What Google actually penalises in AI-written content

Google doesn’t penalise content for being written by AI. It penalises three named patterns, and only one of them mentions AI at all — as an example, not as the offence. The offence is scale without value. The policy text is short, it’s public, and reading it settles most of the arguments teams are still having about this.

Does Google penalise content for being written by AI?

No. The helpful content documentation addresses automation directly and the strongest thing it asks for is a disclosure you might consider making: if automation was used to substantially generate the content, think about saying so. That’s guidance about reader expectations, not a ranking consequence.

The word that does the work in the spam policy is not AI. It is many. Search the policy for generative AI and you find it twice, inside a list of examples under a heading about producing pages at scale. The tool is incidental. What’s being described is a volume of pages that exists to be found rather than to be read.

What does the spam policy actually say?

Three of the sixteen are about content production. Here they are in Google’s words rather than anyone else’s, because the paraphrases circulating are consistently stricter than the source.

Scaled content abuse. “Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users.” The listed example that mentions models reads: “Using generative AI tools or other similar tools to generate many pages without adding value for users.” Both halves are load-bearing. Many pages. No added value.

Site reputation abuse. “Site reputation abuse is a tactic where third-party content is published on a host site mainly because of that host’s already-established ranking signals, which it has earned primarily from its first-party content.” In November 2024 Google clarified that no degree of first-party involvement changes the third-party nature of the content. Editing a sponsored section doesn’t make it yours.

Expired domain abuse. “Expired domain abuse is where an expired domain name is purchased and repurposed primarily to manipulate search rankings by hosting content that provides little to no value to users.” This one has nothing to do with how the content was produced and everything to do with why the domain was bought.

When do quality raters mark a page Lowest?

Only when three things hold at once. The September 2025 search quality rater guidelines assign the Lowest rating where main content is auto or AI generated “with little to no effort, little to no originality, and little to no added value”. That is a conjunction. All three qualifiers have to hold, and the number of published summaries that quote the first clause then drop the other two is remarkable.

The practical consequence is that a page needs to fail on all three counts to land on the floor. Genuine effort with no originality isn’t it. Original data presented lazily isn’t it either.

Raters don’t act on individual pages, and no rater score demotes a URL. The guidelines are how Google describes the bar its systems are built to approximate, which makes them the clearest published statement of what the systems are trying to do — and no substitute for the systems themselves.

Is there still a helpful content update to survive?

There isn’t, and this trips up a surprising number of content plans. Google’s ranking systems guide states that in March 2024 the helpful content system evolved and became part of the core ranking systems. There is no separate classifier to recover from, no site-wide flag to have lifted and no recovery event to wait for.

What replaced it is less comfortable, because a signal folded into core ranking is continuous. A page or a site is being assessed on every crawl rather than on a schedule somebody announces. The upside is that improvement is also continuous, so there is no queue to join.

Is E-E-A-T a ranking factor?

Not directly, and Google's own documentation says so in writing.

Read the qualifier rather than the headline. “Not a specific ranking factor” is not the same as “does not matter”: the same sentence says Google uses signals that identify content showing it. What you cannot do is optimise a number called E-E-A-T, because there isn't one.

E-E-A-T is vocabulary from the rater guidelines. It describes what raters should look for when judging whether a page seems trustworthy, which means there’s nothing to optimise directly and no score to raise. What there is, is a set of artefacts raters can actually see, and those are worth building because a reader can see them too: a real byline attached to a real person, an author page that says what that person has done, citations to primary sources rather than to other blogs and a stated method for any figure you produced yourself.

Chasing the acronym produces author boxes with invented credentials. Building the artefacts produces a page somebody can check.

How do you tell which side of the line your plan falls on?

Run the rater guidelines’ three qualifiers against the plan rather than against a finished page, because the answer is usually visible before anything is written.

Effort. Could this page have been produced by someone who knows nothing about the subject, using only what is already on the first page of results? If yes, the effort is in the tooling and not in the page.

Originality. Is there one thing here that exists nowhere else — your own numbers, your own test, a measurement you took, a document you read that others summarised? A worked calculation counts. So does an observation you made yourself, which is why reading a competitor ad library properly produces publishable material and reading three articles about it does not.

Added value. Would a reader be worse off if this page disappeared? Not “would we lose traffic”. Would the reader lose something. A piece that walks through the arithmetic of break-even ROAS with the working shown does something a summary of five other posts cannot, and the difference is visible in about four seconds of reading.

Choosing which pages to write is the boring half of this and the half that decides the outcome. Pull the queries your competitors already rank for, subtract the ones you cover, then sort what remains by whether you have anything of your own to say about it. A keyword tool and a spreadsheet do that job. RIVYL’s SEO tooling does it against a competitor’s live pages instead of a keyword export, and neither of them tells you what to say, which is the only part that has ever mattered here.

What is genuinely dead?

Four things people still build for, none of which return anything.

FAQPage and HowTo rich results. FAQ rich results were restricted to government and health sites in August 2023, and HowTo was dropped entirely. The markup is harmless and it buys you nothing visual in the results any more.

llms.txt. Google does not support it. The comparison usually drawn is to the keywords meta tag: a file describing your site to a consumer that is not reading it.

Word-count targets. Google's helpful content guidance asks whether you are writing to a word count because you heard Google prefers one, and answers plainly that it does not. Length correlates with depth on subjects that are deep, and that is the whole of the relationship.

Chunking prose into Q&A blocks for machines. Google's own documentation says no special formatting or optimisation is needed for its AI features, which answers this by a different route. Clear headings over real prose are already extractable. There are more surviving practices in this shape than most teams expect, and the ecommerce numbers that do not survive checking covers the same failure mode outside search.

Do you need to optimise for AI Overviews?

Google says not. Its documentation on AI features states that there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimisations necessary. Taken alone, that is a vendor describing its own system, so it’s worth putting a third-party measurement next to it.

Ahrefs studied 863,000 keywords and around 4 million AI Overview URLs in March 2026 and found that 38% of AI Overview citations came from pages ranking in the top 10, down from 76% in a July 2025 study. Ahrefs themselves caveat the comparison: their parsing methodology changed between the two, so the datasets are not directly comparable and the drop should not be read as a clean trend line.

Even with that caveat, the single March 2026 figure carries a useful implication for a site with little authority. If most citations are not coming from top-10 pages, being cited does not require ranking first, and the thing that makes a passage citable is that it answers one specific question completely and on its own. That’s not a separate discipline with a new acronym. It’s the same advice about writing a section that makes sense when someone lands on it cold.

The version to remember

Google is not auditing authorship. It can’t, reliably, and it has said repeatedly that it isn’t trying to. What it is looking for is pages produced in volume with nothing in them that was not already somewhere else, which is a description of a business model rather than a description of a tool.

So the question to ask about any publishing plan is not whether a model was involved. It is whether the plan would still be worth executing if search did not exist. If the honest answer is no, the tool used to execute it was never the problem. That test is cheap, it runs before you commission anything, and it is the same discipline as refusing to spend against a number you have not checked.