Google does not penalise AI content. It penalises what careless AI content does.

There is no penalty for content because it was produced with a model. Search guidance has been consistent on this for years: the assessment is about whether the page helps the person who lands on it, not about the process that produced it.

That answer is unsatisfying because it does not explain the traffic people actually lost. The explanation is usually one step removed: pages produced at scale tend to be thin, duplicative, unsourced and stale, and those properties are what quality systems are designed to notice.

So the useful question is not whether the tool is penalised. It is which of the four failures on this page applies to your site, and what the fix is.

Everything below is about those failures, not about authorship, because authorship is the part you cannot change and the rest is the part you can.

The guidance that matters has three parts. Content should be created primarily for people rather than to manipulate rankings. It should demonstrate first-hand experience and be accurate. And automation used to produce content at scale, without adding value, is treated as spam.

Notice what is absent: any requirement that a human type every word, and any statement that generated text is inherently lower quality. The policy targets the value of the output, not its origin.

Guidance also evolves, and the wording changes while the principle does not. Anyone quoting a specific line should date it, because a sentence from four years ago may describe a policy that has since been rewritten.

The practical reading is this: the standard is the same one that applied to cheap outsourced content in 2012. If a page exists mainly to occupy a keyword, it is at risk, whoever wrote it.

It survives because it offers a clean cause for a messy outcome. A traffic drop after publishing forty generated pages feels like the tool was punished, when the likelier cause is that forty near-identical pages diluted the site and none of them earned a link.

Vendors on both sides amplify it. Detector companies sell certainty about authorship; agencies sell human-written content as insurance. Neither has an incentive to explain that the risk sits in the editorial decisions.

Search representatives have repeatedly said the same thing in public, which has not settled it, because anecdotes travel faster than guidance and every drop has an obvious suspect.

The result is a lot of effort spent on concealing tool use, which changes nothing that matters, and very little spent on sourcing claims and differentiating pages, which changes everything.

First, thin content: pages that answer a question the reader already knew the answer to, or restate a definition without adding anything. Length is irrelevant; the test is whether the page delivers a fact the reader did not have.

Second, scaled duplication: many pages generated from one template to cover keyword variations, city names or product permutations, where each page adds nothing the others do not. This is explicitly in scope of spam policy.

Third, unreliable information: claims with no source, facts that were true two years ago, health or financial advice with no accountable author. Trust failures are the ones that cause the sharpest losses.

All three are production decisions. They can be made with a model, with freelancers or with an in-house team, and the outcome is judged the same way.

There is no field in a page that records how it was written, and no signal that reliably indicates it. Stylistic detection is unreliable enough that using it as a ranking input would create more errors than it fixes.

This cuts both ways. It means a careful AI-assisted page is not disadvantaged. It also means there is no benefit to disguising the process: rewriting generated text to sound human spends effort on the part that is not measured.

Where authorship does matter is accountability. A named author with a profile, a real contact route and other work on the same topic is a trust signal. A page with no owner is a page nobody will maintain.

If your content affects money, health or safety, the expectation of clear ownership is stronger, and the honest move is a byline plus a note on how the page was reviewed.

These are different events and they have different fixes, which is why assuming the worst wastes time.

A manual action appears in the search console's manual actions report and names a specific policy. If that report is empty, no human penalty was applied, and the loss came from ranking systems reassessing the site.

An algorithmic decline shows as a gradual or date-aligned drop with no message. It usually reflects a site-level pattern: too many similar pages, weak evidence across the domain, or a competitor now serving the query better.

Check the report before theorising. It is the one piece of evidence that distinguishes the two, and it takes a minute.

In practice, sites that lose visibility after heavy generated publishing tend to share a set of properties. None is a penalty criterion; together they describe a site that stopped being useful.

Start with the search console. Confirm whether a manual action exists, then look at which queries and pages lost impressions rather than positions, because losing impressions means the site stopped being considered at all.

Then audit the pages that lost traffic. Not the whole site: the ones with the steepest decline. Check whether each still answers its query better than the results above it, and whether anything on it is now wrong.

Look for duplication across your own pages. Consolidating three overlapping pages into one strong page is the most common effective fix, and it is unglamorous enough that most people try something else first.

Improve the survivors rather than deleting everything. Pruning without adding evidence leaves a smaller site that is still unreliable, and the next decline arrives on schedule.

There is no requirement to disclose, and no penalty for doing so. There is also no benefit to concealment. The decision is editorial and belongs to your readers.

A practical approach is a short process note where it matters: this page was drafted with assistance and fact-checked by a named editor, with sources listed. That tells a reader what they need to know about reliability, which is the actual question.

For research, journalism and education, disclosure is often required by the venue rather than by search. Check the policy you are publishing under, which is a different document from a search guideline.

Whatever you choose, keep it consistent across the site. A disclosure on one page and silence on the next reads as an accident rather than a policy.

If you publish pages that answer real questions, contain something only you could supply, name an owner and cite their sources, the tool you used is not a risk factor worth planning around.

If you publish hundreds of pages that restate what already ranks, no amount of human rewriting will save them, because the problem was never the writing style.

The work that protects a site is the same work it always was: fewer pages, better evidence, clear ownership, honest scope. Generated drafting makes that work faster, and it makes the shortcut faster too, which is the whole risk.

Judge your own pages by that standard before anyone else does. It is the only review that changes the outcome.