Google's AI Generated Content Policy: Four Rules From the Documentation

There is no rule against writing with a model. There is a rule against publishing at volume with nothing in it, and teams keep reading the second as the first. Getting the distinction right is what separates a model that helps your site from one that quietly damages it.

Everything below is quoted from Google's published pages rather than summarised from somebody else's summary. It is written for a founder or a content lead who drafts with a model. Four rules come out of it: three long-standing, one added recently. Search is the scope here, and none of it is legal advice.

What the documentation actually forbids

Start with what Google encourages. Its page on generative output calls the technology useful "when researching a topic" and for giving shape to work that is already original. The caution on that same page is about scale without substance: tools used this way to produce "many pages without adding value" can breach the scaled content abuse policy.

So the chain runs: many pages, nothing in them a reader could not find elsewhere, policy triggered, visibility at risk. The production method is not the trigger at any step. The missing substance is. A freelancer filling a template from a spreadsheet lands in exactly the same place.

Rule one: originality is the test, and the author is not

Scaled content abuse, in the spam policies, is page generation whose main point is "manipulating search rankings and not helping users". The helpful-content guidance supplies the test a reader would apply, asking whether a page offers "original information, reporting, research, or analysis" and, where it draws on other sources, whether it avoids copying or rewriting them and instead adds substantial value.

That second question is the one that disqualifies most AI drafts. A model trained on the open web is structurally good at the thing the test rejects: competent restatement of what already ranks. Which means every piece needs one input that cannot be restated, your own data, a named person who said something, a worked example with real numbers, or a position you are willing to defend. Our note on where a human still has to be in the loop covers where that input gets added.

Rule two: fluent and correct are unrelated properties

Google states the mechanism directly. These systems "don't retrieve facts", it says, and instead predict likely wording from training data. The same page calls manual fact-checking critical before anything publishes, and pushes that review into the parts nobody reads: the title tag, the description, the schema markup and the alt text on every image.

The reason this fails so often is that fluency reads as confidence. A wrong sentence in clean prose survives a skim, which is exactly how most review happens. The repair is dull. Open the original document next to the draft and clear one claim at a time. In its 2026 study of business-to-business teams, Content Marketing Institute reports that roughly nine in ten now draft with AI, while under four in ten credit it with better results. Dropped review is one plausible explanation for that gap, and the research itself does not assign a cause.

Rule three: disclosure is a question Google now writes down

The helpful-content page asks outright whether "the use of automation, including AI-generation, self-evident to visitors through disclosures or in other ways", and whether a page carries a byline where a reader would expect one. It also treats invented creator identities, including "AI-generated headshots, made-up names, or false credentials", as deception rather than a style choice. That page was last revised on 5 October 2026, so it is current guidance rather than a historical artefact.

Disclosure can also be a legal question depending on where you publish, which is a separate matter from search. Our walk through the EU transparency code in ordinary language is a starting point, and a qualified professional is the next step.

Rule four, added while most people were not looking

The spam policies page now names "attempting to manipulate generative AI responses in Google Search" in its opening scope, alongside manipulating rankings. The page carries a revision date of 28 August 2026. That single clause is the most consequential change to this documentation in a year, because it means the tactics being sold as answer engine optimisation are now inside the stated scope of a spam policy rather than outside it.

Google's own guide to generative features goes further and rejects the premise of a separate discipline: "optimizing for generative AI search is optimizing for the search experience, and thus still SEO". It also sets out the eligibility conditions plainly, that a page must be indexed and able to appear with a snippet, while noting that meeting them guarantees nothing about being crawled, indexed or served.

Two documents, two different questions

	The spam policies	The helpful-content guidance
The question it asks	Was this built to manipulate a ranking or an AI response?	Would a reader find this useful and trust it?
What sets it off	Volume with nothing added	Low effort, borrowed substance, hidden authorship
Who it applies to	Any method of production	Any method of production
Last revised	28 August 2026	5 October 2026

The AI-visibility checklist you can delete

Here is the position: the file-and-markup layer of the current AI visibility advice is waste, and Google has said so in its own documentation. The guide states that "you don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search". That is a direct answer to an industry that has spent a year selling llms.txt files, AI-specific schema and parallel Markdown versions of every page. If you have those tasks in a backlog for Google visibility specifically, cut them and move the hours onto the originality test in rule one.

The fair counter-argument is that Google is not the only reader. Other assistants crawl and cite independently, publish their own guidance, and a file Google ignores may still be read somewhere else. That is true, and it is the only honest reason to keep one. It is a different justification from the one these tasks are usually sold with.

This call moves on one simple, public event: Google adding a machine-readable file or an AI-specific markup type to its own documented requirements. The guide's revision date of 10 July 2026 is the version this rests on (Google). Re-read it before any quarter where you plan to spend real engineering time on this, and check the revision date first.

Five questions, ten minutes, before anything publishes

• Name the one thing in this piece a competitor could not reproduce by reading the same ten sources.
• Did a human verify each figure, name, quote and link against the original document, rather than against another article?
• Does the metadata, from the title tag through to image alt text, describe what the page really contains?
• Is it self-evident who wrote this, and is the way automation was used recorded somewhere you could show a reader?
• If a customer asked how this was produced, could you answer without hedging?

Five yes answers promise nothing about visibility, and no Google document promises it either. What they do is keep a page clear of the boundary these policies genuinely define. For the periodic re-read, our checklist for these pages tracks what changed, and the claim check on rewriting tools covers the gap the documentation leaves on rewrites.

The standard did not change, only the speed

Add something of your own, verify it, and be straightforward about how it was made. Those were the rules for a human writer and they are unchanged for a model. What a model changes is how quickly you can break them at scale, and that is precisely the behaviour the policy names.