AI content is text produced wholly or partly by a language model. Google does not penalise it as such, which makes the penalty claim one of the SEO myths still doing the rounds: none of its policies cites the origin of a text as a reason for demotion. But the question is badly framed, because the real risk is not a penalty. It is that the content gets indexed, collects a few impressions for two or three months and then disappears without ever bringing a click, on a results page where AI Overviews have already reduced the clicks available.
This article brings together four studies that measured the effect on large samples, explains why they reach conclusions that look opposite, and adds first-hand data from the blog you are reading.
What Google says, word for word
The official position sits in two documents. The first is the guidance on using generative AI content, which says that "using generative AI tools or other similar tools to generate many pages without adding value for users may violate Google's spam policy on scaled content abuse".
The second is the spam policies (last updated 28 August 2026), which define scaled content abuse as when "many pages are generated for the primary purpose of manipulating search rankings and not helping users".
The subject of the rule is scale, not the tool. A thousand pages written by hand to fill a site with variants of the same keyword fall under the same definition as a thousand pages generated in a batch.
Then there are the quality rater guidelines, updated in January 2025. They assign the lowest rating, "Lowest", when a page's main content is "copied, paraphrased, embedded, auto or AI generated, or reposted from other sources with little to no effort, little to no originality, and little to no added value for visitors". The passage was reported by Danny Goodwin on Search Engine Land. Note the sequence: automatic generation appears in a list alongside copy-and-paste and paraphrasing, and the condition that triggers the rating is the last one, zero effort. Raters do not give scores that feed into the algorithm, but the document is the most explicit description Google has ever given of what it considers poor content.
The studies contradict each other, and the reason matters
Anyone looking for confirmation will find both the figure that clears AI and the one that condemns it, often in the same week. Here are the four largest samples published so far, side by side.
| Study | Sample and method | Main finding |
|---|---|---|
| Ahrefs, July 2026 | 1 million pages from the top 10 results of 100,000 SERPs, June 2026; 150,000 with enough text for analysis. Proprietary detector | 5.3% of pages in positions 1-3 are entirely AI. The average share of AI text rises from 27.1% in first position to 30.9% in tenth |
| Semrush, April 2026 | 42,000 blog articles across 20,000 keywords, collected November 2025. GPTZero detector | Fully AI text holds first position in 9% of cases, human text in 80.5% |
| SE Ranking, March 2026 | 20 new domains, 100 AI articles each, 20 niches, no human editing, 16 months of observation | 71% of pages indexed in the first month. From the third to the sixth month only 3% remain in the top 100 results |
| Rankability, 2026 | 487 results on competitive keywords, two detectors (GPTZero and Originality) | 83% of first-page results are classified as written by humans |
Ahrefs and Semrush appear to say the opposite of each other. In fact they measure different things. Ahrefs, in a study by Ryan Law and Xibeijia Guan, looks at the share of AI text within each page and finds that almost all ranking pages are hybrids: in their sample of 900,000 English-language pages from April 2025, 74.2% already contained at least a trace of automatic generation and only 2.5% were entirely AI. Semrush, in Margarita Loktionova's study, measures a binary classification of the whole page instead, and the gap is concentrated entirely on first position: from fifth place down the difference narrows.
There is a caveat that both teams state and that is worth repeating. AI content detectors make mistakes. Search Engine Land says so explicitly when discussing the Semrush data: detection tools are notoriously inconsistent and can misclassify both human and generated text. Ahrefs adds that its own detection method will differ from the way Google detects it. These are studies on the probability that a text looks generated, not on the certainty that it is. If you are interested in how these tools work, we cover them in the guide to AI content detectors.
The data that explains the contradiction: the sixteen-month curve
The most useful of the four studies is also the least cited, because it is the only longitudinal one. Bogdan Babiak and the SE Ranking team bought 20 new domains, with no history and no backlinks, and published 2,000 generated and never edited articles on them, a hundred per domain, on low-competition informational queries across twenty sectors.
The first month looks like a success: 71% of pages are indexed, 122,102 impressions arrive, 80% of the sites rank for at least a hundred keywords. Between the third and the sixth month, 3% of pages remain in the top hundred results. By the sixteenth month the twenty sites have accumulated 1,092,079 impressions in total and 1,381 clicks. That is 0.7 clicks per published article, in a year and four months.
66.9% of the pages are still indexed. No penalty, no mass deindexing. As soon as something better arrives, those pages simply drop and do not come back. It is the difference between content that has been removed and content that is irrelevant, and in reports the second is much harder to see.
Indexed, ranked and clicked are three different things
Most discussions of the topic stop at the first step. But there are three steps, and AI helps a lot with the first, a little with the second and hardly at all with the third.
We can see it on our own site. The Italian Visilay blog grew quickly over the last year, with articles built on verified sources but with a fast production process. These are the real Search Console figures for pages under /it/blog/, traffic from Italy, over the three months from 27 May to 24 August 2026.
| Metric | Value |
|---|---|
| Blog pages with at least one impression | 38 |
| Total impressions | 6,827 |
| Total clicks | 12 |
| Average CTR | 0.18% |
| Average position | 50.8 |
| Pages that received at least one click | 2 of 38 |
Thirty-six pages out of thirty-eight collected zero clicks in three months. They were not penalised: they are indexed, they have measurable visibility, and they sit at position fifty. Nothing comes from there. The traffic that counts is concentrated on two URLs, and they are the two that answer a question on which we had something of our own to say.
Publishing this does not make us look good, but it is the most honest figure we can put in an article on this subject, and it is exactly the curve SE Ranking measured across twenty domains. Volume brings impressions. Impressions are not traffic. If you want to see the opposite case, the article on SEO for manufacturing describes a project in which a single keyword moved from position 88 to 2 with a share of voice of 25.55%: there, the work was not about volume.
The UK context behind the international data
Applying a statistic collected on US results pages to another market is a mistake that costs. The UK shares the language of most of these samples, so the studies above describe its SERPs more closely than they describe Italian or German ones. Adoption of AI inside British companies, however, is still uneven.
According to the ONS article Artificial intelligence in UK businesses: 2023 to 2026, published on 20 July 2026 from the Business Insights and Conditions Survey, the share of UK businesses with ten or more employees using at least one AI technology rose from around 12% in September 2023 to around 35% in June 2026. Among businesses with 250 or more employees it reaches 49%, among the smallest (0 to 9 employees) 28%. Large language models are the most used technology (18% of businesses), followed by visual content creation (16%).
Put in editorial terms: two thirds of UK businesses with ten or more staff still report no use of AI at all, while the English-language SERPs they compete on are already full of hybrid content from US publishers. In a niche where your direct competitors still publish little, AI-assisted content can still gain ground, but the competition for informational queries is not only local, and the window is narrower than the adoption figures suggest.
Where AI actually pays off in the workflow
In the Semrush survey of 224 professionals, 64% report a human-led workflow with AI assistance and only 19% say AI improves the quality of the text. 70% cite speed as the main benefit. The practical consensus among people who do the job is that the model is useful in the side tasks, not in the writing.
- Research and clustering: grouping hundreds of queries by intent is a job the model does in minutes and a person in days. The method stays the one described in the guide to keyword research, only the tool changes.
- Checking intent: having the top ten results summarised to understand what format searchers expect, before deciding your own. We have a dedicated guide to search intent.
- First drafts of descriptive sections: definitions, glossaries, explanations of stable concepts. These are the parts where there is nothing original to say, and the model writes them well.
- Structural review: asking what is missing compared with competitors is more useful than asking it to write.
- Translation and adaptation: with the limit that a straight translation of a piece written for another market almost never ranks without new evidence for the local one.
What the model cannot do is bring the raw material: a number measured on a real project, a screenshot, a mistake made and described. That is why the experience signals described in the E-E-A-T framework remain the only structural defence, and why the first question before publishing is not "how does it sound" but "what does it contain that is not available elsewhere". The full method is in the guide on how to create SEO content.
The second market: generative answers
One figure overturns the common intuition. Ahrefs analysed a million SERPs with AI Overviews and the pages cited as sources: 3.6% are entirely AI-generated, 8.6% entirely human, 87.8% hybrid. The correlation between the share of AI text and citation order is 0.017, in other words zero. The study, by Si Quan Ong and Xibeijia Guan, is available on the Ahrefs blog.
Generative systems do not discriminate between sources based on how they were written. They choose based on structure, clarity of statements and the presence of citable data. Which is also why a table in native HTML gets read and a table inside an image does not. Anyone working on visibility in generative answers already knows this: the rules are in the guide to generative engine optimisation and in the one on how to appear in ChatGPT answers.
How to tell if your content is holding up
Three checks in Search Console, to run every quarter on everything you have published in the last eighteen months.
- Impressions to clicks per URL. A page with hundreds of impressions and zero clicks is not a half success: it is at position thirty or fifty. It is in the same situation as the blog pages in the table above.
- Ninety-day trajectory of average position. The pattern to fear is the SE Ranking one: a good start, a slow and steady decline, no recovery. If you see it on many URLs at once, the problem is systemic and concerns the type of content, not the single page.
- Share of URLs with at least one click. It is the metric nobody looks at and the only one that tells you whether you are building an archive or a warehouse. If you publish thirty articles a year and two bring traffic, you have a topic selection problem, not a writing problem. The selection criteria are in the guide to blog SEO strategies.
There is an implication the industry rarely spells out. If AI content is neither punished nor rewarded, the marginal cost of producing one more page has collapsed for everyone, competitors included. The advantage no longer lies in publishing more than them, because they can too. It lies in publishing the few things they cannot write, and in removing from the site the ones anyone could. The second half of that sentence is the part almost nobody does. If you want to know which pages on your site fall into the second category, it is the kind of analysis we start every SEO project with.
Frequently asked questions
No. None of Google's policies cites the origin of a text as a reason for demotion. The spam policies target scaled content abuse, meaning many pages generated for the primary purpose of manipulating rankings, however they were produced. The Ahrefs study of July 2026 on a million pages finds no correlation between the share of generated text and position.
It has not been shown to do so reliably, and Ahrefs points out that its own detection method differs from Google's. Commercial detectors such as GPTZero and Originality make mistakes in both directions: Search Engine Land, discussing the Semrush data, describes them as notoriously inconsistent. The useful operational question is not whether Google can tell, but whether the text adds something that is not available elsewhere.
Google has not stated a threshold. Ahrefs data from June 2026 shows that 82.2% of pages in positions 1-3 contain less than 50% text classified as generated, and that pages with more than 80% AI text are indexed in 40.35% of cases against 49.28% for the rest. The share is not a rule, it is a consequence: mixed pages win because someone put something of their own into them.
Yes. In the Ahrefs analysis of a million SERPs with AI Overviews, 3.6% of pages cited as sources are entirely AI-generated and 87.8% are hybrid, with a correlation between the share of AI text and citation order of 0.017. Generative systems select on structure and citable data, not on who wrote the text. That is why a table in native HTML gets read while a table inside an image does not.
Measure them first. In the sixteen-month SE Ranking study, 66.9% of pages stayed indexed while bringing no traffic: they did no direct harm, they simply did not perform. The practical criterion is the share of URLs with at least one click in the last ninety days. Pages with zero clicks and an average position beyond thirty get rewritten with your own material or merged, and are deleted only when the topic is already covered better elsewhere on the site.
