Google penalises purpose, not tools

Google does not have an AI content penalty. It has a scaled content abuse policy, and the thing that trips it is why the pages exist, not what typed them.

The spam policies doc, last updated 15 May 2026, defines scaled content abuse as many pages generated for the primary purpose of manipulating search rankings and not helping users. It explicitly lists generating numerous pages through generative AI tools as one way to do that. It also lists paraphrasing other people's work, stitching feeds together, and pumping out pages with a template and a spreadsheet. A human-written page with no first-hand value sits in exactly the same bucket as a machine-written one.

That distinction matters because it changes what you audit. If you go looking for AI in your content, you will find it everywhere and learn nothing. If you go looking for pages that exist to catch a keyword rather than answer a person, you will find the actual liability, and some of it will have been written by a person on your payroll in 2021.

The evidence on unedited AI at scale

SE Ranking ran a 16-month experiment worth knowing about. Twenty brand-new domains, 2,000 articles, zero human editing, zero link building. Around 71% of the pages were indexed in month one and the sites pulled over 122,000 impressions. By month three, only 3% of pages were still in the top 100. The August 2025 spam update briefly pushed that to 20%, and it did not hold. Almost every impression the experiment ever earned arrived in the first two and a half months.

Read that as a shape, not a verdict. Unedited AI content at volume gets indexed, gets a short burst, and decays. It does not get manually penalised in most cases. It just stops mattering, which is worse in a way, because nobody gets a notification about it and the pages sit there diluting the site for years.

AI search does not care who typed it

Here is the counter-intuitive part. Ahrefs analysed one million AI Overview results and 1.9 million cited URLs in July 2025. Of the pages cited, 3.6% were pure AI, 8.6% were pure human, and 87.8% were mixed. In a general sample of 900,000 new web pages, 25.8% were pure human. So the pages getting cited are less likely to be pure human writing than the web at large.

Google is not detecting and demoting AI at the citation layer. It is selecting for something else, and AI assistance is neutral on that axis.

Graphite backs the volume picture. Across 65,000 English URLs from Common Crawl, AI-written articles passed human-written ones in November 2024 and have sat near a 50/50 split since, not climbing. The web did not get flooded past the point of no return. It settled.

The sibling policy brands actually trip

Scaled content abuse gets the headlines. Site reputation abuse catches more real businesses.

That policy covers third-party content published on your domain, riding on ranking signals your own team built. Google began issuing manual actions for it in 2024. The classic shape is a publisher renting out a coupons subfolder. The shape a medium-sized business hits is quieter: a partner-run resources hub, an outsourced glossary sitting under your domain, a syndicated jobs board, a white-labelled comparison section someone in sales signed off in 2022.

Ask one question about every subfolder on your site. If this content were on its own domain with no borrowed authority, would it rank at all? If the honest answer is no, and you did not commission it to serve your customers, it is a risk sitting inside your best asset.

Expired domain abuse is the third sibling: buying a dead domain for its old links and repurposing it. Less common in-house, very common in agencies of the acquire-and-redirect variety. Worth asking whoever manages your migrations.

Google names the format it does not want

The May 2026 generative AI guidance does something unusual. It gives a worked example of the content type not to publish, and the example is something like 7 Tips for First-Time Homebuyers.

That is the listicle. Commodity content, in Google's own word. The format is not banned, obviously. The point is that a page which restates what forty other pages already say has no reason to be retrieved, cited or clicked, and no amount of on-page work fixes the underlying problem, which is that the page contains nothing.

The audit that finds the actual risk

Pull every URL from Search Console with a 16-month window. Sort by clicks ascending. Then, for each page in the bottom half, answer three things, in writing, in a spreadsheet:

1. 1.What does this page contain that could only come from us? A number from our own jobs, a photo of our own work, a rule we learned the hard way, a price, a failure. If the answer is nothing, the page is commodity content regardless of who wrote it.
2. 2.Who commissioned it and for what? Pages commissioned for a keyword are the ones that read as manipulation. Pages commissioned because a customer kept asking the question almost never do.
3. 3.Is it ours? Third-party or partner-produced content on your domain gets a separate column and a separate decision.

Anything failing all three is a delete or a merge, not a rewrite. Rewriting a page that has nothing to say produces a longer page that has nothing to say.

What this means for how you use AI

Use it where it does not touch the thing that makes the page worth reading. Structuring an interview transcript, drafting the boring connective sections, producing eight headline options, cleaning up a technical explainer your engineer dictated in one pass, all fine. Google's own citation data says the mixed page is now the normal cited page.

Point it at a request to write twelve articles about our service area and you have built a liability with a three-month half-life.

The tool was never the question. Whether the page deserved to exist always was.