Three quarters of the recommendations went to sites you compete with illegally

Across ten European countries and thirty prompts per platform, Investigate Europe found that roughly three quarters of chatbot replies recommended gambling sites with no licence in the country being asked about. Some of those sites appear on regulators' own blacklists. In France and Poland, eighteen of twenty-one prompts returned an unlicensed operator.

What was actually tested, and by whom

Two separate pieces of work, published within a day of each other in March 2026, and worth separating because their methods differ.

The Guardian's UK test is the one most widely quoted. The Investigate Europe dataset is the more useful one commercially, because thirty prompts across ten markets in national languages is close to how an operator would measure its own visibility.

The numbers, by platform and by country

Read that as an operator rather than as a reader. In France, a licensed brand is not competing for twenty-one recommendation slots. It is competing for three.

Why do AI models recommend unlicensed casinos over licensed ones?

Not because the models prefer them. Because those operators produce more of the content a model can retrieve, and because the language they publish is the language a recommendation is made of.

• /Offshore operators publish aggressively and cheaply, with no compliance read slowing anything down. Volume reaches a training set.
• /The marketing language is superlative and comparative, which is exactly the shape a model reaches for when asked which is best. Investigate Europe found chatbot replies repeating offshore operators' own marketing wording.
• /Licensed operators are restricted in what they can claim, so their content is flatter and less quotable. Compliance is a competitive disadvantage in a retrieval system that does not know what a licence is.
• /Curaçao and Anjouan registrations are cheap and fast, so the volume of sites carrying that content is large and growing.
• /There is nothing in the retrieval chain that checks a licence register. A model has no reason to know which of two similar-looking casinos is legal in the country it is answering about.

What does this cost a licensed operator?

The consumer harm is the serious part and it is being handled by people better placed than an SEO agency. The commercial consequence is not being discussed at all, so here it is plainly.

• /Every recommendation that goes to an unlicensed site is acquisition you paid a licence fee to be eligible for and did not get.
• /The player who follows one is harder to win back, because they are now in a funnel with looser checks and larger bonuses than you are permitted to offer.
• /Your competitor set in AI answers is not the licensed operators you benchmark against. It is a much larger pool that does not share your constraints.
• /None of this appears in a rank tracker. A brand can hold position three organically and appear in none of the recommendations for the same question.

There is a regulatory dimension for operators too. The investigation was cited in the House of Commons on 19 March 2026, and a UK government spokesperson said AI platforms must protect users from illegal content under the Online Safety Act. Where that lands is not knowable yet, but the direction is toward platforms carrying more responsibility, not less.

What can actually be done about it

Less than the market will tell you, and more than nothing. Four things, in the order they are worth doing.

• /Measure it in the languages your markets speak. A prompt set written in English will not reproduce what a Polish or French player is shown, and the country-level spread here is wide.
• /Get the licence into the sources a model can reach. A regulator register entry, trade coverage naming the licence number, and a page stating it in full are all retrievable. Almost no licensed operator publishes its authorisation number in a form that survives retrieval.
• /Publish the things an unlicensed site cannot. Verified withdrawal times, real limits, the outcome of a failed verification. These are checkable claims, and checkable claims are what a model cites.
• /Report unlicensed recommendations to the regulator in your market. Several already collect them, and a recommendation naming a blacklisted site is evidence.

What will not work is any of the usual answers. Schema does not tell a model which licences are valid. An llms.txt file has not been confirmed as read by any major provider. And no amount of on-page work outweighs a competitor set that publishes ten times the volume with none of the constraints.

The measurable part is the part worth buying. How we run it, including the prompt bands and the counting rules, is in the published methodology, and the reason retrieval and citation are reported separately is set out in the denominator piece.

Sources for the figures here: platform and country breakdowns from Investigate Europe, March 2026; the five-platform UK test and the quoted chatbot responses from The Guardian, 8 March 2026; the parliamentary citation from Hansard, 19 March 2026. Where the two investigations report different figures it is because they used different prompt counts, which is stated in the table above.

Related on this site

• /AI Visibility Index methodology — the prompt bands, sampling and counting rules, published in full.
• /When ChatGPT does not search — 57.8% of runs answer from training data, which is where offshore volume lands.
• /The denominator problem — why a visibility percentage needs its pool stated.
• /Compliance rules — the eleven jurisdictions we decline and why licences are verified before any work.
• /Generative engine optimisation — entity work and third-party presence, which is what moves this.
• /LLM visibility monitoring — retrieval, citation and share across six surfaces.
• /Responsible gambling content — signposting as a deliverable rather than a footer link.
• /Free AI visibility audit — your own share, in your own market, in 48 hours.

Questions