Google's guidance on using generative AI content now tells publishers that the manual review covers structured data. The line went into the page on 1 October: "This review also applies to metadata like <title> elements, meta description elements, structured data, and alternate texts for images, which can appear in Search results." Google's changelog files the edit as housekeeping, bringing the documentation "in sync with our presentations we use at our developer events".
The addition is three sentences and no new policy
Two of the three explain the mechanism. Generative models, the guide says, "predict a likely sequence of words based on their training data", which means their outputs "may contain inaccuracies (also known as hallucinations)". The third sentence is the instruction, and Google did not soften it: "It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing."
Everything underneath that stayed put. No ranking signal was announced, no manual action, and no change to how the spam policy on scaled content abuse works. The page still sends readers to the same two sections of the Search Quality Raters guidelines: 4.6.5, on scaled content abuse, and 4.6.6, on main content produced "with little to no effort, little to no originality, and little to no added value". The version this replaced had stood since October 2025, a detail Barry Schwartz picked out at Search Engine Roundtable the evening the edit appeared.
This is a documentation change rather than an enforcement one. What makes it worth reading is that it names the files where the problem actually lives.
Markup is the output nobody proofreads
AI-written prose on a menu page has readers. The person who pasted it in sees it, a manager scrolling the site sees it, and a diner who orders the dish finds out soon enough whether the description was true. Mistakes in visible copy get caught by the ordinary business of running a restaurant.
JSON-LD has no readers. It lives in a script tag, renders nothing on the page, and reaches only crawlers and the assistants reading behind them. A model asked to produce menu markup returns values that are well-formed and confident, because predicting a likely sequence of words is what the mechanism does, which is the explanation Google has now written into its own documentation. Nobody at the restaurant is positioned to notice when one of those values is wrong.
Prose gets corrected by use. Markup does not, which is why an edit this small is worth the attention. The review instruction already applied to body copy; extending it to structured data names the output where a fabricated value survives longest.
Menu schema carries claims heavier than the prose
Schema.org's MenuItem type takes offers for price, nutrition for nutritional information, and suitableForDiet for dietary suitability, defined in the vocabulary as "a dietary restriction or guideline for which this recipe or menu item is suitable, e.g. diabetic, halal etc." There is no dedicated allergen property on the type, so allergen information ends up inside the description or the nutrition object, written as prose and consumed as fact.
A price that has gone stale in the markup is an irritation. A generated suitableForDiet value is a claim the restaurant has published about what is in its food, in a format built to be read back to somebody who asked the question because the answer matters to them. A kitchen switches supplier; the JSON-LD still says what it said last month.
Google's framing is the useful part: the fix is manual. Markup emitted from the menu someone already keeps current, built out of the dish names and prices and dietary flags a person edited when the kitchen changed, puts the fact-check where the facts are. Generating it separately from a model produces claims about the food that nobody at the restaurant has read, in the format assistants trust most.
