What actually happened in the experiment?
Rick Tousseyn at OtterlyAI bought two fresh domains, pointed an AI writing tool at them, and published roughly 1,000 blog posts on each in under a day. No human editing. The subject was capybaras, chosen because nobody was competing over it. Both sites went live in early April 2026 and were tracked until late July.1
The first site did well for a while. Google indexed 960 of its pages and it grew to over 21,000 weekly impressions by mid-May. Then on 26 June it fell from 859 daily impressions to 85, and within days to single digits. The second site died faster: it peaked at 2,426 daily impressions on 5 April and dropped from 1,629 to 15 on 9 April.1
The part worth pausing on is how it ended. Google Search Console recorded zero manual actions on either domain. No penalty notice, no message, nothing to appeal. The traffic simply stopped.1
Google does not need to send a notice. Its spam policy covers what it calls scaled content abuse, defined as generating many pages "for the primary purpose of manipulating search rankings and not helping users", and it names "using generative AI tools" as one way people do it. The stated consequence is that sites "may rank lower in results or not appear in results at all".2
One thing to hold in mind while reading the rest: OtterlyAI sells AI visibility tracking, so a study showing AI visibility behaves differently from Google is useful to them commercially. The underlying figures come from Search Console, Bing Webmaster Tools and server logs, which are checkable, and the author is open about what the experiment failed to isolate. Take the numbers, and treat the framing with the usual caution.
Which engines cited it, and which ignored it?
This is the finding that gets quoted badly. "AI search did not penalise it" is true and almost useless, because the citations were not spread evenly. They landed on three engines and were absent from four.
| Engine | Site B | Site A |
|---|---|---|
| Microsoft Copilot | 2,718 | 543 |
| Claude | 714 | 1,005 |
| Perplexity | 501 | 628 |
| Google AI Mode | 0 | 19 |
| Google AI Overviews | 0 | 11 |
| ChatGPT | 0 | 0 |
| Gemini | 0 | 0 |
Copilot was the outlier, and by August it had gone further. Bing Webmaster Tools credited one of the domains with 42,817 Copilot citations, making it the single most cited domain in its whole category at 14% share, ahead of Wikipedia and Britannica. Bing had never deindexed it, and gave it 91,104 impressions in July alone.1
That is not a coincidence. Microsoft's own documentation states that Copilot "use[s] generated search queries sent to the Bing search service to ground responses in web data".3 Copilot inherits whatever Bing has indexed. Bing kept the sites, so Copilot kept quoting them, long after Google had made the same pages invisible.
Why did ChatGPT visit 4,808 pages and cite none of them?
Because being crawled and being included are two different things, and it is easy to mistake one for the other when you are reading server logs. ChatGPT-User made 4,808 page visits across the two domains, more than every other crawler combined, and produced almost no citations.1
OpenAI runs separate crawlers for separate jobs, and says so plainly. OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features". ChatGPT-User is something else: when a user asks a question, ChatGPT "may visit a web page with a ChatGPT-User agent".4 One builds the index. The other is a courier sent out mid-conversation.
GPTBot arrives
It gathers material that may be used to train the models. Nothing to do with whether you appear in an answer.
OAI-SearchBot arrives
This is the one that decides whether your pages can show up in ChatGPT's search results at all.
ChatGPT-User arrives
Someone has asked a question and ChatGPT has gone to fetch a page in the moment. That is a visit happening, and nothing beyond it.
You read the logs
Heavy crawler traffic looks like progress. In the experiment it came with zero citations.
If you take one operational lesson from the experiment, it is that crawler hits in your logs are not a scoreboard. We made a related point in what actually makes an AI name your business: the engines are reading the open web as it already describes you, and the traffic they generate on the way in tells you very little about the answer that comes out.
How many UK adults are actually on each engine?
This is where the experiment stops being a curiosity and starts being a decision. The engines that rewarded the AI content are small in the UK. The one that ignored it entirely is not.
Ofcom, using Ipsos iris measurement of UK internet users aged 18 and over, reports that in June 2025 12,990,346 UK adults visited ChatGPT, against 2,551,353 for Microsoft Copilot, 1,745,485 for Google Gemini, 560,979 for Perplexity and 345,735 for Claude.ai.5
UK adults visited ChatGPT in June 2025, roughly five times the reach of Copilot. It cited the AI-written sites zero times.
Read those figures carefully, because they are narrower than they look. Ofcom is counting visits to the standalone websites and apps, so the Copilot number excludes Copilot answers inside Bing, and the Gemini number excludes Gemini answers inside Google Search. Nor are the figures additive, since plenty of people use more than one. And a visit to ChatGPT is not necessarily a search.
Gemini's real footprint is the caveat that matters most. Separately, Ofcom reports that about 30% of searches now show an AI overview, and 53% of adults say they see these summaries often.6 Google's AI answers reach people who never chose an AI product at all, and those answers cited the test sites 30 times in four months.
The wider adoption picture is not in doubt. Ofcom's survey of 7,533 UK adults, with fieldwork from 29 September to 28 November 2025, found 54% of UK adults now use AI tools such as ChatGPT, Copilot or Gemini, rising to 79% of 16 to 24 year olds.7 The audience is real. It is just concentrated somewhere the AI-written content never reached.
Does any of this apply to a plumber in Preston?
Less than you would hope, and it is worth being straight about why. The experiment tested informational content in a category with no real-world businesses in it. Nobody asks an AI which capybara site to hire.
The capybara sites
- No real-world entity behind the domain
- No address, phone number or trading history
- No reviews and no Google Business Profile
- Competing on informational questions
- Nothing to lose if the domain disappears
A UK service business
- A named, registered business people can verify
- An address, a phone number and a service area
- Reviews, and a profile the engines already read
- Competing on "who should I call" questions
- One domain, and years of custom attached to it
Two other limits are worth stating. The sites cannibalised each other, which the author says made a second variable in the design impossible to isolate. And the whole thing ran on English-language, US-centric queries in an uncontested niche, so it says nothing about how an engine chooses between four roofers in Cardiff.
What does transfer, exactly and without adjustment, is the downside. Google's own guidance on AI content says the concern is content produced with "little to no effort, little to no originality, and little to no added value", and that generating many such pages "may violate Google's spam policy on scaled content abuse".8 A capybara site can absorb that. A business whose enquiries come through one domain cannot.
What we found on 27 UK service business sites
We had assumed the common failure was accidental blocking, so we went and checked. It is not, at least not in this sample. On 22 August 2026 we took 31 independent UK service business websites, found through ordinary searches across six trades and six cities, and fetched the robots.txt and homepage of each. Twenty-seven returned both.
Not one of the 27 blocked a single AI crawler. Not OAI-SearchBot, not GPTBot, not PerplexityBot, not ClaudeBot, not Google-Extended. In fact none of them mentioned OAI-SearchBot at all, which means every one of those sites is fully open to the crawler that decides ChatGPT inclusion, entirely by accident.
Volume was not the problem either. The median homepage carried 1,067 words of readable text in the raw HTML, before any JavaScript ran. These businesses are not short of words. What they were short of was the plain facts an engine needs to tell one of them apart from another.
Twenty-seven sites is not a study and we are not presenting it as one. They were picked from search results rather than sampled at random, they lean towards businesses that already rank, and a robots.txt that permits a crawler is no guarantee the crawler ever calls.9 Missing structured data is also not a penalty: Google is explicit that no special markup is required for its AI features.
But the shape of it is hard to miss. Ten of the 27 never say what kind of business they are in a form a machine can read. Fourteen never give their hours. If an engine has to choose between two roofers, one of whom has stated its address, its hours and its rating, and one of whom has 1,000 words of prose and none of those things, the choice is not really about the prose.
So what should you actually publish?
Less than you think, and almost none of it generated. The three things that move first are boring and you can do them yourself this week: claim and complete your Google Business Profile, make sure your address, phone number and opening hours appear identically everywhere they appear at all, and ask your last twenty customers for a review. None of that requires an agency and all of it outranks a blog schedule.
After that, write the pages that answer what people actually ask you on the phone. What you charge, which postcodes you cover, how quickly you can come out, what happens if it goes wrong. One honest page of that beats a hundred generated ones, and it is the sort of thing a model cannot produce on your behalf because it does not know your prices.
Frontpage AI is a UK studio that builds and manages websites for independent service businesses, structured so that Google and AI answer engines can read, trust and name them. We work only with UK service businesses such as trades, clinics, salons and instructors, and charge from £148 per month on a 30 day rolling contract with no setup fee. We do not sell AI content packages, for the reason this whole article describes. If you want the detail of how we approach it, that is on our GEO page, and the results we can actually show are on partners.
The thing the experiment really demonstrates is not that AI content works or does not work. It is that a domain can look healthy on one set of dashboards while it is being quietly erased on another, and nobody sends you a letter when it happens. That is a bad trade for a business with one website and one phone number.