Does Google Penalize AI Content? What Actually Gets Sites Deindexed

The most-cited number in this entire debate is a sample of 14 websites.

You have almost certainly seen it: 100% of websites deindexed by Google's update showed signs of using AI. It gets quoted as proof that Google hunts AI-generated text. It launched a thousand blog posts, several product categories, and a persistent low-grade panic in marketing departments.

The underlying research is real, and the people who ran it were honest about its size. The AI-detection portion of that study looked at 14 deindexed websites, 7 of which had AI content in over 90% of their sampled articles. The industry kept the headline and dropped the sample size.

It also answers the wrong question. This post is about what the evidence actually supports, checked at primary sources: what happened in March 2024, what ranking data shows in 2026, what Google's policy literally says, what changed in enforcement this year, and the publishing standard that comes out the other side.

The Stat That Broke the Industry's Brain

Start with what genuinely happened, because it was not nothing.

In March 2024, Google shipped a core update alongside three new spam policies. Ian Nuttall was monitoring 49,345 websites at the time. 837 of them, about 1.7%, were completely deindexed, removed from the index rather than merely demoted. Those sites accounted for more than 20.7 million organic visits per month and an estimated $446,552 in monthly display advertising revenue. Real businesses, gone in a week.

Originality.ai then examined deindexed sites for signs of AI generation and reported that 100% of them showed it. That finding is where the panic came from, and it has a structural problem that no amount of sample size would fix.

The study sampled only sites that had already been deindexed. That measures the probability that a site used AI given that it was deindexed. The number a publisher actually needs is the reverse: the probability of being deindexed given that you use AI. Those are different quantities, and the first tells you almost nothing about the second unless you also know the base rate of AI use among the sites that survived. The study never measured that.

The gap matters enormously here, because the base rate is huge. Graphite's analysis of 65,000 URLs drawn from Common Crawl found that more than half of new web articles have been primarily AI-generated since November 2024. If AI has touched a majority of everything published, then finding traces of it on 14 penalized sites is close to what chance alone would produce.

None of this means those 837 sites were unlucky innocents. They were mostly doing something genuinely bad, and AI was how they did it at volume. But "AI was the instrument" and "AI was the trigger" are different claims, and only the first one is supported.

What the Ranking Data Actually Shows

The stronger evidence comes from studies that look at the whole distribution rather than only the casualties.

Ahrefs ran the largest of these in June 2026. They pulled 1,000,000 pages from the top 10 positions across 100,000 search results, of which roughly 300,000 were in their crawler database and 150,000 had enough text to analyze. Their findings are hard to reconcile with an AI penalty:

Look at the last two rows of the position data. If Google were detecting and demoting AI text, the AI share would climb steadily as you move down the results page. It climbs 3.8 percentage points across ten positions. That is the signature of a weak correlation with something else, not of a filter doing its job.

Ahrefs' own conclusion: "Google is not against AI content; it is against bad content, but confusion arises because AI content and bad content overlap a significant amount of the time." To their credit, they also state the limit of their method plainly: "AI content detection is not perfect, and the way we detect AI content will be different from how Google does (if it does)." Every AI-detection study in this space, including the one in the previous section, inherits that caveat.

Now put Graphite's two numbers side by side, because together they are the most useful pair of figures in this entire discussion. Graphite used Surfer's detector on 65,000 English-language Common Crawl URLs, classifying an article as AI-generated at a 50% threshold, with a measured 4.2% false-positive rate.

• More than 50% of new web articles are primarily AI-generated.
• Only 14% of articles ranking in Google Search are. 86% are human-written.
• For AI answer engines the split is similar: 82% of articles cited by ChatGPT and Perplexity are human-written.

That gap, between writing the majority of new content and winning a small minority of the rankings, is the crackdown people are describing. But it is not a penalty. Nothing is being deducted. An enormous volume of AI content is simply failing to clear a quality bar that was always there, and failing in a way that is now statistically visible because the volume is so large.

The practical translation: your odds are not set by whether a model touched the draft. They are set by whether the page clears the bar. Most AI content does not, because most AI content is produced by people whose goal was volume.

What Google's Policy Actually Says

It is worth reading the actual text, because it is more specific than the commentary suggests.

From the spam policies: "Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users." The named example is "using generative AI tools or other similar tools to generate many pages without adding value for users." Critically, the policy applies whether the pages were made by automation, by people, or by both. Production method is explicitly not the trigger.

From Google's February 2023 guidance, which has never been retracted: "Rewarding quality content, however it is produced, has been core to Search for many years." And: "Appropriate use of AI or automation is not against our guidelines. This means that it is not used to generate content primarily to manipulate search rankings." Danny Sullivan has restated the position consistently since: "We focus on the quality of content, not how content is produced."

The January 2025 update to the Search Quality Rater Guidelines is where people usually find their smoking gun, and it repays careful reading. Section 4.6.6 instructs raters to assign the Lowest rating when all or almost all of the 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 to the website."

Read the sentence structure rather than scanning for the acronym. "Auto or AI generated" sits in a list of production methods, and the whole list is then qualified by three conditions that must hold: little to no effort, little to no originality, little to no added value. Reposting human-written content with no effort fails the same test. The guidelines say directly that the use of generative AI alone does not determine the effort or the page quality rating.

Those three words are the entire game. Effort, originality, added value are the variables Google is describing, and none of them is a property of the tool.

Two related policies from the same March 2024 batch are worth knowing because they catch different people: expired domain abuse (buying a domain for its residual authority and repurposing it) and site reputation abuse (hosting third-party content on your domain mainly to borrow your ranking signals). More on that second one shortly, because it has produced the most famous casualties by far.

What Actually Changed in 2026

Here is the part that has not been written up anywhere, and it comes straight from Google's own Search Status Dashboard.

Compare that to 2025, which had exactly one spam update: it launched 26 August and took 26 days and 15 hours to finish.

Two things changed, and both are measurable rather than inferred. The cadence tripled, from one spam update a year to three in the first eight months. And the rollout duration collapsed, from over 26 days to under 3.

The second change is the more informative one. A spam update that completes in 19 hours and 30 minutes is not crawling the web and forming fresh judgments. It is publishing a classification that had already been computed. Enforcement has moved from a slow, visible event to something closer to a continuous background process that gets flushed to production every few months.

For anyone running a content program, that has a concrete consequence: the lag between publishing something that violates the policy and having it evaluated is shorter than it used to be, and shrinking. The "publish aggressively and see what sticks" strategy always carried risk, and it now carries that risk on a much shorter fuse.

One honest qualification. Google did not announce new spam policies alongside the June or August 2026 updates, and both Search Engine Land and Search Engine Journal reported the August update as a routine one. So this is a change in enforcement tempo, not a change in the rules. The rules are still the ones quoted in the previous section.

A Note on Where These Numbers Come From

While researching this post, we kept running into confident figures: sites publishing AI pages saw 50 to 80% traffic drops after the March 2026 update, AI content farms lost 60 to 80% of their traffic. They appear across many pages, phrased almost identically.

None of them name a sample size, a methodology, a date range, or a traffic data source. We could not trace a single one to a study. They may be directionally right, and we have deliberately left them out, because a number you cannot trace is not evidence.

The irony is instructive rather than funny. A large share of the pages currently ranking for "does Google penalize AI content" are themselves the thing the question is about: confident, well-formatted, unsourced, and assembled quickly. The topic has attracted exactly the kind of content it warns against.

That also happens to be the most reliable diagnostic available to you. If your page makes a numeric claim that cannot be traced back to whoever generated the number, that is the tell. Not the model that wrote the sentence.

Manual Action or Algorithmic Demotion

If your traffic has already dropped, this is the first question to answer, and skipping it is why so many recovery efforts go nowhere. Two entirely different mechanisms produce similar-looking charts and demand opposite responses.

A correction worth making explicitly, because a lot of guides get it wrong: Google's manual actions report does not list a manual action type called "Scaled content abuse." Readers go looking for that label and conclude they are fine when they do not find it. The labels that actually appear for this class of problem are "Thin content with little or no added value", "Major spam problems", and, for third-party content hosted on your domain, "Site reputation policy". Practitioners including Glenn Gabe have documented scaled-content and AI-content cases in detail, but the string in the report will be one of the above.

Site reputation abuse deserves its own note, because it has claimed the biggest names. Enforcement began on 7 May 2024, and on 19 November 2024 Google closed the loophole publishers had been using, stating that third-party content deployed to exploit a host's ranking signals violates the policy "regardless of whether there is first-party involvement or oversight of the content." Forbes Advisor, CNN Underscored, Fortune Recommends, WSJ Buyside and Men's Journal were all hit. This policy remains manual-action-only rather than algorithmic. If you rent out a subdirectory to an affiliate operator, this is your exposure, and it has nothing to do with whether the content is AI-written.

For a manual action, the recovery path is defined: fix the underlying problem properly, then submit a reconsideration request with specific examples of the bad content you removed and the good content you added. Google's own wording on timing is that "most reconsideration reviews can take several days or weeks." For an algorithmic demotion there is no queue to join and no request to file. You clean up and wait for a core update to re-evaluate the site, which given the 2026 schedule means months.

The Publishing Standard That Survives

Everything above converges on three variables from the rater guidelines: effort, originality, added value. Here is how to make each one checkable rather than aspirational.

Match publishing velocity to editorial capacity. This is the ratio that gets sites in trouble, and it is not pages per month. It is pages per month divided by the number of qualified people who can genuinely stand behind them. A two-person team publishing 200 posts a month is making a claim nobody can back. Google's policy targets pages generated "for the primary purpose of manipulating search rankings," and output that no informed human reviewed is the clearest available evidence of that purpose. Set a rate your review capacity sustains, then hold it even when a tool makes exceeding it trivial.

Apply an added-value test before publishing. One question, answered honestly: what is on this page that is not already on the first page of results for this query? Acceptable answers include original data, first-hand testing, a named practitioner's judgment call, a correction to something widely repeated, or a synthesis nobody has assembled before. Not acceptable: better formatting, more comprehensive coverage, or updated for 2026.

Source at the origin. Every number links to whoever generated it, not to a blog that cited it, and carries its sample size. When you cannot find the sample size, say so or drop the stat. Section one of this post is what that habit produces, and it took one afternoon.

Include experience that could not have been synthesized. This is the first E in E-E-A-T and the hardest thing to fake. Screenshots from your own account. Numbers from your own engagements. The thing that broke in production and what it cost. A model cannot generate this, because it never happened to a model.

Put a real name on it. Named accountability with a traceable track record, not "Admin" or a stock-photo persona.

Prune as seriously as you publish. If you already have an archive of thin pages, deleting or consolidating them is usually higher-leverage than writing new ones. People resist this because the pages feel like assets. Content that never ranks is not neutral; it is part of what a site-wide quality assessment is looking at.

And where AI is genuinely fine, which is most places:

The dividing line was never whether a model touched the text. It is whether a person with relevant knowledge made the decisions and can defend the result. That standard predates AI by about twenty years. What changed is only that clearing it is now a competitive advantage, because generating things that fail it became free.

How We Do This at Keplaris

We build AI systems and automation for a living, so we have no interest in telling anyone to avoid these tools. We use them daily. What we have learned building content and search infrastructure is that the leverage is in the parts nobody wants to automate: sourcing, judgment, pruning, and being willing to publish a correction when the popular number turns out to rest on 14 websites.

This post is the method, not a description of it. Every figure here was checked at its origin, two widely-repeated claims were corrected, one set of numbers was excluded for being untraceable, and the enforcement timeline came from Google's dashboard rather than from someone else's summary of it.

If you want the underlying mechanics of how search engines and answer engines actually evaluate content, we have written the technical versions: semantic SEO covers entities and topical authority, answer engine optimization covers how AI engines select sources, and generative engine optimization covers the peer-reviewed evidence on which tactics actually lift AI citations. If bot traffic is muddying your analytics while you try to diagnose a drop, agentic bot traffic covers that too.

If your traffic has fallen and you cannot tell whether you are looking at a manual action, an algorithmic demotion, a measurement artifact, or an ordinary seasonal dip, that diagnosis is a concrete piece of work with a concrete answer. Talk to us and we will go through it with you, in your own data.