Does Google Penalize AI Content?

AI now writes about half of everything published on the web, but only seven per cent of what ranks first.

That isn’t a contradiction. It just isn’t a penalty.

What Google Actually Said, In Order

Google’s position changed quickly. In April 2022, it treated automatically generated content as spam. By February 2023, it was saying the method didn’t matter; the quality did.

That shift is the key to the whole question. Google stopped judging content by whether a machine helped produce it and started judging what the content was for, what value it added, and whether it was being produced at scale to manipulate search results.

An “AI Penalty” Isn’t Something Google Has

Open Search Console and look at the manual actions report. There are sixteen things Google can flag you for.

“AI content” is not one of them. Neither is “scaled content abuse.”

Sites taken down for mass-producing AI articles get labelled Major spam problems or Thin content with little or no added value, categories that are older than generative AI and were written for people with article spinners.

That takes about thirty seconds to verify, and it is the single most useful thing you can know about this question. The thing you are afraid of is not a category Google possesses.

Which leaves the distinction that the rest of this post turns on.

A manual action is a penalty. A human at Google looked at your site, decided it broke a rule, and you get a message about it.

A core update demotion is not a penalty. Nothing looked at you, nothing decided anything about you, and you get no message at all. Your traffic is simply lower on Thursday than it was on Wednesday.

Almost everybody searching this question is worried about the first thing and exposed to the second.

50% Of Everything Published, 7% Of What Ranks

Graphite ran the two halves separately: one study sampled Common Crawl to measure what gets published, and a second sampled the results pages to measure what ranks. Sampling the crawl rather than the rankings sounds like a technicality and is actually the whole ball game. Every other study in this field samples pages that currently rank, and you can’t measure whether something gets removed by surveying the things that weren’t removed.

Half of what gets written. A seventh of what gets found. A fourteenth of what wins.

Something is removing AI content from search at enormous scale, and there is no penalty anywhere in the system doing it.

Ahrefs Published The Gap And Titled It “No Punishment”

The studies everyone cites aren’t lying. They are answering a narrower question than the one you asked, and the headline does the rest.

I don’t think Ahrefs is being dishonest. I think they mean a specific thing by “punish”: a manual action, a flag, a message in Search Console. By that definition they are completely right.

The reader searching this question means something else by it. They mean “will this cost me traffic.” And the same article says yes, by nine points of indexation and a factor of three in impressions.

The Studies Are Measuring A Detector, Not Your Writing

The more basic problem is the detector itself.

OpenAI launched its own AI text classifier in January 2023. It lasted six months. The company pulled it for its “low rate of accuracy”, despite having access to the models, the training data and the way the text was generated. OpenAI couldn’t reliably spot its own output.

The commercial detectors are trying to do the same thing with less information. In peer-reviewed testing of fourteen of them, the tools were “neither accurate nor reliable.” A little editing was enough to throw them off. Stanford researchers found another problem: they “consistently misclassify non-native English writing samples as AI-generated.”

Those mistakes aren’t evenly distributed. Detectors are suspicious of low perplexity: plain words, predictable sentences, familiar structure.

In other words, competent professional writing. The prose most likely to rank is also the prose the detector is primed to mistrust.

Now take the most repeated number in this debate: 100% of the sites deindexed in Google’s March 2024 update showed signs of AI content.

It sounds conclusive until you ask what the detector finds on sites that weren’t deindexed. The study had no control group. Ahrefs estimates that about 74% of new web pages contain detectable AI content. Test roughly 200 pages on each site against a base rate that high and you would struggle not to find at least one hit.

That 100% figure is what you would expect even if AI had nothing to do with the deindexing.

Nobody Tells You, And Most People Don’t Come Back

This is the part the reassuring articles leave out, and it’s the part that should actually govern your decisions.

Glenn Gabe tracked around 400 sites flattened by the September 2023 helpful content update. Eleven months later, 22% had recovered as much as a fifth of what they lost.

Not recovered. Recovered a fifth. Roughly one site in five.

The “you’d know because Search Console would tell you” reassurance has a hole in it too. Gabe documented a 20,000-URL site that was fully deindexed roughly a day before the notice showed up. And a core update demotion never generates a notice at all, because there is nothing to notify you about.

Danny Sullivan was asked directly whether sites get back to where they were. His answer is the truest sentence anyone at Google has said about this:

You can’t predict that every site will recover to exactly where they were in September because September doesn’t exist anymore.

Use My Content Pipeline Auditor

None of the above is an argument against using AI. It’s an argument about what the exposure actually is, and the exposure is a pipeline property rather than a writing-tool property.

Google does tell you what to do about it. Here is the whole of it, from the closing section of the same February 2023 post, signed by the two people whose names bracket this entire timeline:

Source: https://developers.google.com/search/blog/2023/02/google-search-and-ai-content, retrieved September 2026.

Be original. Be helpful. Ask who, how and why. Fine. But none of that gives you a test you can run. You can satisfy every word on paper and still publish the exact pages Google demotes. These six questions do.

Notice that none of the six mentions AI. That isn’t a rhetorical trick. Google’s actual policy doesn’t mention it either, except as one example among several of how you might produce content at scale.

The Thing That Will Actually Cost You

On August 30, 2026, Google stopped applying its site reputation demotions to users in the EEA after pressure from the European Commission. Its own policy page now says the consequences “vary depending on the location of users.”

The same page. Actioned for a reader in Ohio, untouched for a reader in Ireland.

Whatever a penalty is, it is a policy choice that a regulator can reverse, not a law of physics, and not a stable thing to build a content strategy against. Which is why “will I be penalized?” was always the wrong question. It asks about Google’s intentions when what you needed to know was the mechanism.

And the mechanism turns out not to need an opinion about AI at all. Google never built an AI detector. It built a proxy for what unedited AI output looks like: unoriginal, commodity, produced at volume, adding nothing that wasn’t already on the web. That proxy catches the same pages without ever having to know how they were made.

So no, nobody is going to penalize you for using AI. There is no flag, no message, no category, and nothing to appeal.

You will just be quietly outranked by someone whose page contains something a model could not have known.

Write that thing. The rest of it was never the risk.

The Questions Underneath This One

Can Google detect AI-generated content?

It doesn’t try. Google has never shipped an AI detector, and its own March 2024 policy says it often can’t tell whether a low-quality page was produced by automation. What it built instead is a proxy for what unedited AI output looks like: unoriginal, commodity, produced at volume. That catches the same pages without ever having to know how they were made.

Is AI-generated content against Google’s guidelines?

No. Google’s position since February 2023 is that appropriate use of AI or automation is not against its guidelines. The line it draws is purpose, not method: content generated primarily to manipulate search rankings is spam, whether a model wrote it or a person did.

Do I have to disclose that a page was written with AI?

There is no requirement, no markup for it, and no ranking effect either way. Google’s guidance only encourages disclosure where a reader would reasonably want to know how something was made. Disclose it because it’s true, not because you think it buys you something. It doesn’t.

Will AI content rank on Google?

It already does. About seven per cent of pages ranking first read as AI-generated, against roughly half of everything published, so it ranks at about a seventh the rate it gets written. Using AI earns no advantage of its own. Google’s own phrasing is that it’s just content.

Will Search Console tell me if I’ve been penalized?

Only for a manual action, and not always in time. One 20,000-URL site was fully deindexed roughly a day before the notice appeared. A core update demotion generates no notice at all, because nothing was decided about you.

If you’re working out what any of this means for how your business uses AI, that’s what I do. You might also want the piece on why AI watermarks can’t prove what people want them to prove, which is the same argument applied to authorship instead of rankings, or the guide to the tools worth using.

Every claim here was verified against the primary source in September 2026: Google’s own blog posts, spam policies and rater guidelines, the studies as published by Ahrefs, Semrush, Graphite and Originality.ai, and the peer-reviewed detection research. The one exception is flagged in the timeline. This moves fast. Check before you rely on it.