Google says to fact-check AI content, metadata included

On 1 October 2026 Google updated its guide to using generative AI content in Search. It now calls it "critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing", and it says the review also covers metadata: title elements, meta descriptions, structured data and image alt text. If AI writes any of those on your site, a person should read them before they go live.

Google announced no new policy with it. Its changelog says the edit brings the documentation in line with the talks it gives at developer events. What's new is that the review is written down, and that it reaches the parts of a page a CMS can fill in on its own.

What Google changed in its AI content guidance

Search Engine Journal compared the page with its December 2025 version. The update added three sentences on fact-checking, starting with the reason: a generative model predicts a likely sequence of words from its training data, so what it writes can be wrong. Google uses the word hallucinations. The line on metadata changed too, from "this includes metadata" to "this review also applies to metadata".

The page also points to sections 4.6.5 and 4.6.6 of the Search Quality Rater Guidelines, on scaled content abuse and on content made with little effort, originality or added value. And it warns that using AI "to generate many pages without adding value for users may violate" the scaled content abuse policy.

Why the metadata line matters most

When a person approves an article, they read the body. Metadata can skip that step. A plugin writes titles and descriptions for every page, a template fills the schema, an upload tool writes the alt text, and none of it has to pass in front of a person. Google's own wording says why that's risky: these are the parts "which can appear in Search results". A wrong number in a meta description can be the first thing a searcher reads about you.

A review pass for the four places Google names

Titles and meta descriptions

Export them as a list, from your CMS or a crawl, and read them together. Errors that hide on a single page stand out in a column: last year's date in forty titles, a "free" your pricing page no longer offers, a feature name you changed in spring. Then check each claim against the page it describes. Our guide to title tags and meta descriptions that get clicks covers what a good one says once it's accurate.

Structured data

Google's structured data guidelines say the markup "must be a true representation of the page content", and that you shouldn't mark up anything readers can't see. That's where AI-filled schema goes wrong: asked for a rating count, an author or a date, a model will produce a plausible one whether the page has it or not. Run the page through the Rich Results Test to catch broken markup, then compare every value with what the page shows: author, dates, prices, ratings, FAQ answers.

Image alt text

Alt text has two readers: someone using a screen reader, and Search. Generated alt text can describe a generic scene ("a person working on a laptop") or repeat the page's keyword. Open each image and ask whether the text tells someone who can't see it what's in it. If it doesn't, rewrite it as one plain sentence.

The facts in the body

Google asked for accuracy before this update, and the body is still where most claims live. Check every number, name, date, quote and link against a source you opened yourself, and drop any claim you can't trace. Start with the claim the piece rests on, which the brief should already name, as we suggest in how to write a content brief that gets a usable first draft.

Make the review a step in publishing

If the review depends on someone remembering it, a busy week is all it takes to skip it. Put it in the workflow: a named person reads the body and the metadata, then changes the status to published. For metadata written in bulk, review the template first, then read the full list it produced before it ships.

Then size the output to the review. If your team can't read everything it publishes, it's publishing more than it can stand behind, and that's the situation Google's line on generating many pages without adding value describes. One page you've checked does more for you than ten you haven't, the point we made in write one page that ranks instead of ten that don't.

If you want help building a content workflow that uses AI and still passes its own fact-check, that's the kind of work we do.