Google says AI content must be fact-checked before publishing. What that means for your process

What changed on 1 October

Google's Search Central page on using generative AI content was updated on 1 October 2026, and its wording on review is now blunt. "It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing," the page says.¹

The same sentence goes on: the review "also applies to metadata like <title> elements, meta description elements, structured data, and alternate texts for images." Elsewhere the page warns that using generative AI "to generate many pages without adding value for users may violate Google's spam policy on scaled content abuse." It suggests telling readers how content was created. For e-commerce sites it points to Merchant Center policies: AI-generated images must carry IPTC metadata marking them as made by a trained algorithm, and AI-generated product titles and descriptions must be specified separately and labelled as AI-generated.¹

None of this bans AI-assisted writing. What it does is move the responsibility onto a person, by name, before the page goes live. Our view: that is the right line, and a team that cannot show the check happened does not yet meet it.

Why it lands now

The update arrives as AI-written text becomes a measurable share of the web. Pew Research Center sampled about 490,000 English-language pages from Common Crawl between January 2021 and July 2026. In its July 2026 snapshot of 10,000 pages, 10% showed significant signs of AI authorship. Counting only pages published after ChatGPT's release, the share was over one-third.² Our September note on AI-written content covers what that means for citations.

Pew used a detection tool, so the figures describe likely authorship, not certainty. They still make the point: a search engine and an AI answer engine now read a web where much new text was drafted by a model, and accuracy is the scarce thing.

The timing also follows enforcement. Google's September 2026 spam update began on 24 September, globally, with a rollout window of up to two weeks.³ Spam updates and written guidance are separate things. Read together, they tell publishers which behaviour Google expects and which it acts against. We covered the earlier updates this year in our note on Google's 2026 spam updates and AI content.

What should "manually fact-check" mean in practice?

Manual fact-checking means a person compares each factual claim with a source they have opened, before publishing. Rereading a draft for tone is editing, not fact-checking. The difference matters because models write fluent sentences around wrong numbers, dates and names.

A workable standard has four parts. Every statistic, date, price and named fact is traced to a source the checker opened. The checker is named. Anything that cannot be traced is cut or rewritten as opinion. The check is recorded, so it can be shown later to an editor, a client or a regulator.

Our view: the record is the part most teams skip, and it is the part that turns "we review everything" from a claim into evidence. A spreadsheet with the claim, the source URL, the checker and the date is enough to start. Firms that sell expertise can go further and publish the method, as our note on what clients expect from professional-services firms suggests.

A fact-check you can show someone

1. Mark the claims

   Highlight every number, date, price, name and quotation in the draft.

2. Open the source

   Check each claim against the primary page, not a summary of it.

3. Cut or fix

   Remove anything you cannot trace, or rewrite it as a labelled opinion.

4. Record it

   Claim, source, checker, date. Keep it with the page's history.

5. Disclose the method

   Tell readers how the page was made, in one plain sentence.

Scaled content: where the line sits

The scaled content warning targets pages produced in bulk that add nothing a reader could not get elsewhere. Templates that swap a city or a product name into the same paragraph are the classic case. So are summaries of other people's articles with no new data, view or method.

The test Google gives is value to users, not the tool used. A hundred AI-assisted pages, each built on your own product data and checked by a person, are a different thing from a hundred pages generated from a keyword list overnight. If you cannot say what a page adds, it probably adds nothing.

This also matters beyond Google's own results. AI answer engines cite pages that state facts clearly, and a wrong figure on your page can be repeated in an answer about you.

How this applies to AI-assisted posts, including ours

Every post on this site is drafted with AI from sources the Sigzen AI editorial team selects, then fact-checked against every linked source and edited by people. That line appears at the end of each article. Google's update describes the same order of work, and we think every publisher using AI help should state theirs as plainly.

For a marketing team, the practical changes are small. Add a named reviewer field to your publishing checklist. Keep the claim log with the page, and include titles, meta descriptions, structured data and alt text in it. Label AI-generated product data separately in Merchant Center feeds, with IPTC metadata on generated product images. Then audit the pages you have already published with AI help, starting with those that state prices, specifications or statistics.

What the guidance does not say

The page does not say AI-assisted content ranks lower, and it does not require a disclosure label on every page. It recommends telling readers how content was made where that helps them. It does not define how a fact-check must be documented; that part is up to you.

It also does not change what a good page is. Accuracy, a clear answer and something a reader cannot find elsewhere were already the standard. The difference now is that Google has written down who is responsible for the first of those, and when.