Most business owners we talk to are quietly confident about their reviews. Two hundred five-star ratings on Google. A 4.9 average. A steady drip of new ones every month. The assumption: if Google likes those signals, ChatGPT and Perplexity must like them too.
They do not. Or more accurately, they like a different part of them. AI engines read the text of your reviews, not just the star count, and most reviews are written in a way that gives the AI almost nothing to work with. A 5-star “Great service, highly recommend!” is invisible to the same engine that will happily quote your competitor's 4-star review because it mentions a specific procedure, a named provider, and a real outcome.
This is the gap. Star count gets you past a quality filter. Review text decides whether you get cited. Here is how to close it.
How AI engines actually parse reviews
When ChatGPT, Claude, Gemini, or Perplexity reads a review, it is not counting stars. It is running named-entity recognition on the text, extracting services, products, people, and outcomes, then scoring sentiment in context for each entity it finds. A single review can contribute to your visibility for the procedure named, the practitioner named, the location named, and the outcome described, but only if those nouns are actually in the text.
The same engine reads your competitor's reviews the same way. When a buyer asks “who does Invisalign in Naperville,” the AI is not picking from a list of 4.9-star practices. It is picking from a list of practices where someone wrote the word Invisalign in a positive review recently.
Recency compounds the effect. The BrightLocal 2026 Local Consumer Review Survey confirms what we see in our scans: consumers and AI engines both heavily discount older reviews.
The three attributes of an AI-useful review
- It names the specific service, procedure, or product. Not the category. The exact thing. “Invisalign” beats “orthodontics.” “Estate planning trust” beats “legal services.”
- It names the practitioner or staff member. “Dr. Kim,” “Sarah at the front desk,” “hygienist Maria.”
- It includes a concrete outcome or detail. A timeframe, a number, a before-and-after, a specific result. “Finished in 14 months,” “saved us $4,200 in taxes,” “walked out the same day with the crown.”
Hit two of three and the review starts pulling weight. Hit all three and the review becomes a citation magnet for the exact buyer asking the AI about that procedure in that city.
Before and after
Dental practice.
AI-useless: “Best dentist in town! Staff is amazing. Highly recommend.”
AI-useful: “Dr. Patel finished my Invisalign treatment in 11 months and fixed the gap between my front teeth. Front desk got me a same-week appointment after I broke a crown.”
Estate planning attorney.
AI-useless: “Great lawyer. Very professional. Got everything done.”
AI-useful: “Hired Maria Chen to set up a revocable living trust for our family. She walked us through funding step by step and turned the documents around in two weeks.”
Financial advisor.
AI-useless: “Very helpful and knowledgeable. Trust them completely.”
AI-useful: “Worked with David on a Roth conversion strategy for my early retirement. He modeled three scenarios, picked the right one for my tax bracket, and flagged a backdoor option I had no idea existed.”
Both reviews in each pair are five-star. Both make the customer happy. Only one of them gets cited when a buyer asks an AI engine for a recommendation.
The review prompt template
Customers will write the better review if you ask the right question. Use one of these templated prompts instead, adjusted for your vertical.
For service businesses: “If you have 60 seconds, we would love a review on Google. The most helpful ones mention the specific service you came in for and the person who helped you.”
For practitioner-led practices: “Would you be willing to leave a Google review? It helps future patients most if you mention which procedure you had, which provider you saw, and how it went.”
For high-trust verticals (legal, financial, medical): “The reviews that help future clients most mention what you came in for and the outcome, even at a general level. No need to share anything you would rather keep private.”
What AI engines actually ignore
- Star count in isolation. 200 generic five-stars and 30 specific four-stars produce different citation outcomes.
- Generic praise. “Amazing,” “wonderful,” “the best” carry almost no extractable signal.
- Source platform, beyond a point. Google reviews carry the most weight because volume is highest there, but we do not see meaningful platform-of-origin weighting once Google is healthy.
- Review length past a point. Three solid sentences with the three attributes outperforms ten paragraphs of generic praise.
- Owner responses, mostly. Matters for human readers and health signal but the AI is not extracting much from the response itself.
How to audit your current reviews
Run your last 30 Google reviews through a four-step audit.
- Count how many name a specific service, procedure, or product. Not a category.
- Count how many name a practitioner or staff member by name. First name is enough.
- Count how many include a concrete outcome, number, timeframe, or before-and-after.
- Count how many were posted in the last 90 days. This is the recency baseline.
If fewer than half hit two of the three attributes, your review base is underpowered for AI citation regardless of star count. Both fixable inside one quarter with a new prompt and a steady ask cadence.
Where to start
Pick the prompt template above that fits your vertical and ship it this week. Then run a free AEO Grader scan to baseline where your current reviews are landing across ChatGPT, Claude, Gemini, and Perplexity. Re-scan in 60 to 90 days.
For the broader fix list alongside reviews, five highest-leverage AEO actions cover schema, llms.txt, directory consistency, and answer-format content. For the structured-data layer specifically, see structured data for local businesses. Practice-specific implementations live in AEO for dentists and AEO for law firms.