Skip to content
← Blog
Search and websites ·

Google wants AI content fact-checked. Here's how to do it on a small site.

Google's guidance now calls a manual check of every AI-assisted piece critical, from the body text to the alt text, but it leaves you to decide what a check is.

A desk with a laptop showing a draft article beside a highlighted printout, set up to fact-check AI content

Google now expects you to fact-check AI content by hand before it goes live, and that includes the text visitors rarely see. On 1 October 2026 Google added three sentences to its guidance on using generative AI content. They say it is "critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing". As PPC Land reported, the review covers titles, meta descriptions, structured data and image alt text as well as the article body.

If you draft your news with ChatGPT or a similar tool, this applies to you. Search Engine Journal and PPC Land both covered the change. This guide explains what changed and what it means for rankings. It then sets out a short checking routine you can use on your own site.

What Google changed in its AI content guidance

The edit is small and sits in one place: the section headed "Focus on accuracy, quality, and relevance" on Google Search Central's page about generative AI content. That page was first added on 21 May 2025. It's the only page in Google's documentation logged as updated on 1 October 2026.

The three new sentences explain why the check matters. The first says "generative models don't retrieve facts, but predict a likely sequence of words based on their training data." The second follows on: "Because of this, generative AI outputs may contain inaccuracies (also known as hallucinations)." The third sets the expectation: "It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing."

Search Engine Journal compared the new text with a version dated 10 December 2025. Relevant Audience compared it with an archived copy from 20 September 2026. Both found the three accuracy sentences were new, and both found the paragraph citing the raters' sections 4.6.5 and 4.6.6 unchanged.

The wording is stronger than usual. Search Engine Roundtable pointed out that "critical" is a term Google rarely uses in its search help documentation. Google's changelog describes the edit as adding information from the Search Quality Raters guidelines, "to get our documentation in sync with our presentations we use at our developer events". It doesn't say which events. None of the three new sentences mentions the raters' guidelines.

The opening line of the page also stays the same. It says generative AI "can be particularly useful when researching a topic, and to add structure to original content." Google isn't telling you to stop using these tools. It's telling you to check what they produce.

It covers more than the words on the page

The second part of the edit is easy to miss. Google reworded an existing sentence so the review extends to title elements, meta description elements, structured data and image alternate text. Its reason is that each of these can be displayed in Search results.

On a small business site, these are often the parts nobody reads. A plugin suggests a meta description, or an image tool writes the alt text. If any of that came from an AI tool, it's now inside the review Google describes.

Structured data needs particular care, because it usually comes from a template. PPC Land cites ClickUp's blog, which lost 97.6% of its traffic in April 2026 with broken JSON-LD (a common format for structured data) across more than 7,000 pages on one template. It then published 2,815 more posts in the same format.

If you sell through Google Merchant Center, the page restates separate rules. AI-generated images "must contain metadata using the IPTC DigitalSourceType TrainedAlgorithmicMedia metadata", and AI-generated product data such as titles and descriptions "must be specified separately and labeled as AI-generated." Merchant Center made the image tag mandatory in February 2024 and added structured_title and structured_description attributes for disclosure in April 2024. In July 2026 Google rolled out an AI label setting across five advertising products, Merchant Center among them, though the Search page edited on 1 October still describes only the metadata route.

Will skipping it hurt your rankings?

Not directly, as far as anyone can tell. PPC Land notes that no ranking change, penalty or enforcement mechanism came with the edit. Nothing on the page says whether Google's systems can tell a checked page from an unchecked one.

The timing has prompted questions. Google's September 2026 spam update began on 24 September, the fourth spam update of the year, with a rollout window of up to two weeks. Earlier 2026 spam updates finished in 19.5 hours in March, about 48 hours in June and three days in August. Search Engine Roundtable said the new wording raised the question of whether the September update targets this kind of content. Google hasn't drawn that connection.

The direction has been clear for some time, though. By 9 April 2025, Google's search quality raters were told to flag pages whose main content came from automated or generative AI tools as potentially lowest quality. Raters are contractors, a workforce estimated at 14,000 to 16,000 people. Google adds a caveat: rater guidelines "are not a guide to ranking first in Google" and raters' scores "don't directly influence ranking." According to Search Engine Journal, Gary Illyes said in an April 2024 podcast that generative AI output always has to be fact-checked, and in an August 2025 interview that AI content needed human curation.

Our view: treat the new sentences as the standard Google will judge your pages against, whether or not a penalty is attached. A checked page is also one your customers can trust, and that matters to them more than any ranking.

Why AI-written errors get published

AI drafts are fluent. A wrong date reads exactly like a right one, so errors pass a quick skim. Research from OpenAI and Georgia Tech tied hallucination partly to evaluation methods that reward guessing.

Connecting a model to sources reduces the problem without removing it. A preregistered Stanford RegLab study of legal research tools built on retrieval-augmented architecture recorded hallucination rates between 17% and 33%.

Marketers see this in their own work. In a February 2026 survey by NP Digital, 47.1% of marketers said they encounter AI inaccuracies several times a week, and 36.5% admitted hallucinated content had been published. Only 23% felt confident using AI output without human review.

Readers notice too. Earlier NP Digital research, from July 2025, found 25% of users had encountered errors in Google's AI Overviews, with inaccurate answers making up 51% of reported problems.

For a local business, the cost of one wrong claim is easy to picture. In July 2026 an AI Overview said Anna Mae's Bakery would close on 31 July. The business denied it. Google told CBC News the vast majority of AI Overviews meet its accuracy bar. That mistake came from Google's own system. Even so, it shows how a single invented fact about opening times reaches customers. The same kind of mistake in your own news post carries your name.

A checking routine for a small business site

The page sets the expectation and stops there. PPC Land lists what it leaves out: it doesn't say what counts as a factcheck, sets no sampling threshold, names no responsible role and requires no record of review. Those gaps are yours to fill. For a small site, a short written rule is enough.

  1. Write from a source you can link. Give the tool the article, announcement or document you're writing about, and ask it to work only from that. Don't rely on what the model remembers.
  2. Check each claim against that source. Go through the draft sentence by sentence. Every figure, date, name and quote should appear in the source. If it doesn't, narrow the claim or cut it. Never fill the gap with a guess.
  3. Open the parts around the article. Read the title, meta description, image alt text and any structured data before you publish. After any change to a template, look at how the structured data appears on one live page.
  4. Name the person who approves. One person, by name, reads the final version and decides it goes live. On a small team that might be the owner; at an agency, the account lead for that client.
  5. Keep a note of what was checked. A line in a shared document is enough: the piece, the source, who approved it and when.

Here's how step 2 works in practice. Say an AI draft includes the sentence "Half of everything on the internet is now written by AI." Check it against the source. A Graphite study in October 2025 found 50% of new written content on the internet is likely produced by AI. The draft has dropped "new" and "likely", and turned an estimate into a fact. A supported version would be: "A Graphite study found 50% of new written content on the internet is likely produced by AI." It's less dramatic, and it's true.

A claim-by-claim check compares the draft with the source, not with what sounds right.

The Desk follows the same routine. Every piece is written from a fact sheet taken from the source, and a separate check compares each claim with those facts. A person approves each piece, and nothing is published by itself.

See how news written in your own voice and approved by you works for small business websites.

Written with AI by Corantr Media Desk from its sources' facts, checked claim by claim, and approved by a person before it was published. How we write.

More from the blog