AI Content on Your Site: Google Now Wants Manual Fact-Checks 2026
On 1 October 2026 Google revised its help page on using generative AI content on websites. The key change is short: anything a model writes should be fact-checked by a person before it goes live, and that covers not just body copy but also titles, meta descriptions, structured data and image alt text.
AI content is not banned. But the human review step is now spelled out, and any business producing service or product descriptions in bulk should take another look at its workflow.
What exactly changed
Google’s documentation changelog lists the 1 October entry as an update to the generative AI guide “with information from the Search Quality Raters guidelines”. Search Engine Journal compared the new text with the 10 December 2025 version and found three sentences added to the section on accuracy and quality.
In plain terms, they say:
- generative models predict a likely sequence of words rather than look up facts, so their output can contain errors, often called hallucinations;
- it is critical to check every piece of AI-generated content for accuracy and trustworthiness by hand before publishing;
- that review also applies to metadata that can show up in search results: the
<title>element, meta description, structured data and image alt text.
According to SEJ, the rest of the page stayed the same. Google still describes AI as useful for researching a topic and structuring original work, and still warns that mass-producing pages with no value for people may break its spam policy.
- 1Before 1 October 2026a general call for accuracy, metadata only mentioned in passing
- 2Addedan explicit requirement to fact-check by hand before publishing
- 3Clarifiedthe review covers titles, descriptions, markup and alt text
Sources: Google Search Central, Search Engine Journal
Why it matters now
There is a lot more AI-written text online than a few years ago. In August 2026 Pew Research Center published an analysis of roughly 490,000 English-language pages from the Common Crawl archive (snapshots from January 2021 to July 2026), scored with the Open Pangram detector. Pew found signs of AI authorship on about 10% of all sampled pages and on more than a third of pages published after ChatGPT’s release.
The share is growing fastest on commercial domains:
- .com 9,35 %1,63 %
- .org 4,59 %0,9 %
- .edu 1,03 %0,56 %
- .gov 0,76 %0,34 %
Source: Pew Research Center, August 2026 (six-month averages, Open Pangram detector)
Pew itself cautions that detectors make mistakes and sometimes flag human writing as machine-generated, so treat these figures as a trend rather than an exact count. The takeaway for site owners is still clear: generic AI copy is everywhere, and it is getting harder to stand out with it.
Will Google penalise AI content?
Short answer: not for using AI as such. What matters is the result, not the tool.
Google’s spam policy defines scaled content abuse as producing many pages mainly to manipulate rankings rather than to help users, regardless of how the content was made. The examples listed in the policy include:
- using generative AI to create many pages without adding value for users;
- scraping content and running it through automated transformations such as synonym swapping or translation;
- stitching together text from different pages without adding anything;
- pages stuffed with keywords that make little sense to a reader.
The translation point matters for multilingual businesses, including many in Ukraine that run sites in two or three languages. Our view: machine translation is not a violation in itself, but hundreds of unreviewed translated pages clearly raise the risk.
The guide also points to sections 4.6.5 and 4.6.6 of the rater guidelines, which deal with scaled content and main content made with little effort, originality or added value. Google adds that rater scores do not directly affect rankings; they are used to evaluate how well its systems work.
Separate rules for online stores
If you sell through Google Merchant Center, the requirements are stricter and more specific than in organic search:
| AI-generated element | What Merchant Center requires |
|---|---|
| Product image | Keep the IPTC DigitalSourceType tag (for example TrainedAlgorithmicMedia) and do not strip it, even for images made in Product Studio |
| Product title | Submit it in structured_title with digital_source_type set to trained_algorithmic_media |
| Product description | Submit it in structured_description with the same digital_source_type value |
In practice these tags often disappear without anyone noticing: image optimisers, CDNs and compression plugins strip metadata. If you use generated product photos, check that the tags actually survive all the way to your feed.
What to check before publishing
This is the checklist we recommend for any text an AI tool has touched. It is our advice, not an official Google list.
- Numbers, dates, prices, laws and version names. Check each one against a primary source; if there is none, cut the claim.
- Your own business details. AI is good at inventing plausible prices, lead times, warranties, addresses and phone numbers. Compare them with what you actually offer.
- Title and meta description. They appear in search results, so a mistake there is visible before anyone clicks.
- Structured data. Prices, ratings and availability in the markup should match what the visitor sees; run the page through the Rich Results Test.
- Alt text. It should describe that specific image, not list keywords.
- Translations. Every language version should be reviewed by someone fluent in it, at least on a sample basis.
- Your own contribution. What does the page offer that ten similar pages do not: an example, hands-on knowledge, a comparison, photos of your own products?
It also helps to tell readers how a piece was produced. Google’s guide suggests explaining the use of automation where your audience would expect it.
Building it into how you run the site
Checking one article is easy. Keeping things in order when content changes every week is harder. A solid technical base helps: an admin panel with edit history, dedicated fields for SEO metadata, and structured data generated from the same records as the page itself, so the price in the copy, the markup and the feed never drift apart.
For catalogues with hundreds of items we would suggest not generating everything at once. Start with the best-selling products and key categories, review them properly, and only then scale up. If you are about to build an online store, put those fields and checks into the brief from day one.
What to do next
- Use AI, but make sure a person checks every fact you publish.
- The review covers titles, descriptions, markup and alt text, not just the body copy.
- Bulk AI pages and unreviewed machine translations are the main risk under the spam policy.
- In Merchant Center, AI-generated images and descriptions must be labelled with dedicated attributes.
Want to know whether your site has unchecked AI copy or broken metadata? Get in touch and we will take a look together and tell you where to start.
Author: Sergey Filatyev, founder of the Veb-Dev studio.
Tags
- AI content
- generative AI
- Google Search
- fact-checking
- SEO for business
- Google spam policies