On October 1, Google updated its guidance on AI-generated content, and the change is short but quite concrete: it now says it is critical to manually fact-check and review everything made with AI before publishing it, and that this review also applies to titles, meta descriptions, structured data and image alt text. My read: the practical risk of publishing with AI lies in the errors a team pushes out at scale without looking at them, and the new guidance points right there. Here is what changed, what didn't, and how I would set up content review so it doesn't turn into a bottleneck.
What changed in Google's guidance on October 1?
The Search Central changelog says the guide on using generative AI content was updated "with information from the Search Quality Raters guidelines," to keep the documentation in sync with what Google presents at its developer events. It sounds like paperwork, but the new text adds three ideas that weren't written down before.
The first explains how the tool works: generative models don't retrieve facts, they predict a likely sequence of words based on their training data. The second draws the obvious conclusion: because of this, models can produce inaccuracies, also known as hallucinations. The third is the instruction: it is critical to manually fact-check and review all AI-generated content for accuracy and trustworthiness before publishing.
The fourth change is a wording change, and it's the one I find most interesting. The previous version said the focus on accuracy "includes metadata." The new one says "this review also applies to" metadata. In other words, manual verification is no longer limited to the body of the article: it covers all the content that can show up in search results.
It's worth being clear about what the update does not bring: there is no new penalty and no system that detects AI-generated content. It's documentation, and it's best read as a more precise description of what Google expects from people who publish with AI.
What is AI-generated content for Google?
AI-generated content is any text, image, audio or video produced by an artificial intelligence model from a prompt, whether with ChatGPT, Gemini, Claude or another tool. It can create something from scratch (an article, a product image) or transform something that already exists (summarize, translate, rewrite a product page for your brand).
For Google, in practice, it's content like any other. Its position since 2023 is that using AI doesn't give a page a special benefit, and doesn't punish it on its own either: what it evaluates is whether the text is helpful, original and trustworthy for the searcher. The October update keeps that position and adds a practical nuance: if the model makes up a fact and you publish it, the error belongs to your site.
Does Google penalize AI-generated content?
Google's short answer is that it depends on what you publish. One of its spam policies, scaled content abuse, prohibits producing many pages that aren't helpful to users with the goal of manipulating rankings, whether a model wrote them, a team of writers, or a mix. The AI guidance says it directly: generating many pages without adding value may violate that policy.
The underlying standard is still E-E-A-T, which puts experience, expertise, authoritativeness and trust above the production method. AI-generated content that has been reviewed, has sources and has something of its own to say can be helpful and rank well. A hundred texts generated without review to cover a hundred variants of a keyword is the pattern the policy describes.
Context matters too. The September 2026 spam update started on September 24, and Google said the rollout could take up to two weeks, so as of October 2 results are still moving. If you want the details, I explained them in the article on the September spam update. There's no evidence that the update and the guidance change are connected, but they do overlap on the calendar, and that's reason enough to calmly review what you've published with AI in recent months.
What do sections 4.6.5 and 4.6.6 of the rater guidelines say?
The updated guide points to two sections of the Search Quality Rater Guidelines, the document used by the people hired to rate the search engine's results.
Section 4.6.5 describes scaled content abuse: lots of content created with little effort or originality, with no editing or human curation, and it mentions generative AI as one example of a tool used for this. Section 4.6.6 calls for the lowest rating (Lowest) when all or almost all of a page's main content is copied, paraphrased or AI-generated with little to no effort, little originality and little added value for visitors.
One detail worth keeping in mind: the same guide clarifies that these ratings don't directly influence ranking. They are used to evaluate how Google's ranking systems are performing. But they show quite precisely what kind of content Google wants its systems to push down, and "little effort" is a criterion any team can check on its own site without special tools.
Why does the review now include metadata?
Because metadata is the part of content that gets generated in bulk today and nobody reads. On a large site, the SEO team rarely writes 3,000 meta descriptions by hand; a tool or an AI script produces them, someone checks the first five and the rest goes straight to production. The problem is that these texts are seen in search results before your page is, and the guidance now includes them in the review.
| Element | Where it shows | Typical AI error | What I would check |
|---|---|---|---|
| Title | Search results and browser tab | Promises something the page doesn't have, or repeats the same formula across hundreds of URLs | That it describes the actual content and isn't duplicated |
| Meta description | Snippet under the title | Makes up a fact, a price or a condition | Every figure against the source |
| Structured data | Rich results and AI systems | Marks up a review, price or date that isn't in the visible content | That everything marked up is visible and true |
| Alt text | Images, accessibility, image search | Describes an image that doesn't match, or stuffs keywords | That it describes what the image shows |
For structured data, the guidance also asks you to follow the general guidelines and the policies for each search feature, and to validate the markup. If you want the context on why schema also matters for AI, I covered it in this article on structured data.
Do you have to disclose that content was made with AI?
In Google Search, there's no general obligation to label content. The guidance does suggest giving context about how the content was created when it makes sense for your audience, for example by explaining how you used automation. It's transparency as good practice, the same way IBM or Microsoft recommend it for their own documentation.
Where there are concrete rules is in other products. YouTube asks creators to disclose AI use when a video makes it look like a real person said or did something they didn't, or recreates a realistic scene that didn't happen. And for ecommerce, Merchant Center requires labeling what's generated with AI: images must keep the IPTC DigitalSourceType metadata with the value TrainedAlgorithmicMedia, and AI-generated product titles and descriptions go in separate attributes (structured_title and structured_description) marked as AI-generated.
This hits retail teams that generate product pages in bulk directly. A few days ago I shared on LinkedIn the example of a backpack meant to carry a 16-inch laptop in the rain whose description reads "Designed to join you on every adventure." Glad to hear it, but does the laptop fit? Does it get wet? A description like that has no error to catch, it just doesn't say anything. And a generated description that does say something (that it's waterproof, for example) has to be true, because it now travels to Merchant Center, to AI Mode and wherever the product is shown outside your site. If you sell online, in this guide on agentic shopping I explain why that text matters more and more.
What is AI actually useful for in a content team?
AI-generated content is useful for a lot, SEO included. It speeds up initial research, organizes ideas, proposes structures, produces a first draft and makes it possible to produce content at a scale a human team can't reach (product pages, translations, variants for social media). None of the pages ranking for this search recommends stopping using it, and neither do I.
What changes is where you put people. A model can write a very convincing description of a backpack it has never seen, so human work shifts toward verification, editorial judgment and brand voice. For an SEO team, that means the time that used to go into writing now goes into reviewing. If you want the full process for writing content with AI, I detailed it in how to use AI to create content without hurting your rankings.
How do you build a review process that scales?
The objection I hear most is that reviewing everything by hand doesn't scale. That's true if "reviewing" means one person reads every word of every page. But there's a middle path. This is how I would set it up:
- One owner per content type. Someone with a name signs off on articles, another person on product pages and another on metadata. If nobody owns the meta descriptions, nobody reviews them.
- A source for every fact. Every figure, date, price or condition in the content has to come from a source you can open (official documentation, your product database, a cited study). If the AI brought it in without a source, it gets deleted or verified.
- An automated check before publishing. Rules a machine can check without getting tired: length, terms the brand doesn't use, that every figure appears in the source, that the schema validates. Whatever fails the check doesn't get published.
- A human sample of bulk output. For metas, alt text and product pages, a person reviews a sample from every batch. If the sample has errors, the whole batch gets reviewed.
- A log. What was published, with which artificial intelligence tool, who reviewed it and when. It will come in handy the day an update moves something and you need to understand what happened.
An example close to home: the content on this very blog is written with artificial intelligence agents, starting from a brief built with search data. Before publishing, a script checks length, terms, entities and topic coverage, and if it doesn't return ALL OK, the article doesn't go out and waits for review. It's not perfect (the script doesn't know whether a fact is true, that's what the source rule covers), but it keeps a record of what was checked on each piece.
Gary Illyes, from Google, had already said publicly that AI content needs to be curated by humans and fact-checked. What's new is that it's now written into the official documentation, with metadata included.
My suggestion for this week doesn't need a budget: pull a sample of 20 meta descriptions or product pages of AI-generated content from the last three months and check them against the source. If you find errors, you already know where to put the first owner of your content review process.