AI text and search engines: where the line actually falls

People ask whether AI text gets penalised as though the answer were a secret. It is published, and it is unambiguous. The difficulty is that the wording in the documentation is shorter than any summary of it, and summaries keep dropping the one phrase that decides everything.

What Google's documentation says

Google does not penalise text for having been generated. It penalises purpose. In its own words: "If you use automation, including AI-generation, to produce content for the primary purpose of manipulating search rankings, that's a violation of our spam policies" (developers.google.com/search/docs/fundamentals/creating-helpful-content, checked 19 September 2026).

The phrase carrying the weight is "primary purpose". It draws the line at what the material is for rather than at how it was produced. Text written by a model and useful to a reader does not breach the policy. Text written by a person for rankings and useless to a reader breaches it exactly as much as generated text would.

The same page sets out a frame of three questions: who, how, and why. Who means authorship should be clear from a byline and an author page. How means automation is worth disclosing where a reader might wonder how something was made. Why is, by the documentation's own account, the most important: the material should exist to help people rather than to attract search traffic.

Where the line falls in practice

Some fairly concrete things follow from that wording.

• Three hundred pages built from a template against three hundred near-identical queries breach the policy regardless of who wrote them.
• An article with a specific reader and a specific question does not breach it, regardless of who wrote it.
• A paraphrase of three pages from the top ten does not become more useful for having been made by a model, and no less useful either. It was useless and stays that way.

The practical test: if a piece cannot be described as "this text answers this person's question about this", the problem is not the tool. Swapping the tool changes nothing there, because the question stays exactly the same.

What breaks in model-written text

Penalties are far from the main risk. Much more often AI text simply does not work, for recognisable reasons.

It contains no fact the competitors lack. A model writes from the averaged contents of the internet, so by default it produces what everyone else has. The page comes out correct and exactly as useful as the ten that already exist.

Numbers look like facts and are not. This is the expensive one. A model will confidently write a price, a market share, or a date because such a figure is plausible. Every one has to be checked.

The structure fits an expectation rather than a query. A model reproduces the shape of "introduction, eight subheadings, conclusion" well and guesses badly at what people actually ask on a given query, because it never saw the results page.

Everything sounds the same. The same constructions, the same rhythm. Readers notice before search engines do.

A working division of labour

The split that holds up at volume: the model handles form, a person owns the substance and the facts.

Note the third row. It is what separates a piece that works from one more paraphrase: your own figures, your own experience, specific examples. A model will not invent them, and asking it to is exactly how you get a plausible fabrication.

What a model genuinely does well

The list is shorter than the marketing promises and more useful.

• Turning a list of queries into readable subheadings.
• Cutting a finished text without losing the meaning.
• Reading large exports: forty thousand rows you need a pattern out of.
• Translating while preserving the argument, with a native speaker reading it after.
• Drafting a section once a person has written the points it has to make.

What all five share is that something the model does not itself have goes in as input. Where there is no input, the output is the average of the internet. Testing that takes a minute: ask a model to write a section with no input facts at all and compare the result against the first page of results on the same topic. The overlap is larger than anyone would like.

Does AI use need disclosing

Google's documentation advises disclosing automation where a reader might wonder how something was made. That is not a requirement to label every paragraph. It covers the cases where the method of production changes how the material should be read: a generated product review, an automatically assembled summary, a synthetic testimonial.

Reviews and any testimony written in a person's voice deserve their own note. A generated review is a fabrication however plausible it reads, and the risk there is not only a search risk.

How this sits in Yandex

Yandex has published noticeably less about AI text than Google has, and there is no point inventing a position for it. What can be said from observable behaviour: Yandex leans harder on user behaviour, and behaviour is where empty text fails first. Someone opens the page, does not find an answer, and goes back to the results.

So in practice the requirement is the same, expressed through a different mechanism. What else differs between the engines is covered in Yandex and Google.

The pre-publication check

1. Every number carries a source and a date in the sentence itself. A number without a source gets deleted.
2. The piece contains at least one fact absent from the first three results on that query.
3. The structure is checked against the top ten: page type and length match the format the engine treats as an answer.
4. The text has been read aloud. A uniform rhythm is audible immediately.
5. Internal links are in place and point at pages that exist.
6. The piece can be described as "answers this person's question about this".

Check the sixth item first. It filters out the pieces that should never have been started.

The second item deserves more attention, because it is the uncomfortable one. A fact competitors lack rarely comes from reading: it is your own numbers, a specific case you worked through, a result you measured. When there is no such fact, the honest conclusion is that there is nothing to write on the topic yet, and that is a normal state. Half the topics in any plan are waiting for something to say.

That also shows where a model really saves time. It removes the blank page and takes on the form, while gathering the substance stays exactly where it was. Counted across the whole piece, the saving is about a third rather than the multiples the marketing promises.

What does the work, and what it costs

Prices from the gbseo.ru catalogue as of 19 September 2026; "retail" is the vendor's own price for the same subscription. The full set is 4 999₽ a month against 36 700₽ at retail. The calculator on the homepage prices any subset. How these stages fit a quarterly schedule is covered in our piece on the content plan.

FAQ

Does Google demote AI text?

For the fact of generation, no. For automation whose primary purpose is manipulating rankings, yes, and that is stated directly in its documentation.

Do AI detectors work?

As grounds for a decision, no. They are wrong in both directions, and accepting copy on their percentage replaces a question about usefulness with a question about style.

Can a model's translation be published?

Yes, once someone who reads the language has been through it. An unread machine translation gives itself away with constructions the language does not use, and that loses readers before it loses rankings.

How much editing does a model draft need?

In practice, roughly as long as the generation took, on any piece with real substance in it. The saving comes from not facing a blank page rather than from the editing.

Should an article state that AI helped write it?

Where a reader might wonder how it was made. For an ordinary expert article with a named author that is usually unnecessary; for a generated summary or product review it is appropriate.