Auto-generated content is text produced by a program rather than written by a person: AI models, templates filled from a database, translation scripts, article spinners. Google does not ban it. Google penalises using it at scale to manipulate rankings without adding value, which is a narrower line than most people assume.
What the guidelines actually say
Google’s spam policies cover scaled content abuse, defined as producing many pages primarily to game rankings rather than to help people. The March 2024 rewrite of that policy was deliberate. The old wording targeted automation. The new wording targets scale plus low value, whatever produced the pages. A person hand-writing 500 thin pages breaks exactly the same rule.
Where automation is fine
- Data-driven pages. Weather, stock levels, sports results, flight times. The data is the value and nobody wants it typed out by hand.
- Templated product pages. A standard structure with genuinely different specs, images and availability per item.
- Drafts a person then edits. A model writes a first pass, a specialist corrects the facts, adds what only they know and puts their name on it.
- Translation with review. Machine translation checked by somebody who speaks the language, rather than published raw.
Where it goes wrong
The pattern that gets sites demoted is volume without substance. Two hundred location pages where only the town name changes. Question pages generated from a keyword export that restate the question and answer nothing. Articles that summarise what is already on page one without adding a figure, an example or a view.
Those pages tend to rank briefly, then vanish, and they drag the rest of the domain with them. Quality signals apply sitewide. A site that is 80% filler gets read as a filler site, including the twenty good pages on it.
AI search punishes thin pages harder
When a model picks sources to cite, restating the consensus is the fastest route to being ignored. Models cite pages that contribute something: original data, a specific process, named examples, a clear position. Generic content gets summarised out of existence because the model can already produce that text itself. Teams scaling content for AI search usually discover this the expensive way.
One test before you publish
Ask what this page holds that a reader could not get from the first three results, or from asking a chatbot directly. If the honest answer is nothing, the page is a liability, whoever wrote it. Add the missing thing or leave it unpublished.
Common questions
Will Google penalise AI-written content?
Not for being AI-written. Google has said publicly that it rewards quality content, regardless of how it was produced. What gets penalised is content made at scale mainly to rank, which describes plenty of AI output but by no means all of it.
Can Google detect AI writing?
To some degree, probably, but detection is not the basis of enforcement. The public detection tools are unreliable in both directions. Google’s systems weigh behaviour and quality signals across a whole site, which is harder to fake than a writing style.
Do we need to label AI-assisted content?
Google does not require it. Disclosure matters where a reader would expect human judgement, such as medical, legal or financial advice. For a product description, nobody cares and nobody asks.
How much editing is enough?
Enough that the page says something the model could not have known. A client result, a real price, a process you actually follow, a mistake you have watched someone make. If the edit pass only tidies phrasing, it has added nothing.
