Does Google know your article was written with AI, and what that means for your site

Plenty of contradictory stories go round about AI content. One source says Google cannot see anything, another warns that every word out of a model is dangerous. The reality is more level headed, and Google writes it down itself: if you use automation, AI generation included, to make content with the primary purpose of manipulating your positions, you are breaking the spam policy, checked on 25 August 2026. A detection mechanism with layers, thresholds or a score is nowhere published by Google, so we do not describe one here either.

What Google does publish are the questions you should ask yourself as soon as a text is largely produced automatically: is it clear to the reader that automation was used, and can you explain why that approach helps this reader. On top of that, rolling out pages at scale without value of their own is its own breach in the spam policy, under the name scaled content abuse. So a text in which a specialist has added their own examples, data and nuance stands differently from a text that adds nothing. That has nothing to do with banning AI, and everything to do with rewarding real expertise.

Three levels you hold in your own hands

• Text level: what does this page add that is not elsewhere, and is your own experience in it.
• Page level: the build of the page, internal references and checkable sources, so that a reader can see where you got it from.
• Scale level: no hundreds of pages saying the same thing, because at Google that is a breach in its own right.

At text level you have the quickest influence through your content process. The trick is not to avoid every form of AI, but to bring your content into the zone where Google clearly sees that a person is in charge.

Google does not share every detail of its systems. Even so, on the basis of research, patents and practical experience a reasonably clear picture emerges of the signals that may count towards whether a text moves towards generic model output or towards human expertise.

• Perplexity and predictability: by running your text through a model of its own, Google can estimate how likely each word is. Texts that are extremely flat, or strikingly odd, stand out.
• Rhythm and sentence variation: human writers alternate short and longer sentences. Models tend towards a calm rhythm with no peaks. Pleasant to read, but also a recognisable pattern you can deliberately break.
• Overlap with known model output: if your page matches earlier generated output on many points, that is a clear signal that little of your own has been added.
• A lack of concrete examples: generic explanation with no practical cases, data or names is easy to generate. Your own situations and observations give your text a fingerprint that is hard to copy.

Looking at text alone is not enough. So Google uses behavioural signals as well. Do visitors stay, do they click on, do they come back to the results and search again for the same subject? A text that looks tidy to a model but that people click away from quickly probably ends up lower over time. The other way round, an imperfect text with strong content and good engagement can stay steady.

How you use AI largely decides how your content is seen. In practice we see three common patterns. The chances are you recognise yourself in one of these scenarios.

AI as an assistant, a person as the author

You use models for ideas, outlines and suggestions, but the final text is written by your team. You add plenty of practical examples and every important piece goes past an editor. This is the healthiest form: you use the speed of AI while your brand story and your expertise stay clearly visible.

Hybrid text with light editing

You let a model write the first version and mainly adjust language and layout. You add an example here and there, but the base stays recognisable as model output. This approach works as long as you stay critical. The more you lean on standard phrasing, the greater the risk that Google marks your content as middling.

Mass publication without editing

You publish large numbers of pages pasted almost straight out of a model. Tone, structure and examples look much alike and change little from subject to subject. This is exactly the territory Google watches. Not because AI was used, but because the content is hard to tell apart from generic text. What it means for your broader approach when text loses value as evidence is in our analysis of the post text economy: that one is about the ground text is losing to video, and about the difference between how a machine reads your site and how a customer does.

The question is not whether you may use AI. The question is how you do it in a way that makes your brand stronger instead of flatter. You can lay this plan straight over your own content process.

• Decide where AI may and may not write. Put models to work on research, structure and generating questions. At the same time, set down which parts are always written or rewritten by a specialist.
• Start with a human angle and a clear position. Write down briefly what you want to say, which position you take and which situations you definitely want to name. That stops the model from shaping your story instead of the other way round.
• Fill in your own examples and data systematically. Refer to figures, cases, clients or products that exist only with you. Those are exactly the things a model does not come up with by itself.
• Break the rhythm a model chooses. Alternate longer explanation with short observations. Ask a question in passing, make a sharp remark, or point out that in practice something often runs differently from the theory.
• Build an AI ready structure. Work with clear headings, summaries, lists and explanatory blocks. Combine that with schema data. Adding an llms.txt as a tidy table of contents is fine, but do not count on it: no large AI provider has committed to fetching that file from external sites, and Google says Search ignores it.
• Have an editor look at the text deliberately. Does the text sound like something you would say in a conversation? If the answer is no, it needs another pass in which your voice comes through more clearly.

This plan connects to the rest of your AI strategy, such as your AI visibility score and the way you use llms.txt to signal which pages may be used as a source. Would you rather set this process up together? Then look at our service AI and automation.

With the right prompts you use AI as an editor rather than as an automatic text machine. Paste your own text under the prompt in a model of your choice and have it run these five checks:

• Generic model style: which paragraphs feel most interchangeable and where are concrete examples, data or names missing? Ask for five rewrite suggestions that make the text sound more like a human account.
• Rhythm and variation: is the average sentence length roughly the same everywhere? Have it point out three places where a shorter sentence or a direct observation works better.
• A sharper position: which parts stay too cautious or too general, and which concrete positions suit this subject?
• Unique fingerprint: which passages could you also find on other sites, and which parts are genuinely specific to your brand or situation?
• Balance between AI help and human input: have it estimate on a scale of 0 to 10 how strongly the text feels like bare model output, plus the three changes that would lower that score most.

Use the feedback as an extra layer of checking, not as a verdict. You decide what ends up on the page.

With an llms.txt file you give language models extra context about your site. In it you indicate which pages are suitable as a source, which content you would rather not see used as a basis and how the writing process is arranged. Describe for instance that AI helps, but that experienced specialists always carry the final responsibility and check and add to all the content. What matters most is that you describe honestly how you work and which parts of the site represent your expertise.

Add schema markup to that, such as Article markup with your organisation as author and publisher. That helps search engines understand better what the article is about and what role your organisation plays. Transparency about your editorial process, on an about us page for instance, gives models context that can work in your favour.

Do the same in your own knowledge base: give the subject of AI detection a page of its own and refer to it from related articles. That creates a logical network of pages that visitors and AI assistants can both follow. Our own knowledge base is built the same way.

Most people arrive at this subject the other way round: not before publishing, but after a drop. Something happened, AI was involved somewhere, and the two get connected. Before you rewrite anything, take that connection apart, because acting on the wrong cause costs you months.

First check whether it is your pages at all

Open Search Console and compare a long enough period with the same period a year earlier, not with last month. Then look at whether the drop sits across the whole site or on a handful of pages. A site wide drop that starts on one date usually points at something outside your content: a core update, a technical change, a redirect that went missing during a redesign. A drop on specific pages is far more likely to be about those pages.

Then check what those specific pages have in common

Line up the pages that lost the most and read the first two paragraphs of each. If you cannot tell them apart, you have your answer, and it is the scale problem from DOC.K.03 rather than the fact that a model was involved. If they are all in the same subject cluster, the more likely explanation is that somebody else now answers that question better than you do.

Rewrite the few that matter, not everything

The instinct after a drop is to go through everything, and that instinct is wrong. Pick the pages that carried real traffic or real enquiries and give those your full attention: your own figures, a case you handled, the exception that only comes up in practice, the part of the answer other pages leave out. Ten pages rewritten properly beat a hundred pages touched up.

What to do with the rest

Pages that never brought anything and add nothing are not worth rewriting either. Merge them into a page that does earn its place, and redirect the old address to the new one. Removing them without a redirect throws away whatever little they had. Leaving them there is not free either, because a site full of interchangeable pages says something about the site as a whole.

One warning about timing. Whatever you change, do not expect the graph to answer you next week. Give a rewritten set of pages a full quarter before you judge it, and change one thing at a time, otherwise you will never know which move did the work.

Everything above is about risk, which makes it easy to forget why anybody started using these tools. Used in the right place they take hours out of a week. Used in the wrong place they hand you work that looks finished and is not, which is more expensive than doing nothing.

Where it genuinely helps

• Getting started. The blank page is the most expensive part of writing. A rough outline you disagree with is still faster than nothing, because disagreeing is easier than inventing.
• Asking you the questions a reader would ask. Give a model your draft and ask what a sceptical customer would still want to know. That is a genuinely useful second pair of eyes and it costs you a minute.
• The boring layers. Summaries, headings, alternative titles, tightening a paragraph that ran long, turning your notes into an ordered list. None of that carries your expertise, so none of it needs to come from you.
• Translation and tone. Turning something you already wrote into another language or another register, with a person checking the result, is one of the safest uses there is.

Where it costs you without showing it

• Facts, numbers and names. Anything a model states confidently still has to be checked against a source, and checking takes longer than looking it up yourself would have. Assume every figure is wrong until you have seen where it came from.
• Your own positions. A model will produce a balanced view of any question, and a balanced view of everything reads as an opinion about nothing. The parts where you take a side are exactly the parts to write yourself.
• Anything about your own business. Prices, guarantees, delivery, what you do and do not do. A model fills these in plausibly, and plausible is precisely the problem: it will be on your site, not on theirs.
• Volume for its own sake. The moment the tool makes it cheap to publish more, the question stops being can we and becomes should we. That is the point at which this article stops being about writing and starts being about the spam policy.

The pattern behind both lists is the same one running through this whole page: automation is safe where it handles form, and risky where it handles substance. Keep that line clear and the tooling question mostly answers itself. How we set that division up as a working process, rather than as a rule on paper, is what our service AI and automation covers.

Does Google penalise all AI content automatically?

No. Google looks at the value of the content and not at how it was made, checked on 25 August 2026. An article produced with the help of AI but clearly added to with expertise, examples and good structure can perform perfectly well. What Google is strict about is pages that offer no real added value, that are written mainly for search engines, or that are clearly part of a spam network.

Does Google use AI itself to recognise AI content?

Which systems Google uses for that it does not publish, so we make no claim about it. What is stated: there is no list of banned tools, and the question is whether a page adds value for the person reading it.

Should I delete older AI articles from my site?

Deleting is not always necessary. Start with the pages that matter for your traffic and revenue and read them back critically. If they are very generic, you are better off rewriting them and adding practical examples, data and a stronger structure. Pages that get hardly any traffic and add little value you can merge or remove from the index.

Are public AI detectors a reliable measure of what Google sees?

Public detectors give a rough indication, but they have limits. They sometimes mark human text as AI and the other way round. See them as an extra signal, not as the truth. For your SEO strategy it is more useful to look at content, at visitor behaviour and at your position in the market.

How much AI may I use before it becomes risky?

There is no fixed line. A practical rule of thumb: every important page has to carry a clear human stamp. Can you no longer explain which passages come from your own experience? Then you make it harder for the algorithm to see the text as a reliable source. So use AI above all as an accelerator, not as a replacement for your expertise.

How quickly do I see an effect when I change my AI approach?

The speed differs per site, per sector and per level of competition. You often see signals in engagement and behaviour on the page within a few weeks. Changes in positions and in visibility in AI answers usually build up in steps. Keep the new approach up consistently and review it each quarter.

Want to read more about how Google is changing? Look as well at what Google Search Live means for trades in 2026 or at how to advertise as a physiotherapist without breaking the UK GDPR or the CAP Code.

The honest summary of this article is that the question is not whether a model touched your text, but whether a reader can tell that somebody who knows the subject stood behind it. That is a judgement, and judgements are easier to make about someone else's pages than about your own.

If you want that read on your own site, send us the address and the two or three pages you are least sure about. We will tell you which of them look interchangeable, which ones carry something only you could have written, and what we would rewrite first. Including when the answer is that they are fine as they are. You can reach us through the contact page, where you pick a slot in the calendar yourself, or by phone. A first conversation carries no price and no obligation.