AI-written content: what it actually does to your rankings
Google does not have a rule against content written with AI. It has a rule against something these tools make very easy to do at scale, and the difference decides whether the way you are using them is safe or expensive. This is what the spam policy actually says, what genuinely gets a site demoted, and the checks that matter far more than any of it.
The claim everyone repeats, and what is wrong with it
“Google penalises AI content.” You will have heard it from someone selling copywriting, and you may have heard the opposite from someone selling an AI tool. Both are overstating their case, and the truth is more useful than either.
Google does not have a rule against content written with AI. It has a rule against a thing that AI makes very easy to do at scale, which is not the same claim, and the difference decides whether the way you are using it is safe or expensive.
This article sets out what the policy actually says, what genuinely gets a site demoted, why AI drafts tend to fail on their own merits well before any policy is involved, and where the tools genuinely save time in real work. It is written by somebody who uses them daily and has had to build a checking process around them.
What the spam policy actually says
The relevant rule in Google’s spam policies for web search is called scaled content abuse. Its own definition describes many pages generated for the primary purpose of manipulating search rankings rather than helping users, typically large amounts of unoriginal content of little value — and then adds the clause that settles the argument: “no matter how it’s created”.
That clause cuts both ways, and people tend to only quote the half that suits them.
It means a page is not in trouble because a model wrote it. It also means a page is not in the clear because a human did. The production method is explicitly not the test. The test is whether there are many pages, whether they are unoriginal, and whether they exist to rank rather than to help.
The policy does name generative AI directly in its examples — using such tools to generate many pages without adding value for users. Note every qualifier in that sentence. Many pages. Without adding value. Neither of those describes one carefully edited article on a subject you actually know.
What actually gets a site demoted
In practice, the sites that get hurt share a recognisable shape, and it has very little to do with which tool produced the words.
They publish a great deal very quickly — dozens or hundreds of pages appearing where there were none. The pages are near-identical to each other, usually one template with the town, the trade or the product swapped. They say nothing that could not have been assembled from other pages already ranking. And they exist because somebody wanted to cover a keyword list, not because a customer ever asked the question.
Any one of those is a warning sign. Together they are the pattern the policy describes, and it is entirely possible to produce it by hand. Agencies were churning out near-duplicate location pages long before these tools existed, and those pages were demoted too.
The corollary is the reassuring part. One genuinely useful article a month, edited by someone who knows the subject, is not the thing the rule is aimed at, regardless of what helped draft it.
Most AI drafts fail long before any policy applies
Worrying about penalties is mostly a distraction. The far more common outcome is simply that the page never ranks, because it has nothing in it worth ranking.
An unedited draft has a set of predictable weaknesses:
- It says what everything else says. These tools produce the consensus of what has already been written. If your page is a well-phrased average of page one, there is no reason for Google to prefer it to the pages it averaged.
- It has no first-hand experience in it. It cannot tell you what you learned from a job that went wrong, what a supplier in your area actually charges, or which fix works in the housing stock you deal with. Those are the only parts of a page a competitor cannot replicate.
- It hedges everything. Drafts are full of “it depends” and “there are many factors”, because that is the safest thing to say when you do not know. Readers came for a position.
- It is confidently wrong at the edges. Which brings us to the part that has cost me the most time.
None of that is a reason to avoid the tools. It is a reason not to publish their output as it arrives.
Invented facts are the actual liability
This is the risk nobody selling AI content mentions, and it is far more immediate than any ranking question.
Working with drafts produced this way, I have had to strip out statistics that were never published anywhere, percentages attributed to reports that do not say that, and on one occasion an entire organisation that does not exist — a plausible-sounding university with a plausible-sounding web address, invented whole and stated as fact. Nothing about it looked wrong. It read exactly like the true sentences around it.
That is the difficulty. A fabricated figure does not arrive flagged. It arrives in the same confident tone as everything else, in the middle of a paragraph that is otherwise correct, and it will sit on your website with your name above it.
For a regulated business this is not merely embarrassing. A clinic, a law firm or a financial adviser publishing an invented statistic has a compliance problem, not a content problem. And for anybody else, a customer who checks one claim and finds it is untrue has learned something about the whole site.
Every number, every named source, every date and every regulatory claim has to be verified against the original before it goes live. Not spot-checked. Every one.
Where these tools genuinely save real hours
I use them constantly, and the useful applications share a feature: the tool is doing work where being wrong is immediately visible, rather than work where being wrong is invisible until a customer notices.
- Structure before writing. Turning a messy set of notes into a sensible running order is genuinely quick, and a bad order is obvious the moment you read it.
- Editing your own draft. Asking what is unclear, what is repeated and what a sceptical reader would push back on. This is the strongest use and the most underused.
- Scale tasks with a verifiable answer. Drafting title tags and meta descriptions to a character limit, or working through a list of pages looking for a specific fault. You can check the output mechanically.
- First drafts of things only you can finish. A skeleton you then load with the specifics that came out of your own work — the ones that make the page worth reading.
What they are poor at is the thing people most want from them: producing finished, publishable, factually reliable pages without anybody who knows the subject reading them properly. That is precisely the job where the failure is invisible until it is expensive.
The checks to run before anything goes live
- Verify every fact against its original source. Numbers, dates, named bodies, regulations, quotes. If you cannot find the source, remove the claim rather than soften it.
- Delete anything that could have been written about any business in your trade. If a competitor could publish the paragraph unchanged, it is not earning its place.
- Add at least one thing only you could have written. Something from a real job, a real objection you get asked, a real constraint in your area.
- Check it does not duplicate your own pages. The most common self-inflicted wound is publishing a near-copy of something already on your site, which splits the signal rather than adding to it. This is one of the things a proper audit looks for.
- Read it aloud. Hedging and padding are audible in a way they are not on screen.
- Ask whether it answers the question a customer actually asked — not the keyword, the question.
If a page cannot survive that list, the answer is not to publish it with a disclaimer. The answer is that it was not ready, and publishing it makes the site slightly worse rather than slightly bigger.
What this means for a small business
If you run a business and somebody has offered you fifty AI-written pages for a low monthly fee, you now have the vocabulary to evaluate it. Ask how many pages, how different each one genuinely is, and who verifies the facts. The pattern being sold to you is the pattern the policy describes.
If you are writing your own content with these tools, the honest position is that they have made drafting faster and made checking more important, and the second has grown roughly as much as the first has shrunk. The total time saved is real but smaller than advertised.
And if you are choosing between publishing twenty average pages and four genuinely useful ones, choose the four. That was true before any of this existed. What has changed is only that producing the twenty is now cheap enough to be tempting, which is exactly why the ordinary discipline of fewer, better pages has become more of an advantage rather than less.
The businesses that will do well out of these tools are the ones using them to publish things nobody else could publish, faster. Not the ones using them to publish more of what everybody already has. If you want that done properly, it is the writing itself that has to carry the difference — the tool never will.
AI content questions, answered plainly
Does Google actually penalise AI-written content?
Not for being AI-written. The relevant rule is scaled content abuse, which targets many pages produced mainly to rank rather than to help, and its wording explicitly says this applies no matter how the content is created. A human can breach it and a model can avoid it. The test is scale, originality and purpose, not authorship.
Can Google tell whether AI wrote my page?
Treat the question as irrelevant rather than trying to answer it. Google does not need to identify the tool to assess whether a page is unoriginal, thin, duplicated across a site or unhelpful — and those are the things actually being measured. Optimising to look human is solving the wrong problem; producing something worth reading solves both.
Should I disclose that AI helped write an article?
There is no search requirement to, and adding a disclaimer does not make a weak page acceptable. Disclosure is a trust decision for your own audience rather than a ranking one. What genuinely matters is that a named person stands behind the accuracy of what is published, because that is who owns the consequences if a claim is wrong.
Is it safe to generate location pages for every town I cover?
This is the single riskiest use, because it produces exactly the pattern the policy describes: many near-identical pages differing by a place name. If each page genuinely says something different and true about that place, it can work. If the only variable is the town, you are building duplication, and it tends to weaken the pages you already have rather than add to them.
How do I check a statistic an AI tool gave me?
Find the original publication and read the figure in place. Not a blog citing it, not a summary, not the tool’s own restatement of it — the source document. If you cannot locate it in a few minutes, assume it is not real and cut it. Fabricated figures are usually plausible, specific and attached to a real-sounding organisation, which is precisely why they survive a casual check.
Will AI content hurt rankings on pages I already have?
It can, indirectly. Adding a large volume of thin pages to a site can dilute what the site is understood to be about and split relevance between near-duplicates. The risk is rarely a dramatic penalty; it is a gradual softening across pages that were previously doing fine. If rankings slipped after a bulk publishing push, that is the first thing to examine.
Is there any amount of AI content that is definitely safe?
The question has the wrong shape, and anybody answering it with a percentage is inventing one. There is no published threshold and no ratio to aim for. A site of thoroughly edited, genuinely original articles is fine whatever helped draft them; a site of unedited near-duplicates is at risk whatever the proportion. Judge the pages, not the process.
Should I delete AI-written pages I have already published?
Judge them one at a time rather than as a batch. A page that gets traffic, answers a real question and contains nothing invented is worth keeping and improving. A page that duplicates another, attracts nothing and contains claims you cannot verify is worth removing or merging. Deleting everything because of how it was produced discards the ones that were working.
Send the pages and I will tell you which to keep
Send the website address and say roughly how the content was produced. You get a written answer: which pages duplicate each other, which ones contain claims that cannot be verified, which are earning traffic, and which would be better merged or removed. It costs nothing and there is nothing to sign. If the honest answer is that the content is sound and the site simply needs links, that is what the answer will say.