AI-written content and Google: the policy, in Google’s words

• There is no penalty for AI assistance as such. Google has said it rewards quality content however it is produced.
• There is enforcement against pages produced at scale with nothing of their own. The policy is called scaled content abuse and it does not care who or what produced them.
• Google does not require you to disclose that a model helped. Its structured data has no field for it.
• A page written with a model, from your facts, edited by you, is an ordinary page and is judged like one.

Everything else on this page is the detail behind those four lines.

The policy that matters here is scaled content abuse, updated in March 2024. It describes producing many pages primarily to manipulate rankings rather than to help people, whatever produces them: automation, humans, or both.

Read that last clause twice. The 2024 wording deliberately closed the gap where a publisher could argue that a human had touched each page. A hundred thin pages written by freelancers at five pounds each are the same violation as a hundred generated overnight.

Two neighbouring policies come up in the same conversations. Site reputation abuse is third-party content hosted on a strong site to exploit its ranking. Expired domain abuse is buying a domain for the links it earned under a previous owner.

Any advice you read that dates from before September 2023 and says Google detects and demotes AI writing was describing a rule that Google has since worded differently.

Google has said it does not use AI detection as a ranking signal, and that its systems judge quality. It is worth understanding why that is not a loophole.

A model produces text from patterns. Left to itself, it writes what the average page on the subject already says, because that is what it learned. The result is a page that repeats the top ten results with different words: no data of its own, no worked example, no opinion, nothing first hand.

That page fails on quality, and it would fail the same way if a person had written it from the same three sources. The model did not cause the problem. It made the problem cheap to produce at scale, which is exactly the pattern the policy names.

1. 1

   Fine: one page, your facts, your edit

   You write a brief for “how to invoice a client abroad”, list the facts your support team repeats, have a model draft the sections, then add the mistake your customers actually make and check every figure. The page holds something no other page holds.

2. 2

   Grey: forty pages from one template

   You generate a page per integration, each with the same four paragraphs and a swapped product name. It survives if each page carries real setup steps, real limits and a screenshot. It is scaled content abuse if the only difference is the name.

3. 3

   Over the line: a thousand pages nobody asked for

   You generate a page for every city, or every combination of two nouns, to catch long-tail queries. Nobody searching those queries is served by the result, and the family exists to rank. This is the pattern the policy describes.

The difference between the second and the third is not word count or tooling. It is whether the page could be written at all without the data or the experience behind it.

Google does not require AI disclosure. There is no “written by AI” field in its structured data, and no policy asking for a label. Its guidance is to use author bylines where readers would expect them, which on a product site means guides, opinion pieces and anything that makes a claim about experience.

The practical question is whether a reader would feel misled. A founder who signs a guide they briefed, edited and stand behind is not misleading anyone. A byline invented for a person who does not exist, complete with a generated portrait, is a trust problem whatever the tooling.

• Sign pages with a real person who can answer for them.
• Link the byline to a page that says who they are.
• Do not invent authors, credentials or portraits.
• Date the page, and change the date when the facts change, not to look fresh.

Before publishing anything a model helped with, answer these. If two answers are no, the page is not ready.

1. Does this page hold at least one thing that exists nowhere else: your data, your screenshot, your worked example, your opinion with the reason?
2. Is every figure on it traceable to a source I opened?
3. Would I send this page to a customer who asked the question?
4. If I published fifty more like it, would each one still pass the first question?
5. Is there a person whose name belongs on it?

The fourth question is the one that separates a library from a spam farm, and it is the one that gets skipped when the writing becomes cheap.

Two things can happen, and they are different. An algorithmic demotion happens inside the core and spam systems: nothing appears in Search Console, traffic to the family falls, and recovery usually waits for a later update after the pages are fixed or removed.

A manual action is a human review decision. It is listed in the Manual actions report with the type and the affected pages, and it is lifted by fixing the problem and filing a reconsideration request.

If a page family is demoted, the fix is rarely a rewrite of the same pages in the same shape. It is deciding which of them deserve to exist, making those genuinely useful, and removing the rest.