What Google Actually Penalizes About AI Content

Google’s spam policy targets the purpose a page was built for, never the tool that typed it, and two tests drawn from Google’s own wording decide whether yours is safe to publish.

Stated plainly, the question is narrow: what does Google penalize about AI content. Google answered it in March 2024, when it named scaled content abuse and wrote that the rule applies “whether automation or humans are involved.”

That reframes the problem. AI is not a category Google grades. It is one of several ways to produce a page, and the grading starts only after the page exists.

The uncomfortable part is that the reframe cuts both ways. A page a person wrote by hand can fail the policy, an AI-assisted page can clear it, and the two tests below are what separate them.

Quality, Not Authorship: What Does Google Penalize About AI Content?

Google’s answer sits in one Search Central post, and it is unusually direct.

Google’s March 2024 announcement of the scaled content abuse policy states the rule in a sentence. Producing content at scale is abusive “if done for the purpose of manipulating search rankings,” and that “applies whether automation or humans are involved.”

The same post answers the AI question head on. Asked whether the policy changes how Google views AI content, Google replies that its “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.”

Two clauses carry all the weight there. Purpose decides whether a page is a violation, and production method is ruled out as the deciding factor. That narrows what gets judged rather than excusing anything.

What the Policy Says, Stripped Back

Four statements from that post carry the whole rule:

• The abuse is producing content at scale to manipulate rankings, not the software used to produce it
• It applies whether automation, human effort, or a mix of the two made the pages
• It replaced the older automatically-generated content rule because the origin of low-quality content is often unclear
• Sites that violate a spam policy may rank lower or not appear at all, and a manual action arrives as a Search Console notice

That last point matters for how founders talk about this. A spam policy violation and a bad week after a core update are different mechanisms with different remedies, and treating them as one thing is where most of the panic starts.

Everything after this section is one of the two tests, or a place the two get confused. Scaled Content Abuse Is About Quality, Not About AI reads that policy line by line; this page stays at the level of the decision.

Purpose Is the First Test, and It Ignores Who Typed the Words

The Purpose Test asks one question: was this page made to help the person who searched, or to occupy a ranking.

The policy defines the violation by the primary purpose of the pages, so purpose is what gets established first and everything else is weighed against it.

What Purpose Looks Like on the Page

Purpose is not a mood. It shows up in decisions a reader can reconstruct from the finished page, and three of those decisions do most of the work.

• Who the page answers: A page built for one specific question reads differently from a page built to hold a keyword slot, and the difference is visible in the first screen.
• What it would lose: If three near-identical pages could be merged with nothing lost, the purpose was the keyword rather than the reader.
• Why it exists offline: A page still worth publishing when it cannot rank has a purpose that survives an algorithm change.

Scale is where this gets misread, because volume itself is not the violation. Google’s position on scaled content is that the production method is beside the point: publishing without value is the problem whether a person, a script or a model produced the pages.

A second reason to take purpose seriously has nothing to do with policy. With fewer than a third of US Google searches sending a click anywhere in early 2026, the audience for a page that exists only to hold a keyword is already thin. That is a market condition, not a penalty.

Aleyda Solis turns the same pressure into a prioritisation question: an AI answer satisfies a reader who wants a quick explanation, and does not satisfy one who still needs proof, current data, a real comparison, or somewhere to take the next action.

That distinction is exactly what the first test measures.

Purpose gets a page to the starting line. What keeps it there is the second test.

Evidence Is What a Rater Is Actually Told to Measure

The Evidence Test asks what a page proves once it exists, and Google hands its raters the vocabulary for it.

Google’s rater instructions make the standard explicit. The guidelines tell raters that pages “made up of content created at scale with no original content or added value for users, should be rated Lowest, no matter how they are created,” and that “the use of Generative AI tools alone does not determine the level of effort or Page Quality rating.”

Time spent is not effort. Proof left on the page is.

What Counts as Proof

Four things carry the weight, and a draft carrying none of them could have come from anywhere:

• A number, date or quantity that is specific rather than approximate
• An example that could only come from doing the work, not from summarising other people’s
• A table, chart or diagram built for this page instead of borrowed
• A stated position, including what you would not recommend and why

Google folded the same thinking into the instructions themselves. The rater guidelines were updated to address AI-generated content directly, allowing AI tools while requiring that the page provide unique value.

The same logic explains why the on-page factors a search engine can assess are the only place proof can live. Whatever a page demonstrates, it demonstrates in content, structure and sourcing that a crawler and a rater can both read. Google’s E-E-A-T shorthand, experience, expertise, authoritativeness and trust, is that same idea in four words.

Writing for E-E-A-T takes that apart at page level, and The Evidence of Effort a Reader Can Actually See ranks the evidence types by what a reader notices first. Use AI to Write Without Sounding Like AI handles the sentence-level version of the same problem.

A draft can clear the Purpose Test and fail this one with nobody noticing, because nothing about it is wrong. It could simply have been written by anyone, about anything, for anyone.

Scraping, Stitching and AI Sit in One List for a Reason

AI appears in the spam policy as a peer of two much older tactics, and reading the three in order explains why.

1. Scraping republishes someone else’s page with light changes and no added value.
2. Stitching assembles paragraphs from several pages into one that reads as new and adds nothing.
3. Generative drafting produces a plausible page from patterns in existing content, and fails the same way when nobody edits it toward a point of view.

The common failure is not the mechanism. It is that a reader finishes the page holding nothing they could not have found on the page it came from, which is why duplicate and near-duplicate content was a quality problem long before generative tools existed. Same failure, older tools.

Why the Fix Is Not Using AI Less

So “use AI less” is the wrong correction. The correction is the one that would have applied to a person hand-publishing two hundred thin pages in 2015: every page has to add something the last one did not.

Volume carries a mechanical cost as well as a policy one. Crawl economics explain why mass programmatic AI pages collapse: a site that ships thousands of near-identical pages spends its crawl and indexing allowance on pages that never earn it back.

Preparation is what makes volume survivable. Seer Interactive reaches the same correction from the production side, putting a subject-matter expert’s insight into the draft because models drawing on the same sources produce material that resembles what already exists. I would put it more bluntly: a model cannot supply a fact nobody went and got.

Cadence follows from the same logic, which is the argument in how many blog posts you need to rank, and the gathering itself belongs in the brief, as how to write a content brief sets out.

Not sure whether your library passes either test A free diagnosis reads your published pages against the Purpose Test and the Evidence Test, then names the ones carrying real risk. You get the list for your own site instead of a general standard. Book a free diagnosis

Neither 2026 Update Named AI, and the Log Is Public

Both confirmed 2026 ranking updates sit in Google’s own ranking-update history, and neither is recorded as an action against AI.

The March 2026 core update began on 27 March 2026, completed on 8 April 2026, and ran 12 days and 4 hours. The August 2026 spam update began on 18 August 2026, completed on 21 August 2026, and ran 2 days and 16 hours. Those are the dates and durations Google’s log carries, and the log carries nothing else about either one.

What a core update actually does is set out separately. Google’s core-update documentation says they are “designed to ensure that overall, we’re delivering on our mission to present helpful and reliable results,” and that the changes “are broad in nature, and don’t target specific sites or individual web pages.”

What Google Said About Each One

• Google’s log records the March 2026 core update as a core update and nothing further, with a start date, a completion date and a duration
• Google described the August spam update as a normal spam update rolling out for all languages and locations
• Neither entry names AI-generated or AI-assisted pages as a target

I want to be precise about the limit of that claim. It is a statement about what the published entries say, not a claim about what Google’s systems detect internally, which I have no visibility into.

Thin output published at volume stays exposed to both kinds of update, because it fails the Purpose Test the way any thin content does. What the public record does not support is the stronger claim that either update went after AI as a category.

For anyone deciding whether to keep drafting with AI, the two tests are the stable target. Guessing at the next update is not.

Disclosure Belongs With Counsel, Not With a Ranking Policy

Whether you must tell readers an article was written with AI is a legal question, and it is not one I will answer here.

Google’s own position is a recommendation rather than a requirement. Nothing in the spam policies conditions a ranking on a disclosure statement, and where Google addresses disclosure at all it frames the question as what a reader would reasonably expect. That is a trust argument, not a compliance rule.

Three Questions, Three Different Authorities

• The ranking question: Settled by the two tests in this article, both drawn from Google’s published wording.
• The disclosure question: Sits with regulators and courts in specific jurisdictions, and those rules are moving.
• The right adviser: Qualified counsel who can read the rule that applies where you operate, rather than an SEO article.

The OECD’s AI policy dashboard tracks how national frameworks are developing, and it is a neutral place to watch the shape of it. I take no position on what any of them require of anyone.

Running Both Tests Before You Publish Takes One Pass

Run both tests, because clearing one says nothing about the other.

Test	The question it asks	The check that exposes a fail
Purpose	Was the page built to help the person who searched, or to hold a ranking	Would the page survive being merged into three similar pages with nothing lost? If yes, it fails
Evidence	Does the page carry proof of effort a reader and a rater can both see	Would removing the byline change how much you trust it? If no, it fails

The first check is about redundancy and the second is about attribution. Neither needs a tool, and both are answerable in a minute on a finished draft.

The One-Pass Check

• Ask what a reader can do afterwards that they could not do before
• Find the one sentence a competitor could not publish unchanged, and write it if it is missing
• Confirm every number on the page has a source a reader can open

There is a reason to hold that bar higher than the policy strictly demands. Pew Research found that Google users click result links less often on searches that produce an AI summary, so a page now has to be worth arriving at rather than merely worth indexing.

The production version of this check, brief through publish, is how to produce content that ranks and gets cited, and content that ranks and gets cited by AI applies the same standard to pages that have to serve a citing engine as well as a ranking one.

Neither test cares whether AI drafted the first version. Both care about what is true of the page once it publishes.

Frequently Asked Questions

What does Google penalize about AI content?

Nothing, for the use of AI alone. Google’s spam policy penalizes pages generated primarily to manipulate search rankings rather than help users, and states the violation applies no matter how the content was created. A page fails on purpose and value, not on the tool that drafted it.

Is a Google penalty the same as a ranking demotion?

No. A penalty, which Google calls a manual action, is a human reviewer applying a listed spam policy to a site. A demotion is an algorithmic reassessment of the kind a core update applies. Thin pages usually lose ground through the second, not the first.

Can AI-assisted content rank well in Google Search?

Yes, when the page clears both tests. Google’s long-standing policy is that automation, generative AI included, counts as spam only when the primary purpose is manipulating rankings. A helpful, original page is assessed on what it delivers, not on what produced the first draft.

Do I need to disclose that an article was written with AI?

That is a legal question rather than an SEO one, and disclosure rules are actively changing across jurisdictions. This article takes no position on scope or duty. Google recommends disclosure where readers would reasonably expect it, but check with qualified counsel for the rule that applies where you operate.

How does a core update differ from a spam update?

A core update reassesses how helpful and reliable content across the index is, so pages move up as well as down, and it targets no specific site. A spam update enforces the practices named in the spam policies. Google logs both in its ranking-update history with dates and durations.

Which checks should run before an AI-assisted draft publishes?

Two. Ask whether the page would survive being merged into three similar pages without a reader losing anything, which tests purpose. Then ask whether removing the byline would change how much you trust it, which tests evidence. Failing either one is enough to hold publication.

Continue Reading

More From This Series

• The Evidence of Effort a Reader Can Actually See
• The AI Tells That Cost You Readers, Not Rankings
• Scaled Content Abuse Is About Quality, Not About AI
• The Content Audit That Produces Your Next Ideas
• Making AI a Writing Partner That Earns Its Place

More from TDM Insights

• Writing for E-E-A-T: Prove Experience on the Page
• Use AI to Write Without Sounding Like AI
• How Many Blog Posts Do You Need to Rank?
• How to Write a Content Brief (That Gets It Right)
• How to Produce Content That Ranks and Gets Cited
• Content That Ranks and Gets Cited by AI

Explore TDM Insights Topics

• SEO
• AEO
• Content
• Case Notes