Most small business owners we talk to have landed in the same place on AI writing. They are already using it in some form, and they are quietly worried that Google will work out what they did and bury the site for it.
It is a fair worry. It is also, on the best evidence available right now, mostly the wrong one.
What was actually measured
Ahrefs published a study in July 2026 looking for a relationship between how much of a page is AI-generated and where that page ranks. The sample started at a million pages taken from the top ten results across 100,000 searches in June 2026. Around 300,000 of those were in the Ahrefs crawler database. Of those, 150,000 had enough text on them to run AI detection against.
So the working sample is 150,000 ranking pages, each scored for how much of its content reads as machine-generated.
What it found
Three numbers do most of the work.
- 5.3% of top-ranking pages scored as 100% AI content.
- 9% scored at 80% AI content or higher.
- 54.7% scored under 20%.
Then there is the number that settles the question. If Google were demoting pages for being AI-written, you would expect the amount of AI content to climb as you move down the results. It does climb. Average AI content goes from 27.1% at position one to 30.9% at position ten.
That is a gap of under four percentage points across the entire first page. It is not nothing, and it does point in the direction you would expect. It is also nowhere near large enough to describe as a penalty. Pages at the top and pages at the bottom of page one look substantially alike on this measure.
What the study does not say
This part matters more than the headline, and most coverage of the study skipped it.
Ahrefs is explicit about the limits of its own method. Ryan Law, who ran the study, writes: “Importantly, AI content detection is not perfect, and the way we detect AI content will be different from how Google does (if it does).” That parenthetical is doing a lot of work. Nobody outside Google knows whether Google detects AI authorship at all, let alone acts on it.
He also sets out what the data is good for: “looking at large-scale datasets and directional trends, and not fixating on individual articles and their AI content.”
So the honest summary is narrower than the headline. A large study looked for a ranking penalty against AI content and found no sign of one. That is not the same as proof that none exists, and it is emphatically not a finding that AI content is good.
The finding people skip
Look again at that 54.7%. More than half of the pages ranking at the top of Google are under 20% AI content, and only about one in twenty is fully machine-written.
The pattern that wins is not “no AI” and it is not “all AI”. It is mostly human pages with AI somewhere in the process. That is the shape of the winning set, and it is worth sitting with, because it is not what either camp in this argument tends to claim.
What we would actually do with this
If you run a small business and you are using AI to help with your website, the evidence says the tool is not the problem. Three things still are.
Volume without review. The failure mode is not asking a model to draft a page. It is publishing forty of them without anyone reading them. AI makes it cheap to produce pages nobody checked, and pages nobody checked are how sites get into trouble.
Facts nobody verified. This is big. It’s a monumental problem because most AI output will include made-up “facts” unless the writing skill specifically addresses that issue. Most don’t. A model will produce a confident, fluent, wrong number, and it will read exactly like a correct one. On anything with a price, a date, a regulation, or a statistic, that is a liability, and it is the single most common problem we find when we audit AI-assisted content.
The part only you know. Your service area, your actual process, what your customers ask you on the phone, the job that went sideways in 2019, and what you changed afterward. A model cannot invent any of it, and it is usually the only thing on the page a competitor cannot copy that afternoon.
Use the tool for the draft and the structure. Keep a person on the facts and the experience. On the evidence, Google is not checking who typed it.
One thing worth doing this week
Pick the three pages on your site that matter most commercially, and read them start to finish as if you were a customer. Not for keywords. Read them to determine whether anything on the page could only have been written by your business.
If the answer is no, then that is the problem to fix, and it would still be the problem if a human had written every word.
If you would like us to check the page for you, get in touch, and we will look at your three pages and tell you what we find.
