Ahrefs studied 331,000 pages to answer this. Another study looked at 600,000. Moz wrote a piece asking whether we should stop writing with AI at all.

Meanwhile the answer has been published on a Google page since February 2023, and it is four words long.

"Using AI doesn't give content any special gains. It's just content."

That is Google, on its own developer site, in a FAQ it has never withdrawn. The page it comes from is linked from the live guidance page today.

So no. There is no penalty. Not for using AI, not for the words coming out of a model. Google has said this in writing for three years, in two separate places, and the industry keeps commissioning studies about it.

But something did change. On 1 October 2026 Google added a sentence to that guidance, and the sentence is worth reading carefully, because half of what people are saying about it is right and half of them have not checked.

What I actually did

I did not read a summary of the change. I went and got both versions of the page.

The Wayback Machine has snapshots on both sides of the date. I pulled the 27 September capture and the 2 October capture, stripped them to text, and compared them character by character.

27 September   10,391 characters
2 October      10,753 characters
difference        362 characters

347 new characters, 14 characters reworded, and nothing deleted at all. That is the whole change. On a page of roughly 450 words.

The new sentence

This is what was inserted, verbatim:

"Keep in mind that generative models don't retrieve facts, but predict a likely sequence of words based on their training data. Because of this, generative AI outputs may contain inaccuracies (also known as hallucinations). It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing."

Three sentences. Models predict, they do not retrieve. Outputs may be wrong. Review them before publishing.

Nothing in there bans anything. Nothing describes a consequence. It is a warning about accuracy, and it sits inside a section that was already there, called Focus on accuracy, quality, and relevance.

The half nobody is reporting

Buried in the same edit is a reworded sentence, and it is more interesting than the new paragraph.

The old text said:

"This includes metadata like title elements, meta description elements, structured data, and alternate texts for images, which can appear in Search results."

The new text says:

"This review also applies to metadata like title elements, meta description elements, structured data, and alternate texts for images, which can appear in Search results."

Two words changed. This includes became This review also applies to.

That is not a stylistic edit. It moves metadata from being an example of the advice to being an object of the requirement.

Read the old sentence and metadata is one of the things you should focus on, alongside everything else. Read the new sentence and your title, your meta description, your structured data and your alt text are specifically things you have to fact-check by hand before publishing.

I have not found anybody reporting that change. It is invisible unless you look at the previous version, and reading the current page tells you nothing, because the current page just reads as though it always said that.

Why the metadata sentence matters more than the paragraph

Anyone running a site knows which parts of a page get automated first. Titles get generated by default. Meta descriptions get pulled from the first paragraph or written by a plugin. Alt text gets left empty or filled with the filename.

The body text is the part people actually look at. The four fields Google listed are the part people do not.

And they are the four fields Google displays. A wrong fact in a meta description is shown to a searcher before they ever reach the page. That is why the sentence ends with "which can appear in Search results."

So the practical version of this change is narrower than "review your content" and more specific:

Review the four fields that get surfaced, not just the paragraph you wrote yourself.

What Google did not say

This is the part worth being precise about, because a lot of coverage has gone past it.

It does not ban AI content. The FAQ says the opposite, twice, and that FAQ has not been withdrawn.

It does not describe a penalty. The guidance page names no consequence at all. For consequences it points at the spam policies, which are about mass-produced pages that add nothing, not about which tool produced a page.

It does not say AI content cannot rank. "It's just content." If it is useful and original it can do well. If it is not, it will not.

It does not turn the rater guidelines into a ranking document. The same page carries the disclaimer in its own text:

"These guidelines are not a guide to ranking first in Google; they're used by our search raters to help evaluate the performance of our various search ranking systems, and their ratings don't directly influence ranking."

And it does not say how much reviewing counts. "Manually factcheck and review" is not quantified anywhere.

The thing that actually connects it

Here is the bit that makes the change make sense, and it needs two other documents.

The guidance page already said that using AI to "generate many pages without adding value for users" may violate the spam policy on scaled content abuse. That policy is defined in the rater guidelines, section 4.6.6, as content with "little to no effort, little to no originality, and little to no added value."

I went through that definition in detail in Google's definition of "low value content" is three words, and through the phrase Google uses for it in nine of its own pages. The short version of both is the same: the standard is human effort, and length is not part of it.

But none of those documents said how anybody would ever tell. Effort is invisible. You cannot look at a page and see whether a person did the work.

The October addition supplies the missing piece:

hallucinations      a model invents something, unreviewed
      down
visible errors      evidence nobody checked
      down
little to no effort the rater guideline standard
      down
lowest rating       what that earns

The hallucination sentence is not really a warning about accuracy. It is a way of making the invisible visible. Factual errors are observable, and they correlate with the thing the rater guidelines actually measure, which is whether a person did the work.

I could be wrong about the intent. Google has not said any of that. But it is the only reading that connects the new sentence to the policy the same page already cited.

Where this research is weak, and where it is not

I would rather tell you the limits than have you find them.

Solid. The change is proven, not asserted. Two dated snapshots, one before and one after, with a delta that adds up. Every quote is verbatim from Google's own pages. The metadata rewording is documented, and I have not found it reported anywhere else.

Not solid. The Wayback Machine is an independent archive, not Google. A snapshot proves what the archive captured on a date, not what every visitor was served that day. This is one small edit to one guidance page, not a ranking update, and no ranking system is named anywhere in it. "It is critical" is Google's strongest word in the new text, and the page still carries its own disclaimer that rater guidelines do not directly influence ranking. The 2023 FAQ has not been updated with an October 2026 entry, so the two documents are three and a half years apart and should not be read as one body.

And one thing I cannot verify at all. Whether any of this changes rankings. Nobody can, yet. It was published in October. Whatever effect it has will show up in the data long before it shows up in an announcement.

What I do now

Three things, and they take about a minute each.

1. I read the four fields before publishing, not after. Title, meta description, structured data, alt text. Every claim in them gets checked against the body of the article, because those are the four things Google shows and the four things I used to let a plugin write.

2. I stopped treating the 2023 FAQ as history. It is live guidance, it is linked from the current page, and it says the opposite of what most people believe. If I am going to quote Google on AI content, that is the page to quote.

3. I kept the diff. Two captures, a text strip, and a comparison. That took twenty minutes and it is the only reason I know about the metadata rewording. For a page that gets edited rather than announced, that is the only method that works.

The short version

Google has said since 2023 that AI content is not against its guidelines. On 1 October 2026 it added 347 characters saying you should fact-check what a model produces, and reworded two words that quietly made metadata part of that requirement.

No ban. No penalty. No ranking change announced. Just a page that changed, and a lot of writing about it by people who did not open the previous version.

Sources