What Happened to Sites That Scaled AI Content, 2023 to 2026

• Google's published position since February 2023 is that production method is not the test. Using automation, including AI, mainly to manipulate rankings is.
• Detected AI share barely moves across the top ten in the one large sample reported (27.1% at position one, 30.9% at position ten), but only about 15% of that sample could be scored.
• The accounts of sites that fell are anecdotes or secondary reports. None names a site with dated before-and-after data that this page could check.
• Of nine pieces of evidence graded, two are strong (both are Google's own rules), three are fair, two are weak and two cannot be verified.

Did Google penalize sites that scaled AI content?

The evidence behind this page does not show a penalty for using AI, and it does show a spam policy that applies to scaled pages made mainly to manipulate rankings. The distinction matters because the question searchers type, whether there is an AI content penalty, bundles two different things.

Google's guidance on generative AI content says that using generative tools to produce many pages without adding value may violate the scaled content abuse policy. It does not say that using AI is itself the violation, and the supplied text contains no percentage threshold.

Rankability asserts that Google has applied manual actions to spammy AI-generated content, and Frase says sites with mass-produced articles lost visibility in March 2026. Neither names a site or gives a dated before-and-after figure, so both are claims to weigh, not records to rely on.

What did Google say about AI content, and when?

Google's core position was set out in February 2023 and has not been reversed in the pages supplied, with its generative AI guidance page last updated on 1 October 2026. The table lists every dated step in the evidence and says how firm each date is.

Google's February 2023 post says that about ten years earlier there were concerns about mass-produced human-written content, and that Google improved its systems to reward quality instead of banning human writing. Its stated reasoning applies the same logic to AI: reward quality, whatever produced it.

The evidence does not date the scaled content abuse policy or any specific enforcement action. A reader should not assume the policy existed in 2023 because the 2023 post talks about spam.

How was each piece of evidence graded?

Each piece of evidence was scored from 0 to 2 on four criteria and given a grade from the total of 8. The criteria, the scale and the bands are set out here before any grade is shown, so a reader can disagree with a score and see what it would change.

The bands are 7 to 8 Strong, 5 to 6 Fair, 3 to 4 Weak and 0 to 2 Unverified. Unverified does not mean false. It means the supplied text gives no way to check the claim.

Two limits apply. A grade is attached to one claim from one source, not to the publisher in general, and a Strong grade for Google's pages covers what the rule is, not how sites fared under it.

Which cases and studies hold up?

Only Google's own pages grade Strong, and they answer what the rule is, not what happened to sites. Everything about outcomes grades Fair or lower.

The grades separate evidence that looks similar on the surface. A vendor study and a vendor anecdote can both be called data, yet one states its tool, thresholds and exclusions and the other states nothing.

One of the nine is a named case. The reported SEO heist, in which a founder is said to have built about 1,800 AI articles from a competitor's sitemap, is covered in Kanso's sourced timeline of the SEO heist. That page found every figure came from one secondary source with undefined units and no baseline for the later fall, which is why it grades Weak here.

What do the ranking studies report about AI content in the top results?

They report that detected AI content appears across the whole first page, and they cannot show that Google rewards or ignores it. Two studies are in the evidence, both relayed by Frase and both resting on detection.

Frase reports that Ahrefs, in June 2026, sampled a million pages from 100,000 results pages and ran detection on the 150,000 with enough text. The average detected AI level was 27.1% at position one and 30.9% at position ten, and 5% of pages in positions one to three were detected as fully AI-generated. Frase adds that 27.5% of top-three pages sat between 20% and 50% detected.

Frase also reports a Semrush analysis from April 2026 of 42,000 blog posts. Pages classified as purely AI-generated took the top spot 9% of the time, and predominantly human-written pages held it 80% of the time.

Rankability's own study scored 487 top-ranking results with a blend of two detectors. It counts a page as human-written at a blended score of 30 or below, AI-generated at 70 or above and mixed in between, and reports that 83% of the pages it could score read as human-written. Rankability calls this a directional study with a small sample.

These three numbers measure different things. Ahrefs reports an average share of detected AI text per page, Semrush reports the share of position-one posts in each class, and Rankability reports the share of scored pages below a threshold. Comparing 27.1%, 9% and 83% directly would compare three units.

How much do the headline numbers in those studies leave out?

The arithmetic behind them shows how much of each study the headline number leaves out. The table derives six figures from the numbers the sources report, with the inputs shown.

150,000 ÷ 1,000,000 = 15%; 30.9 − 27.1 = 3.8 points; 487 × 0.83 ≈ 404; 100 − 9 − 80 = 11

The 15% is the largest gap. Every Ahrefs figure describes the pages with enough text to score, and Frase does not say how the other 850,000 differ.

On a plain reading, the Semrush figures describe who occupies one rank, not a win rate. Without the share of each class among all 42,000 posts, 80% at position one cannot be compared with a base rate, so it does not show that human-written posts are favoured.

Why can't detector data settle whether Google favours human writing?

Detector data cannot settle it because a detector reports how text reads, not who or what produced it. Rankability says this directly: detector output is classification evidence, not proof of who wrote a page or what will rank.

Frase draws the consequence for the other two studies. A draft that began as AI output and was then researched, edited and given a real voice may no longer read as AI, so it would be counted as human. On that argument the true AI share of winning pages is higher than the detected one, though Frase states this as reasoning, not as a measurement.

The same logic cuts against Rankability's conclusion. If 83% of scored pages read as human-written, that is consistent with Google favouring human writing and equally consistent with well-edited AI drafts passing as human. The evidence cannot tell the two apart.

Frase also gives a warning about its own process. An earlier draft of its article carried a Semrush figure that turned out not to appear in the source, traced to unverified AI-assisted research notes. That is the failure Google's guidance on generative AI warns about, and the reason every figure on this page is attributed to the text that reported it.

What happened to the sites that dropped?

The evidence describes what dropped sites had in common but supplies no checkable record of any one of them. This is the weakest part of the picture, and the grades say so.

Frase says that after the March 2026 core update, sites publishing mass-produced articles saw large drops. It lists what they shared beyond AI: hundreds of near-identical pages, no first-hand experience and nothing new compared with the first result. It also says plenty of sites using AI did not drop, a point it notes tends to get left out of the story.

Rankability describes two pages of its own. A page generated with ChatGPT for a local training keyword and scored as 100% AI was removed from Google's index after a recent update, was replaced with human-written content and was reindexed within hours and ranking in the top ten. A second AI-written page performed poorly for months until rebuilt by a human writer. Neither anecdote carries a date or a traffic figure.

SEO Sherpa claims that Bankrate and CNET used AI to scale content while keeping their rankings, but the supplied text gives no figures, dates or source. It grades 0. For the one named case with numbers, the reported SEO heist, the sourced timeline and its unverified gaps are on Kanso's page on that case.

What can be said fairly is that the three vendor accounts point the same way: scale without added value is what went wrong, not AI. That agreement comes from publishers who each sell a content or SEO product, and it is a pattern in claims, not a finding.

Is a traffic fall a penalty, a demotion or a page that is not indexed?

They are three different problems, and the evidence supplied does not say which one hit any of the sites above. A Reddit thread on scaled content observes that many people confuse not ranking with being de-indexed.

The distinction changes the fix. A page that is not indexed needs a crawl and indexing diagnosis, a page that ranks lower needs a quality diagnosis, and a manual action needs a different response again. Kanso's guide to finding which stage is broken covers the first, and its seven-step diagnostic for a traffic drop covers the second.

Google also says something relevant about raters. Its guidance page points to the Search Quality Rater Guidelines on scaled content abuse (section 4.6.5) and on main content made with little effort, originality or added value (section 4.6.6). It states that these guidelines are not a guide to ranking first and that raters' ratings do not directly influence ranking.

What survived from 2023 to 2026?

Google's purpose-and-quality test survived unchanged in substance, and so did the practice of using AI under human review. The pages supplied show no reversal of the February 2023 position.

Three things in the guidance carry the most weight for a content team. Google says generative models predict likely word sequences rather than retrieve facts, so all AI-generated content must be manually fact-checked before publishing, including titles, meta descriptions, structured data and image alt text. It says sharing how content was created can help readers, and its 2023 post advises accurate bylines and disclosures where readers would expect them, and says listing AI as the author is probably not the best way to do that.

The third thing is practitioner opinion, labelled as such. Frase argues that what separates AI content that ranks from AI content that sinks is decided before drafting: whether the page contains something not already in the results, whether real experience stands behind it, and whether it was built on current research instead of what a model already knew. Rankability argues for a human-led editorial process, and SEO Sherpa for human review, fact-checking and a byline.

What did not survive is the idea that the question is settled by counting. No source supplied gives a safe percentage, a safe publishing rate or a safe page count.

What is documented and what is analysis?

The first list holds what the supplied sources state, each attributed. The second holds what Kanso infers from them, which a reader should weigh as opinion.

• Documented (Google, February 2023): automation, including AI, used mainly to manipulate ranking violates spam policies, and not all use of automation is spam.
• Documented (Google, page updated 1 October 2026): generating many pages with generative tools without adding value may violate the scaled content abuse policy, and AI output must be fact-checked before publishing.
• Documented (Frase, relaying Ahrefs, June 2026): average detected AI share of 27.1% at position one and 30.9% at position ten, on 150,000 scored pages.
• Documented (Frase, relaying Semrush, April 2026): pages classified as purely AI-generated held the top spot 9% of the time and predominantly human-written pages 80%, in 42,000 blog posts.
• Documented (Rankability, 2026): 83% of the scored pages among 487 top results read as human-written on its detector blend, in what it calls a small directional study.

• Analysis (Kanso): the three ranking figures are in different units and cannot be compared directly.
• Analysis (Kanso): detector-based studies cannot separate Google favouring human writing from edited AI drafts passing as human.
• Analysis (Kanso): the accounts of dropped sites agree on scale without added value, but all come from vendors and none can be checked here.
• Analysis (Kanso): the grades and the five checks below are judgments built from the sources, not measurements.

What should a content team check before scaling AI drafts?

Five checks, each a question with a stop condition, cover the failure patterns the evidence describes. The table was built by taking Google's guidance and the traits Frase says dropped sites shared, and turning each into a question a team can answer before publishing.

For scoring a single draft, a twenty-point editorial rubric for AI-written drafts goes further, and a source-verification workflow for AI marketing content covers the accuracy check.

Kanso's Writer Agent turns Search Console opportunities into long-form drafts in your brand voice, for you to review and publish yourself. It does not publish for you, and Kanso does not promise rankings.

What does this page not show?

This page does not show the traffic history of any named site, and it does not show what Google does internally, because none of that is in the evidence. Six gaps are worth stating plainly.

• The Ahrefs and Semrush studies themselves were not read. Their figures are as Frase reported them.
• The date the scaled content abuse policy began is not in the supplied text.
• No supplied source gives dated before-and-after traffic for a named site that could be checked.
• No supplied source tests whether formatting, structure or tables change outcomes.
• The vendor accounts of the March 2026 core update cannot be tied to specific sites or pages.
• The 2023 to 2025 stretch has little in the evidence beyond Google's February 2023 post.

For the primary detail, read Google's own pages and the Search Quality Rater Guidelines sections it names. This page also does not walk through Google's spam policies clause by clause or set quality gates for programmatic pages, and Google's policy documentation is the place to start for both.

AI content and Google: frequently asked questions

Sources

• Google Search's guidance about AI-generated content, Google Search Central Blog
• Google Search's guidance on using generative AI content on your website, Google Search Central documentation
• Does Google Penalize AI Content? What the Data Shows, Frase
• Does Google Penalize AI Content? An SEO Study (2026), Rankability
• Does Google Penalize AI Content? What You Need to Know, SEO Sherpa