An AI will recommend a business that does not exist, and sound certain

Somebody in your town asked an assistant which braider to book. It answered in about two seconds, with three names, in the tone of a friend who knows the area. Last week two researchers checked whether names like those belong to anybody real.

The paper went up on 16 September. Hazem Ibrahim and Yasir Zaki asked three AI models for local providers across the hundred largest US metropolitan areas, in four fields that happen to have an official government register behind them — primary care doctors, hospitals, nursing homes and financial advisory firms — and then checked every single name the models returned against that register.

For primary care doctors, with the model working from memory alone and no access to the live web, 4% of the names were a clinician actually practising in the city the question named. On the larger commercial model with search switched off, 10.8%. Then they switched web search on, asked the same questions, and the figure went to 63.9%.

What was actually measured

That last sentence is the whole result. The model did not get cleverer between the second bar and the third. It got a way to check.

What this study is, and what it is not

Before anybody builds an argument on it, including me — this is a preprint. It has been posted publicly, not yet peer reviewed. It is one team, three specific model versions, and a collection window of two days in August. Models change weekly; the same audit in March could read differently in either direction.

And it is worth being exact about what it did and did not measure. It measured whether recommended providers exist in an official register. It did not measure websites, it did not test what makes a business get named, and it says nothing whatsoever about how many people are asking assistants for a braider rather than opening Maps. Anyone citing this paper as proof that you need a website — including anyone quoting it at you in a sales email this month — has gone further than the evidence goes.

What it does establish is narrow and solid: when these models cannot retrieve, their recommendations mostly do not correspond to reality, and they say so in exactly the same voice they use when they are right. In the authors’ words, an answer produced without retrieval “often carries no sign that its recommendations were never verified.”

The finding nobody is quoting

The 4% is the number that will travel. It is not the most interesting one in the paper.

In the financial advisory field, the researchers could check something beyond existence: whether a recommended firm had a misconduct disclosure filed against it with the SEC. Among the firms the model named with search switched off, 18.3% carried a disclosure — against a base rate of 5.1% across the register as a whole. Roughly three and a half times over-represented. With search enabled, it fell to 1.7%, below the base rate.

That is the difference between a model being unreliable and a model being biased in a particular direction. Working from memory, it reached for the firms that had generated the most written record — and firms generate written record by being large, by being advertised, and sometimes by being in trouble.

The restaurant half of the study makes the same point from the other side. There the researchers had no register to check against, so they asked a different question: what kind of place does it pick? Recommended restaurants had three and a half to five times the review count of a typical establishment — and a rating advantage of between 0.07 and 0.11 of a star. Essentially nothing.

It was not recommending the best restaurant. It was recommending the most documented one.

This next part is reasoning, not research

I want to mark the join clearly, because this is exactly where most writing about this study will quietly stop being about the study.

The paper does not say that having a website gets you recommended. It did not test that. What it establishes is a mechanism: these systems answer far better about providers they can retrieve, and when they cannot retrieve they fall back on whatever left the deepest impression in the text they were trained on.

From there it is ordinary logic, and you can check the steps yourself. Retrieval needs something to retrieve. If the only place your business exists is a profile inside somebody else’s app, then there is no page anywhere carrying your services, your hours, your area or a way to book you. You are not being passed over in favour of a competitor. There is nothing to pass over.

And the second step, which follows from the restaurant finding: the volume of consistent, public, checkable record about you is doing more work than you would like. That is uncomfortable — it means being findable in quantity can beat being excellent in silence — but it is what the data shows about how these systems choose.

It also rhymes with what Google publishes about its own AI answers: no special markup, no secret file, nothing to buy — just pages that are indexed, crawlable and readable as text.

What that makes worth doing

None of this is exotic, and none of it requires buying anything from anybody:

• Have a page that states the plain facts. What you do, the services by name, the area you serve, your hours. In text, on a page, not inside an image or a PDF.
• Make the records agree. Same business name, same address, same phone number wherever you appear. Contradiction between listings is what makes a machine hesitate to name you — and the same thing makes a human hesitate.
• Complete the free listings you already have. Category, hours including holidays, services written the way a customer would say them. The free setup guides walk through claiming the accounts you should own whoever builds your site.
• Keep asking for reviews. Not for the star average — the study suggests that moves almost nothing — but because the volume of public record is what gets you into the set of things that can be retrieved at all.

And whose address it lives at

There is a second question underneath this one, and it outlasts any particular model. Whatever page represents you — who holds it?

On most subscription platforms you are licensing a site that works while you pay and stops when you stop. The source code is not yours to take. The domain is sometimes registered to the platform rather than to you. Years of paying produce no asset, only a bill that renews.

A site you own means all of it is in your name: the domain, the hosting account, the source code, and every connected account — payments, email, analytics. Nobody can switch it off, raise the rent, or keep your customer list.

That is what I build for small service businesses: a real site at an address you control, with your services and hours written where they can be read, online booking that takes the deposit at the time of booking, and an owner’s dashboard. muriellehairbraids.com is one running live, taking real bookings and real payments. Open it on your phone and go as far as the deposit screen.

The models will keep changing. What does not change is that a system which answers by retrieving can only offer what somebody actually published — and that the page it finds should belong to you.

Common questions