Does Google penalise AI content? What one million ranking pages actually show

Ask most marketers whether Google penalises AI-written content and you will get a confident answer. Ask them for the study behind it and you will usually get a vendor blog post. There is now enough real data to settle the question, and the answer is more interesting than either camp wants it to be.

What Google’s policy actually says

Google published its guidance on AI-generated content on 8 February 2023, and the position has not changed since. It did not ban AI content. The stated test is whether content is produced primarily for search rankings rather than for people, the same test that has governed since the helpful content system launched in August 2022.

The March 2024 core update is where the mechanics changed. Announced on 5 March 2024 and completed on 19 April, it ran 45 days, one of the longest core updates on record, and it did two structural things: it folded the standalone helpful content classifier into core ranking, and it introduced three new spam policies.

The relevant one is scaled content abuse, defined as “many pages generated for the primary purpose of manipulating search rankings and not helping users.” It replaced the older automatically-generated content policy, and the replacement deliberately removed the method of production from the test. Google’s own framing at the time targeted content produced at scale to boost rankings “whether automation, humans or a combination are involved.”

Google forecast the March update would reduce low-quality unoriginal content by 40%, and later reported achieving 45%. The September 2025 update to the Search Quality Rater Guidelines expanded the treatment of scaled low-value content and named generative AI as a possible source, while keeping the test author-neutral: pages with “little to no effort, little to no originality, and little to no added value” get the lowest rating regardless of who or what made them. In May 2026 Google extended the spam policies formally to cover AI Overviews and AI Mode, adding no new rules but applying the existing ones to generative surfaces.

What the data shows

Ahrefs published the largest study on this question in July 2026: one million pages drawn from top-ten positions across a hundred thousand search results, with roughly 150,000 carrying enough text to score for AI content.

The finding that settles the penalty question is the distribution across positions. Pages with 80% or more AI content appear at a nearly flat rate of roughly 8.4% to 11.7% at every position from one to ten. Average AI content rises only slightly down the page, from about 27.1% at position one to 30.9% at position ten.

If Google were penalising AI authorship, you would see a gradient. There is no gradient. There is a very weak slope that is fully explained by editorial investment correlating with rank.

Indexation tells the same story with slightly more edge: low-AI pages were indexed at around 49% and very-high-AI pages at around 40%. A nine-point gap is a real signal about quality correlates. It is not exclusion.

A companion Ahrefs study of 900,000 newly created pages found that 74.2% contained some AI-generated content, but only 2.5% were purely AI and 25.8% purely human. Almost 72% were mixed. The dominant production mode on the web in 2026 is hybrid, which is worth remembering the next time someone frames this as a binary.

So what did get deindexed?

Something clearly happened in March 2024, and plenty of sites lost everything. Understanding what they had in common matters more than the headline.

The most-cited dataset comes from Originality.ai, which checked around 79,000 sites and found roughly 1,446 with manual actions, about 1.9%, and reported that all of the deindexed sites they examined in detail showed signs of AI use.

That study should be read carefully, for two reasons. It comes from a company selling AI detection, and it has no control group. Since roughly 74% of new pages contain some AI content, “100% of deindexed sites showed AI signals” is close to saying “100% of deindexed sites were websites.” Without measuring the AI rate among sites that were not penalised, the study establishes correlation against a near-universal base rate.

The enforcement that is well documented is the site reputation abuse crackdown, and it is instructive precisely because the targets were not AI content farms.

These are major publishers with editorial staff. The violation was structural: third-party commercial content published on a host domain to borrow that domain’s ranking signals. Manual actions targeted specific subdirectories, not whole domains, and the publishers’ editorial sections were untouched. As of 2026 this enforcement remains manual rather than algorithmic.

The lesson is that Google’s actual enforcement in this period was about structure and intent, not authorship, and the biggest casualties were human-written.

The thing that is actually destroying your traffic

While the industry argued about AI penalties, search referral traffic collapsed for an unrelated reason.

Pew Research ran the cleanest study, published 22 July 2025: 900 US adults with tracking software, 68,879 unique Google searches captured in March 2025. When an AI summary appeared, users clicked a traditional result in 8% of visits. Without one, 15%. They clicked a link inside the AI summary in just 1% of visits. Sessions ended after the search 26% of the time with a summary present, versus 16% without.

Ahrefs measured the same effect against Search Console data across 300,000 keywords and found position-one click-through reduced by about 58% when an AI Overview is present. Their April 2025 measurement of the same effect was 34.5%. The damage roughly doubled in eight months.

SparkToro and Similarweb, using US clickstream data from January to April 2026, put zero-click searches at 68.01%, up from 60.45% in 2024 and around 45% a decade earlier. That 2024-to-2026 move is the fastest acceleration of the trend in ten years.

That last figure matters for how you allocate attention. AI Mode was a rounding error in early 2026 while AI Overviews appeared on over a fifth of searches. The traffic loss is overwhelmingly from Overviews.

One counterweight before anyone abandons search entirely: Google still sends vastly more referral traffic than ChatGPT, Gemini and Perplexity combined, by some measures a couple of orders of magnitude more, with AI platforms sitting around a tenth of a percent of total web referrals. Search is shrinking, not gone, and the alternative channels are not yet a replacement.

What this means practically

Using AI to write is not the risk variable. Undifferentiated volume is. This is the same structure as every other enforcement regime. Google penalises low-value content at scale produced to extract value from a channel, and AI is simply the cheapest way to produce that. A hundred thin pages get penalised whether a person or a model wrote them. Ten pages containing something nobody else has do not.

Rewriting does not launder a policy violation. The scaled content abuse policy explicitly names “automated transformations like synonymizing, translating, or other obfuscation techniques.” Running a thin page through a humanizer produces a rewritten thin page. We went through the mechanics of that in our piece on humanizers.

Detection scores are not a ranking input. There is no published evidence that Google detects AI authorship as a ranking signal, and the Ahrefs positional data is inconsistent with it doing so. Third-party AI detectors are useful for internal quality control. Knowing what share of your agency’s output is unedited model text is worth knowing. But they are not measuring anything Google measures.

Plan for the click you will not get. If two-thirds of searches end without a click, content whose only business model is a search visit is in structural decline. What still earns the click is what an AI summary cannot reproduce: original data, first-hand experience, a defensible position, and specifics that exist nowhere else. That has always been the advice; the difference is that the penalty for ignoring it is now immediate.

E-E-A-T is not a score. Google has said repeatedly that it is a concept used by human quality raters, not a ranking factor. The strongest signal in the Ahrefs data was not authorship at all; it was content investment. Pages under 50% AI hold 82% of top-three positions, and low-to-moderate AI pages earned two to three times the impressions, which is a story about editorial effort, not about detection.

The short version

Google does not penalise AI-written content. It penalises content produced at scale with intent to extract search value and no value returned, and it has said so in author-neutral language since March 2024. The data shows AI-heavy pages ranking at every position on page one at roughly the same rate.

Meanwhile the actual crisis, a two-thirds zero-click rate and position-one click-through halved where AI Overviews appear, has nothing to do with who wrote your content and everything to do with whether a summary can replace it. Publishing more machine-written pages is a direct accelerant of that problem, because the more your page reads like a summary, the more completely a summary replaces it.

If you want help building content operations that survive this, that is part of what we do.