Google’s AI content guidance: fact-check what your website actually publishes

Does Google now expect you to fact-check AI content before you publish it? Yes. Since 1 October 2026, Google’s guidance on using generative AI content has said: “It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing.” That is a change to Google’s documentation, not an announced ranking change.

Errors can also enter after the draft has been checked. By the time a page goes live, its claims may appear in an SEO title, a meta description, a social preview, alt text, structured data (information supplied in the page’s code for machines to read) or a product feed, and a plugin or AI tool can generate some of those fields after the draft is approved. How do we check that the published title, images, structured data and product information still match the approved claims?

The answer in brief

• Since 1 October 2026, Google’s guidance on using generative AI content calls manual fact-checking and review before publishing “critical”, and explains that generative models predict words rather than retrieve facts. All sources in this article were checked on 6 October 2026.
• Google hasn’t announced a ranking change, a penalty or a ban on AI content, and its spam policies still target scaled, low-value content “no matter how it’s created”.
• The review covers the <title> element, meta description, structured data and image alt text as well as the body. Those four fields were already listed before the update; the explicit call for manual review is what’s new.
• Check the published page, not just the draft. Titles, previews, alt text, structured data and product feeds can keep an outdated claim or lose a qualification that the approved body copy carried.
• Our recommendation for applying that guidance: record a Pass or Blocked decision for every material claim, with the evidence, an owner and a date. One material error blocks the page.

The examples below show how approved claims change before publication. A five-step review process then shows how to catch those changes.

What changed in Google’s AI content guidance?

We compared the current page with a Wayback Machine copy captured on 27 September 2026, which shows “Last updated 2025-12-10 UTC”. Between those two versions, the change is in the passage headed “Focus on accuracy, quality, and relevance”.

The four metadata fields are not new. The September version already listed them. What’s new is the explanation of why AI output can be wrong and the explicit call for manual review, which now also covers those fields.

Google’s documentation changelog says the guide was updated “with information from the Search Quality Raters guidelines”, to get the documentation “in sync with our presentations we use at our developer events”. We found no other textual change between the two versions. Google still says generative AI “can be particularly useful when researching a topic, and to add structure to original content”.

The update fits with Google’s wider people-first content guidance. That page lists the attributes Google’s quality raters are trained to evaluate. On accuracy, it says: “For informational pages, the content should be factually accurate.” On effort, it gives “using generative AI to produce large amounts of text without manual oversight or curation” as an example of “little to no effort”.

Is AI content against Google’s or Bing’s guidelines?

No, not in itself. Google’s FAQ on AI content says: “Appropriate use of AI or automation is not against our guidelines” (Google Search Central blog, 8 February 2023).

Both search engines target scale without oversight:

• Google: its scaled content abuse policy covers cases “when many pages are generated for the primary purpose of manipulating search rankings and not helping users”, typically “large amounts of unoriginal content that provides little to no value to users, no matter how it’s created”.
• Bing: its Webmaster Guidelines warn that large-scale content generated without oversight, quality control or editorial review “may be excluded from indexing”.

AIWIZ’s view, not something either search engine says: an error on one AI-assisted page needs correcting; it does not, by itself, make the page scaled content abuse. Adding a human name to low-value pages doesn’t make them useful, either. Google’s people-first guidance asks publishers to consider accurate authorship information and AI disclosure where readers would reasonably expect them, and it calls fabricating creator profiles “a form of deception”. The same page also says: “Any form of deception makes a page untrustworthy to both users and our automated quality systems, and is a signal of a low-quality page.” Never invent a byline or credentials on a page.

The same accuracy shapes how AI answers describe your business. Bing’s guidelines say “GEO does not guarantee grounding or citations in AI experiences” and ask that “Facts and definitions are explicit”; our guide to generative engine optimisation (GEO), the work of improving how a business is found, cited and described in AI-generated answers, covers the rest.

Why isn’t the approved draft the same as the published page?

Signing off a draft approves one version of its claims. The page that visitors and search engines receive is assembled from several other parts:

• HTML title and meta description. These may be entered in an SEO plugin or taken from an AI suggestion.
• Social preview (og:title, og:description and image). Compare the shared card with the approved claim, because a generated preview can repeat an overstatement. It matters for search too: Google lists og:title among the sources it uses to generate title links.
• Images, captions and alt text. AI tools can now draft these. WordPress 7.0’s optional AI plugin, for example, can suggest titles and excerpts and draft alt text, as we explained in what’s new in WordPress 7.0.
• Structured data. A theme, plugin or ecommerce platform may generate this separately from the copy.
• Product feeds. These may come from a separate product catalogue that nobody compared with the page.

The risk is that nobody compares those versions. Assign someone to do it before publication. Microsoft Bing makes a similar point in its AI Performance help page: “Ensure text, images, and other media describe the same products, entities, and concepts.”

How does an average become a guarantee?

Hypothetical example: an invented SEO consultancy’s service page. The figures below are made up for illustration and don’t describe AIWIZ or any client. The approved body copy says:

“In our 2025 review of 12 client accounts, average organic enquiries rose 18% within six months. Results varied: three accounts saw no measurable change.”

The sentence reports an average across 12 accounts and says three saw no measurable change. The versions below lose those qualifications.

A short title doesn’t have to carry every caveat, but it mustn’t turn a qualified result into a guarantee. If the qualification won’t fit, choose a different angle, such as “Organic enquiry growth: what our 2025 client review found”. Then bring the description, preview, alt text and structured data into line with it.

In the UK, this goes beyond style. The CAP Code, the advertising rules enforced by the Advertising Standards Authority (ASA), says marketers “must hold documentary evidence to prove claims that consumers are likely to regard as objective and that are capable of objective substantiation” before publication (rule 3.7). Rule 3.9 says marketing communications “must not mislead by omitting significant limitations and qualifications”. Google’s guidance is not UK law. But where a page falls within the CAP Code’s scope, a claim in the title needs the same evidence as a claim in the body.

What happens when the product page, structured data and feed disagree?

Hypothetical example: an invented retailer’s jacket. The product page says “waterproof”. The Merchant Center feed, built from supplier data, says “water-resistant”. The page’s Product structured data still shows last month’s price, because a cache is serving old JSON-LD.

That’s two different problems:

1. A factual claim. “Waterproof” and “water-resistant” are different promises, and no validator can tell you which is true. The ecommerce owner checks the manufacturer’s specification. If it supports only “water-resistant”, that wording goes everywhere.
2. A consistency requirement. Google says Googlebot compares the price in your product data with the prices on your landing page or in your structured data, and that a product with a mismatch “may be disapproved”. Google’s landing page requirements say titles, descriptions and images “don’t always need to be identical” to your product data, but “should refer to the same product or product variant”. The price should match, and availability must be shown clearly.

Technical note for whoever manages the feed: if feed text or images are AI-generated, Google Merchant Center adds labelling rules of its own. Its AI-generated content policy requires AI-generated titles and descriptions to be sent in the structured_title and structured_description attributes, with the digital source type trained_algorithmic_media. AI-generated images must carry the IPTC DigitalSourceType TrainedAlgorithmicMedia metadata. These rules apply to product data submitted to Merchant Center. They aren’t a structured data requirement for every AI-assisted page.

In this workflow, the approver holds publication until the page, the structured data and the feed agree with the approved specification and the current price list.

Can structured data pass validation and still be wrong?

Yes. A clean result in the Rich Results Test shows the structured data is technically sound. It doesn’t show the content is true. Google’s general structured data guidelines say structured data “must be a true representation of the page content”. Among the reasons a rich result may not appear, they list data that is “incorrect in a way that the Rich Results Test was not able to catch”. Bing’s Webmaster Guidelines say the same: “Structured data must accurately represent visible content.”

Ask the technical publisher to compare the structured data with what readers can see: headline, dates, author, image, price, availability and description. Don’t add structured data just because a section has the right shape, or expect it to earn a rich result: Google’s changelog records that FAQ rich results “will no longer appear in Google Search starting May 7, 2026”.

How do you fact-check AI content before it goes live?

Use these five steps as your AI content QA checklist.

1. Verify the claim. For anything that could change a reader’s decision (prices, availability, specifications, results, credentials, dates, quotations, legal points), open the original source and record:

• where the evidence is
• its date
• the market it covers, such as the UK
• the product or document version
• any qualifications

The model can help you find a claim. It shouldn’t be the evidence for it. An AI-supplied citation is only a lead until someone has opened it and confirmed it says what your sentence says.

2. Compare every version of the claim. Check each place it appears:

• body copy and H1
• HTML title and meta description
• social preview
• image, caption and alt text
• structured data
• product feed, if you use one

Preserve the claim’s meaning and its qualifications wherever it appears.

3. Inspect the rendered page, then the live page. A CMS field or editor preview isn’t the final output. Check staging before release. Then, straight after publishing, check the public URL:

• compare the title and metadata in the page source with the approved versions
• in Search Console’s URL Inspection tool, select Test live URL and check the live output against the approved version, then run the Rich Results Test to confirm the structured data is technically valid (the tool’s default result is the last indexed version, not the live page)
• on mobile, check that qualifications stay visible alongside the claims they qualify

Caches and templates can serve something other than what you approved. For a material error found after publication, the publisher corrects every affected output and the reviewer checks them again. If a claim can’t be verified, remove it until it can.

4. Record the decision and who made it. Mark each material claim Pass or Blocked, with a reason, the evidence location, an owner and a date. Use Not applicable only for a check that doesn’t apply, such as a product-feed check on a service page. If the evidence is missing or unclear, the claim stays blocked until it’s resolved. One material error keeps the page blocked, however many other checks pass. That’s why we avoid percentage scores: they let a serious error hide among small passes. Record the reviewer’s name in the approval log.

5. Trace repeated errors upstream. If the same mistake appears on several pages, fix the title template, prompt, plugin setting or product dataset that caused it, then check the other pages it produced.

In a small team, one person may hold several of these roles. What matters is that every decision has a name next to it. Where the contract requires client approval, record the client approver’s name as well as the reviewer’s.

What should you do this week?

1. Turn off automatic publishing of AI-suggested titles and meta descriptions, or add a review step for them.
2. Choose ten pages where an inaccurate price, specification or performance claim could affect a customer’s decision, and check every version of those claims.
3. Add a sign-off field for the live page to your publishing checklist, recording who checked it and when.

Then turn to the pages you’ve already published, starting with those where an error would do most harm: prices and availability, product specifications, health, financial or safety claims, and strong performance promises. For a live product page with a material mismatch, check the approved specification and the current price list. Correct the page, the structured data and the Merchant Center feed together. Recheck the published page and the submitted data, then record who made and who checked the correction. For AI-written titles and descriptions already submitted to Merchant Center, check the label attributes as well as the words. Our view: don’t delete or noindex a page just because AI helped write it. Keep it if it’s accurate and useful, and correct errors wherever they appear.

For the next page you publish, name the person who will check the live output and record their decision. If you’d rather have a page checked against this process, send AIWIZ the URL and the claims you’re unsure about; the live-page checks are part of our technical SEO review.