistency with the site. If reviews praise your weekend availability and your site says nothing about hours, the engine has to pick a source. Make them agree.

The free AEO checker covers the on-site side: crawlability, structure, schema. It does not read your reviews. This article is the part the checker cannot see.

What we fix first

The first items are cheap and the engine reads them immediately.

1. Reply to every unanswered negative review on Google, oldest first, with a fact and a next step in each.
2. Rewrite the templated replies from the last year so each one carries at least one specific detail.
3. Set up a review request that asks for the job and the area, sent within a day of the work finishing, and run it for ninety days before judging it.
4. Clear the secondary platforms that your category's engines actually read. Trades, start with HomeStars. Clinics, start with the rating sites patients already use.
5. Bring the website into line with what the reviews say, so the two sources confirm each other instead of competing.

None of this is glamorous. All of it changes the words available to a model when a buyer in the GTA asks who to call. If you would rather we run the measurement and the review work alongside the rest of your AI visibility, that is part of our SEO and AEO service.

FAQ

Do AI engines actually read the text of my Google reviews, or just the star rating? Both, and the balance is shifting toward the text. In our Toronto field tests ChatGPT quotes the star rating and review count as its stated reason for a pick, and since Yelp licensed its reviews to OpenAI in August 2026 its local answers also carry a review-highlights field built from review text. Google's AI Overviews pull from Reddit threads, which are reviews by another name.

Can replying to a negative review really change what an AI says about my business? It changes the text the engine has to work with. An owner reply on Google Business Profile is public and sits with the review, so a crawler reads both. We cannot promise a specific engine will quote your reply, but an unanswered complaint is guaranteed to be the only version on record.

How many reviews does a Toronto service business need to show up in ChatGPT? There is no fixed number we can honestly give you. In the Toronto categories we have studied, the businesses ChatGPT named carried deep review bases and the assistant cited the count, but the threshold varies by category and by how crowded the shortlist already is. A steady flow of detailed, recent reviews that name the job and the area outperforms a large, stale pile.

Should I ask customers to mention my city or neighbourhood in their review? Ask them to describe the job and where they are, in their own words. A review that says "replaced our furnace in Vaughan in one day" gives the engine a sentence it can match to a geo-bound query, the kind a local business can actually win. Do not script the wording. Coached reviews look like spam to humans and to summarisers alike.

• $3,000 Growth CreditFree assessment · credit applied to onboarding. See if you qualify →

Want the optimization playbook, not just the platform overview? Get our free ChatGPT Ads cheat sheet, context hints, bid-floor tests, industry readiness, and the measurement stack from practitioners running real budgets.

Quick answers