Google just wrote a fake-review rule into review markup. AI answers read a different review layer.

What Google actually changed

The Search Central documentation updates log records one entry for July 24, 2026: a new guideline added to the review snippet documentation covering fake and undisclosed incentivized reviews, with the stated purpose of improving user review transparency. The review snippet page itself (last updated 2026-07-24 UTC) now carries the guideline verbatim:

"Don't include fake or undisclosed incentivized reviews on your page or in your structured data markup. Examples include: Reviews that aren't based on a genuine experience of a product or service · Reviews written in exchange for a benefit (such as money, discounts, vouchers, or free products) that don't clearly and prominently disclose the incentivization." — Google Search Central, review snippet guidelines, added 2026-07-24.

Two things this is not. It is not an announcement about AI Overviews, AI Mode, or any generative surface — Google said nothing of the sort, and we are not going to extend it for them. And it is not, on its face, a new policy so much as an explicit written statement of one: incentivized-review disclosure was already the expectation in Google's broader spam and review policies. What changed today is that it is now enumerated in the structured-data eligibility guidelines, alongside the existing rules that aggregate ratings must not be pulled from external websites and that review content has to be readily visible to users on the page. Review snippets are eligible for Local business among other types, which is why this lands on our buyers' desks and not only on ecommerce ones.

The part that matters operationally: the incentive has to be disclosed, not removed

Read the second bullet again. It does not say a review obtained in exchange for a benefit is forbidden. It says one that doesn't clearly and prominently disclose the incentivization is. That distinction is the whole compliance question for a local service business, because "$25 off your next visit if you leave us a review" and "free whitening for a Google review" are ordinary, widespread practice in the exact verticals we audit most — med spas, dental practices, home services. If a business runs one of those programmes and its site publishes the resulting reviews in Review / AggregateRating markup with no disclosure, that markup is now explicitly outside the guidelines.

Our own cut: which review surface does an AI engine read?

Because the rule is about reviews on your own page, the obvious next question is how much of that page an AI engine ever consults. We pulled a fresh production export today and re-counted. The corpus is 247 completed free AI-visibility checks — 1,468 answers to category buyer questions (never the business's name) run 2026-06-12 to 2026-07-14, with our smoke-test and self-audit rows removed by exact-name filter. The ChatGPT half is 733 answers, every one of which linked at least one domain: 1,511 distinct domains across 3,787 domain-appearances. (This cut is ChatGPT-only for the same reason as our earlier source-map work — the Perplexity API does not return citation URLs in the answer body.)

Review platforms are 423 of 3,787 domain-appearances (11.2%), spread over 24 distinct platforms. The distribution is extremely top-heavy, and the top of it is not where most review programmes point:

Yelp appears zero times in 3,787 cited domains. We checked that number twice, because it is the one most likely to be a filter artifact: every yelp.com URL anywhere in the export belongs to our own "Smoke Test Plumbing" QA row, which the corpus filter removes. It is also consistent with yesterday's cut, where Yelp was named as a recommendation three times in 4,015 business mentions. On this data Yelp is neither what the engine reads nor whom it names.

Three implications for a business or an agency that sells AI visibility

• Fix the markup, but bank the return in the right place. The new guideline is about rich-result eligibility in classic search — real, worth doing, and cheap. It is not a lever on AI citation: in 30.0% of answers ChatGPT read someone else's review page about the business and never touched the business's own site at all. Sell the markup fix as search hygiene, not as an AI-visibility outcome, and you will not have to walk it back.
• The layer AI reads has no disclosure rule. Google now enumerates an incentivization-disclosure requirement for the reviews you publish. No AI engine publishes an equivalent standard for the third-party review pages it retrieves — and those pages are where 43.4% of the answers about our audited businesses got their review evidence. If your client's incentivized-review programme feeds a Birdeye or Trustindex widget rather than their own site's markup, today's guideline does not reach it and neither does any published AI policy.
• Audit the platforms the engine actually cites, per vertical. Birdeye's review pages alone are 187 of 423 review-platform appearances — 44% of the whole class, and the most-cited domain in our entire corpus. Yelp is zero. A review strategy built from the standard local-SEO checklist is pointed at the wrong surfaces for this workload.

Limits

• Citations are ChatGPT-only (733 of the 1,468 answers). The Perplexity API does not return citation URLs in the answer body, so the 43.4% and the platform table describe ChatGPT.
• "Review platform" is a hand-curated classifier — domains whose primary unit of content is a user review or star rating of a business. It counts Birdeye, Trustpilot, Trustindex, ConsumerAffairs, TripAdvisor, Healthgrades, Avvo, Zocdoc, Opencare, WebMD's doctor pages, Trustanalytica and 13 others. Errors run one way: an unrecognised review site is counted as not a review platform, so 11.2% and 43.4% are floors.
• Directory sites that carry reviews are excluded — BBB, Angi, HomeAdvisor, Houzz, Expertise, ThreeBestRated and the vertical directories are counted separately (see yesterday's directory cut, where directory domains are 25.4% of cited sources). Fold them in and the review-adjacent share is far larger.
• Convenience sample, single-sampled per question, weighted toward med spas, dentists, personal-injury law and home services in US and UK metros. Directional, not nationally representative.
• We sell this measurement. AskedAbout sells AI-visibility audits, so check the method — which is why it is stated in full.

You can run the free 60-second check on any business and read the source list out of the answers themselves. Agencies baselining a book of clients can run five at once with the $249 Agency 5-pack — 25 buyer questions × 4 engines × 3 samples per business, white-labeled, including the per-answer source map this post is a corpus-wide cut of.