Engine policy

Google now says fact-check every AI draft, schema included. Here's the review step

By Arnav Mukherjee, founder of TofuBofu · October 17, 2026

TL;DR

  • Since 1 October, Google calls a manual fact-check of every AI draft "critical". Build that step before you scale.
  • The review covers titles, meta descriptions, structured data and alt text as well as the body copy.
  • AI content still isn't penalised for being AI. Google's target remains many pages that add nothing.
  • AI answers lift single passages and fields, so one wrong number in your schema travels without its page.
  • Check five surfaces per draft, each with one question. The checklist is below.

In September a model drafting part of a customer report for us wrote a market-share figure into its own prose. Nobody had measured it. It read like every other number on the page, and our number checker caught it because every figure in a report has to trace back to a named source. That one traced to nothing.

On 1 October Google wrote the same rule into its own documentation. Its guide to using generative AI content now tells you why models invent figures, and says a person has to check every AI draft before it goes live, metadata included. If you're using AI to write the pages you want AI engines to cite, this is the paragraph that sets your process.

What did Google add to its AI content guide on 1 October?

I compared the live page, which shows a last-updated date of 1 October 2026, with the Internet Archive's copy from 27 September 2026, which still carried the previous date of 10 December 2025. The change sits in one section, "Focus on accuracy, quality, and relevance". Three sentences are new:

One existing sentence moved with them. The old page said accuracy "includes metadata like" titles, meta descriptions, structured data and image alt text. The new one says "This review also applies to" the same four. The list didn't grow. What changed is that it now hangs off a manual review, so a person checking only the body copy hasn't met it.

Most of the trade coverage, ppc.land and Relevant Audience among it, reads it the same way: the four items were already listed, and the review mandate around them is the new part.

Is AI-written content against Google's rules?

No, and this update doesn't change that. Google's February 2023 guidance answered the question directly: "Appropriate use of AI or automation is not against our guidelines." The same page defines inappropriate use as content generated primarily to manipulate search rankings.

The 2026 guide keeps that line. It says AI is useful for research and for adding structure to original content, and that generating "many pages without adding value for users" may violate the spam policy on scaled content abuse. It points to two sections of the Search Quality Raters guidelines, 4.6.5 on scaled content abuse and 4.6.6 on content made with little effort, originality or added value, and adds that the raters' ratings "don't directly influence ranking."

The page names no new penalty and no new ranking signal. The October change states what Google expects a responsible publisher to do with AI-assisted content, and that's the bar your process will be judged against when a page goes wrong.

WHAT GOOGLE'S REVIEW RULE COVERS AI draft predicted words MANUAL FACT-CHECK Body copy every number, name, quote Title element no claim beyond the body Meta description no figure the body lacks Structured data matches the visible text Image alt text describes the real image Publish checked facts only From Google Search Central's guide on using generative AI content, last updated 1 October 2026.

This part is our read, not Google's. A blue link sends the reader to your page, where the context around a claim can soften it or correct it. An AI answer lifts a passage, or a field from your markup, and repeats it inside its own answer. The buyer gets the sentence without the page around it.

That makes the surfaces people rarely read the riskiest ones. We learned it the hard way. A G2 rating we'd taken from a secondary summary turned out to be wrong, and it had shipped inside our FAQPage schema. Markup gets far less proofreading than prose, so an error there tends to last.

Google's structured data policies already say "Don't mark up content that is not visible to readers of the page." Read that together with the new review rule and the test for schema becomes simple: it should say exactly what the page says, and what the page says should be checked.

See what AI engines say about you before you publish more

A free scan asks five AI engines your buyers' questions and shows each engine's answer in full.

Check your AI visibility free

What five surfaces should a review of an AI draft check?

One block per surface, same three fields each time, so the person reviewing knows what they're looking for and can't skip a step by accident.

Body copy

The question: Does every number trace to a named source, and does every company, product and person named in it exist?

How to check: Build a table of every figure before reading the prose: value, source, date. Any number with no row gets cut. Read each quotation against its source; a model will happily tidy a quote into something nobody said.

Title element

The question: Does the title claim only what the body supports?

How to check: Titles get rewritten for clicks, and that's where a hedge in the body becomes a certainty in the headline. If the body says a vendor tests a feature, the title can't say it launched.

Meta description

The question: Does it carry any figure or claim the body doesn't?

How to check: Search the body for every number in the description. A description written separately from the post drifts first, because it's seldom reread after edits.

Structured data

The question: Does every field match the visible text, word for word?

How to check: Compare the rendered FAQPage answers with the rendered FAQ on the page, not the template with the template. Then validate the markup, as Google's guide asks.

Image alt text

The question: Does it describe the image that is there?

How to check: A generated alt text often describes the image the model expected. Look at the image, then read the alt text.

Google also asks you to consider telling readers how a piece was made, where that makes sense for your audience. It's a suggestion for web pages. Merchant Center makes it a requirement: AI-generated images need IPTC TrainedAlgorithmicMedia metadata, and AI-generated product titles and descriptions must be labelled as such.

Who should do the review?

Someone who didn't write the draft, and who isn't the model that wrote it. The writer reads what they meant to say. A second reader reads what's on the page. That's how every new post on this blog is made: each figure goes in a table before a word of prose, external numbers come from primary sources we've read ourselves, and a reviewer who didn't write the post signs off before it publishes. It catches things the automated checks can't, because a check can validate a number and still miss a sentence that misstates what the number means.

If you're buying AI-drafted content from an agency or a tool, ask who does this step and what they check. Ask them to answer the five questions above for one recent piece.

Frequently asked questions

Does Google penalize AI-generated content?

No, not for being AI-generated. Google's 2023 guidance says appropriate use of AI or automation is not against its guidelines. The spam policy that applies here is scaled content abuse: generating many pages without adding value for users, however they are made. The 1 October 2026 update names no new penalty and no new ranking signal.

What did Google change in its AI content guidance in October 2026?

Three sentences were added to the section on accuracy, quality and relevance. They say generative models predict a likely sequence of words rather than retrieving facts, that their output may contain inaccuracies, and that it is critical to manually fact-check and review all AI-generated content before publishing. The existing line about metadata now reads as part of that review. The page shows a last-updated date of 1 October 2026; the previous version was dated 10 December 2025.

Does Google's fact-check rule apply to schema markup?

Yes. The guide says the review also applies to metadata like title elements, meta description elements, structured data and alternate text for images, because those can appear in Search results. Google's separate structured data policies also say not to mark up content that is not visible to readers of the page.

Do I have to label content as AI-generated?

For ordinary web pages, Google asks you to consider it rather than requiring it. The guide suggests adding information on how content was created where that makes sense for your audience. Merchant Center is stricter: AI-generated images need IPTC DigitalSourceType TrainedAlgorithmicMedia metadata, and AI-generated product titles and descriptions must be labelled as AI-generated.

Do Google's search quality raters decide my ranking?

No. Google's guide points to rater guideline sections 4.6.5 on scaled content abuse and 4.6.6 on low-effort content, and says the raters help evaluate Google's ranking systems and that their ratings don't directly influence ranking. Read those sections as a description of what Google is trying to demote.

Why does a hallucinated fact matter more for AI search than for classic search?

Our read: an AI engine lifts a passage or a structured field and repeats it inside an answer, often without the surrounding page. A wrong number in your FAQ schema or meta description can travel on its own, under your name, to a buyer who never visits the page that might have corrected it.

What should a review of an AI draft check?

Five surfaces, each with its own question. In the body, does every number trace to a named source and does every named company and product exist? Does the title claim only what the body supports? Does the meta description carry any number the body doesn't? Does the structured data match the visible text word for word? Does the alt text describe the image that is there?

Sources and further reading

Related reading

Anthropic's usage policy now names AI answers. Where is the line for AEO?
Claude now watermarks what it writes. Marked is not labelled.
The schema study everyone is quoting excluded every page that needed schema

Arnav Mukherjee

Founder, TofuBofu