AI content and Google: the day the rule changed its name, not its target

The same question has come up in every meeting for three years: if we produce with AI, will Google penalise us? Almost every answer in circulation is wrong, and none of them carries a date.

So we stopped reading opinions and went looking for the texts. 254 network requests, zero failures: the full Search Central blog corpus since 2022 (166 posts listed, 166 retrieved), the 41 « Ranking » incidents on the Search Status Dashboard opened one page at a time, 9 documentation pages, both quality rater guideline PDFs (182 and 36 pages), and 1,772 Internet Archive captures queried by binary search across two doctrine pages.

From that, 34 quotes kept, each re-verified by string inclusion in the downloaded body. Nothing is quoted from memory.

The result fits in one sentence: no public Google policy targets the production method. And the real story is not where the market keeps looking for it. It fits in a single day, 5 March 2024.

What the rule says, word for word

The applicable rule is called Scaled content abuse. It is a section of the Spam Policies, and its defining sentence closes the debate on its own:

« This abusive practice is typically focused on creating large amounts of unoriginal content that provides little to no value to users, no matter how it's created. »

No matter how it's created. Generative AI then shows up only as the first example in a list of five: « Using generative AI tools or other similar tools to generate many pages without adding value for users ».

Two conditions stack up in that sentence, and the tool is not one of them: volume and absence of value. Ten AI-produced pages that are useful fall outside this section. A thousand hand-written pages that add nothing fall inside it.

The stated consequence is just as generic: « Sites that violate our policies may rank lower in results or not appear in results at all. »

On 5 March 2024, the rule changed its name in one day

This is the part nobody tells, because you have to dig it out of an archive.

Before 5 March 2024, the matching section was not called Scaled content abuse. It was called « Spammy automatically-generated content »: a rule about machines. Afterwards it is called « Scaled content abuse »: a rule about scale. And the previous wording simply vanished from the document.

The switch can be read to the hour. At 10:51 UTC the page still carries the old section. At 18:38 UTC it carries the new one. In between, at 17:02 and 17:03 UTC, the March 2024 core update and the March 2024 spam update start rolling.

Be precise about what the archive establishes and what it does not. It brackets the rewording inside a 7 h 47 window, and the launch of both updates falls inside that window. It cannot say which of the two came first. What it does establish is that they belong to the same day, and that Google's public doctrine on automated content rewrote itself in a matter of hours.

And here is what matters most: the change broadened the rule, it did not harden it against AI. The announcement post says so plainly.

« This new policy builds on our previous spam policy about automatically-generated content, ensuring that we can take action on scaled content abuse as needed, no matter whether content is produced through automation, human efforts, or some combination of human and automated processes. »

And two paragraphs later: « Our new policy is meant to help people focus more clearly on the idea that producing content at scale is abusive if done for the purpose of manipulating search rankings and that this applies whether automation or humans are involved. »

The market read March 2024 as a crackdown on AI content. The text says the opposite: the rule stopped talking about machines and started talking about scale and intent. Google even calls its own position an old one: « Our long-standing spam policy has been that use of automation, including generative AI, is spam if the primary purpose is manipulating ranking in Search results. »

One dedicated post, and it dates from February 2023

Across the 166 Search Central blog posts published since 2022, 34 mention AI, 8 pair AI vocabulary with a penalty term, and exactly one is dedicated to the subject: « Google Search's guidance about AI-generated content », published 8 February 2023, with 37 occurrences of AI vocabulary, far ahead of every other post.

Careful not to turn that into an accusation of silence. Google talks about AI constantly: three recent posts carry 20, 20 and 17 occurrences, in May 2025, December 2025 and June 2026. They simply talk about AI features in Search, not about penalising AI-produced content. On that specific point, the last post pairing AI vocabulary with penalty vocabulary dates from 5 March 2024.

The 2023 post has never been replaced. It says three things, head on:

• « Appropriate use of AI or automation is not against our guidelines. »
• « This said, it's important to recognize that not all use of automation, including AI generation, is spam. »
• « Using AI doesn't give content any special gains. »

The third one deserves a frame above every editorial calendar: AI does not get you penalised, but it does not buy you positions either. It changes production cost, not perceived value.

The reference documentation carries the same sentence, in the same terms: « If you use automation, including AI-generation, to produce content for the primary purpose of manipulating search rankings, that's a violation of our spam policies ».

Google asks its raters to grade without knowing

This is the document that settles the detection question, and it is public: the Search Quality Rater Guidelines, 182 pages, version dated 11 September 2025. Three passages are enough.

• « Likewise, the use of Generative AI tools alone does not determine the level of effort or Page Quality rating. »
• « Generative AI tools may be used for high quality and low quality content creation. »
• « Even if you are unsure of the method of creation, e.g. whether or not the page is created using generative AI tools, you should still use the Lowest rating when you strongly suspect scaled content abuse after looking at several pages on the website. »

Read the third one twice. Google explicitly asks raters to grade a page without knowing how it was produced. The trigger for the lowest rating is not « this was written by an AI », it is « I strongly suspect scaled content abuse after looking at several pages on the website ».

One honest caveat: those raters do not change any ranking directly. The document describes an intent, not an algorithm. But if a reliable AI text detector existed in-house, you would not ask thousands of human raters to judge blind.

There is no « AI content » manual action

The public list of Search Console manual actions carries no such label. The one that does exist is scaled content abuse, and the help page states who issues it: « Google issues a manual action against a site when a human reviewer at Google has determined that pages on the site are not compliant with Google's spam policies ».

A manual action is a human decision, not the output of a detector. The automated anti-spam system, meanwhile, is itself an AI: « For example, SpamBrain is our AI-based spam-prevention system. » So AI-produced content gets judged by an AI, with neither of them concerned about the production method.

A practical corollary, and it applies to every case study you will come across: nobody can prove a site was penalised « for AI content ». Neither Google nor the affected sites publish the label of a manual action. Any such claim extrapolates from a traffic curve. That is why this article names no penalised site.

The seven myths, line by line

Each myth is set against a sentence found as-is in the source, with its date and its URL. The verbatims stay in English: they are the texts Google published.

Only one of the seven deserves a caveat, and the embed carries it: « a page written by AI does not get indexed » is the only one whose refutation rests not on a Google statement but on a measurement by an SEO tool vendor. More on that now.

The only two large-scale public measurements

Ahrefs, on 27 July 2026, states in its method that it analysed « 1,000,000 pages pulled from the top 10 positions in 100,000 SERPs in June 2026 ». Keep that figure: the article's headline says 331k pages, the body says one million pulled, roughly 300,000 present in the crawl database and 150,000 substantial enough for detection. Quote the method, never the headline.

What it found: 9% of top-ranking pages carry at least 80% AI-classified content, and 5.3% are classified 100% AI. Average AI level moves from 27.1% at position 1 to 30.9% at position 10, 3.8 points across ten positions, with no statistical test published. That is a weak relationship, not a null one, and the nuance matters.

On indexation, the rate falls from 49.28% for low-AI pages to 40.35% for very-high-AI pages, which the author calls « a meaningful but far from disqualifying gap ».

Semrush, on 1 April 2026, appears to find the opposite: after grading 42,000 blog posts with a detector, « content classified as fully human-written outperformed content classified as AI-generated or mixed across all top 10 positions ».

The two coexist without contradicting each other. One measures that AI is not eliminated, the other that human still dominates. Neither establishes causality, and above all both rest on an AI text detector whose accuracy is not published. A text generated then rewritten by a human is classified human. Those percentages measure what the detector sees, not how the text was produced.

What is judged is what is served

Two in-house measurements cross-check all of the above, and they say the same thing by two different routes.

The first comes from our rendering audit: across 20 sites served to four User-Agents, 19 return exactly the same text to a browser, to Googlebot, to GPTBot and to ClaudeBot, and the gap between Googlebot and GPTBot is zero on all 20. Nothing in an HTTP response says how the text was written. What is judged is what is served. The method is detailed in our article on rendering and JavaScript.

The second comes from our E-E-A-T audit: across 170 articles measured at 17 French SEO agencies, 75% carry an author byline, but only 56% make that name lead to an author page, 71% cite an external source and 29% cite three or more. Google asks you to make clear who writes and how. The industry selling that advice applies it halfway. The full ranking is in our audit of E-E-A-T signals, and the definition in our glossary.

What to do, in order

1. Stop hunting for the detector. There is none in the published doctrine, and Google's own instructions ask raters to grade without knowing. Time spent « humanising » a text to fool a detector is time stolen from the content.
2. Count volume, not tooling. The useful question: how many pages did you publish this month, and how many carry something no other page already carries? That is exactly the pair the rule targets.
3. Answer the three questions the documentation asks when automation substantially generates content: is the use of automation self-evident to visitors, are you providing background on how it was used, and are you explaining why it was useful? The official wording on disclosure is careful: « AI or automation disclosures are useful for content where someone might think "How was this created?". Consider adding these when it would be reasonably expected. » Useful in one specific case, not mandatory everywhere.
4. Sign your articles, and make the byline lead to an author page that exists. 56% of the agency articles we measured do. It is the cheapest « who » signal to fix.
5. Cite verifiable external sources. 29% of the articles measured cite three or more. It is the most visible « how » signal in the HTML.
6. Never give an AI an author byline. Google explicitly discourages it: « Giving AI an author byline is probably not the best way to follow our recommendation to make clear to readers when AI is part of the content creation process. »
7. Build nothing special for Google's AI features. The dedicated documentation is categorical: « You don't need to create new machine readable files, AI text files, or markup to appear in these features. »

What we did not measure, and why

A documentary audit measures what Google publishes, not what Google does. None of the following appears in this article, and none of it should appear anywhere else without proof.

• The real effect of a penalty. You do not trigger a Google penalty to observe it, and no A/B test is possible on a ranking.
• The name of a site penalised for AI content. No public source supports that attribution.
• Quantified impact of the 2026 spam updates. Three launched this year, all recorded on the official dashboard with their bounds: March 2026 (24 March, 19 h 30), June 2026 (24 June, 2 days 1 h), August 2026 (18 to 21 August, 2 days 16 h). Google publishes their name, date and duration, and nothing else: no announcement accompanies any of the three. Our reading of the most recent one is in our March 2026 spam update analysis.
• The accuracy of AI text detectors. Neither Ahrefs nor Semrush publishes a confusion matrix.
• How widely these seven myths actually circulate. We refuted them; we did not measure who repeats them or how often.

Reproducing the measurement, and its two traps

The audit replays with no key and no paid tool, purely on public pages: 254 requests, about five minutes. Two traps are worth knowing, and they apply to any measurement on Google documentation.

First trap: without the ?hl=en parameter, developers.google.com serves documentation in a language that changes from one call to the next. Portuguese, Spanish, Russian and Chinese were all served on the same passage during a single run. An earlier pass happened to land in English by luck, and would have produced quotes pulled from translations. A quote taken from a translated version is not the position Google published.

Second trap: no documentation page carries a usable modification date. The footer shows « Last updated » followed by the date you consulted it. Out of 9 pages measured, 0 carries a reliable date. That is why every doctrine date in this article comes from the Internet Archive, and why they are upper bounds.

What Vydera does with all this

This article was produced by a human team and AI agents. We say so, because Google's documentation asks you to explain the how and the why when automation substantially generates content, and because it is true.

The « how »: a script collects the public sources, downloads them, and verifies every quote by string inclusion in the downloaded body. The « why »: because one human does not read 166 blog posts, two PDFs totalling 218 pages and 1,772 archive captures in five minutes, and because no undated opinion is worth one sourced sentence.

That is exactly what the rule asks for: value, at a scale that stays defensible page by page. If you want that method applied to your editorial calendar, start here: content marketing.