What the research says about AI search and gambling

AI assistants now answer gambling questions that people once typed into a search box, and the published evidence on how they do it is not reassuring for regulated markets. Across the three studies we review below, chatbots frequently recommended operators without a local licence, sometimes described them as legal, and occasionally explained how to reach sites outside self-exclusion schemes. For a licensed casino or sportsbook, that means AI search is less a new traffic channel than a new place where unlicensed competitors and outdated facts can appear next to, or instead of, its name.

This review sets out what each study did, what it found and how far its findings can be stretched. None of them is a peer-reviewed measurement, and all three were run in Europe or the UK.

Investigate Europe: seven chatbots, ten countries

The largest of the three was published on 9 March 2026 by Investigate Europe with media partners including the Guardian and Público. Over two weeks, reporters prompted seven assistants (Meta AI, ChatGPT, Grok, Google Gemini, Microsoft Copilot, Claude and Le Chat) in the national languages of ten European countries. The prompts were deliberately risky: casinos with the biggest bonuses, sites without age verification, sites outside self-exclusion schemes such as GamStop, and casinos with "fast payouts".

According to the report, about three quarters of replies recommended sites not licensed in Europe. When asked for sites outside self-exclusion schemes or without identity checks, Grok, Meta AI and Gemini recommended unlicensed operators in every country tested. In Greece, ChatGPT listed ten casinos that all appeared on the Greek regulator's blacklist. Claude was described as the assistant most likely to avoid naming unlicensed operators. The report does not state the total number of prompts, which limits how precisely the "three quarters" figure can be used.

KVA in the Netherlands: checked against the licence register

The most methodical study came from the Netherlands. Keurmerk Verantwoorde Affiliates (KVA), a Dutch certification body for responsible affiliates, tested ten tools: ChatGPT, Google AI Mode, Gemini, Copilot, Perplexity, Claude, Grok, DeepSeek, Mistral Le Chat and Meta AI. Its questions were realistic Dutch consumer queries plus research-style sequences, including requests for casinos that do not require registration with Cruks, the Dutch self-exclusion register. Every operator named was checked against the Kansspelautoriteit licence register.

The results, reported by Gambling.com on 2 October 2026, were stark. All ten tools named at least one operator without a Dutch licence. Eight did so in response to neutral or leading questions, without any research framing. Four surfaced unlicensed operators after a single question. Three (Copilot, Le Chat and Gemini) presented unlicensed brands as legal or approved by the regulator, and Copilot produced a "top 50 legal casinos" list that included unlicensed operators. KVA itself describes the sample as indicative rather than exhaustive.

Casinos.com: a small UK test from an affiliate

The third study is the weakest evidentially and worth reading with its source in mind. Casinos.com, a UK casino review site that earns from operator listings, asked ChatGPT, Grok and Gemini how UK players can reach offshore casinos, then followed up. Its article, updated on 16 March 2026, reports that ChatGPT stressed licensed casinos and declined to explain how to bypass geolocation, while Grok produced a list of offshore sites when pressed. No sample size or scoring is given. It is a useful illustration, not a measurement.

How the studies compare

What this means for licensed brands

Two practical conclusions follow, and one thing does not.

First, accuracy is now a brand risk as well as a marketing one. If an assistant tells a Dutch user that an unlicensed site is approved by the regulator, the licensed operators in that market are competing with misinformation as well as with rivals. Brands have no direct way to edit AI answers. The levers they do have are their own pages (a clear licence and registration page, dated and linked to the regulator) and the third-party sources assistants draw on.

Second, the risky prompts in these studies are not marketing prompts. Questions about avoiding age checks or self-exclusion are ones licensed brands should never try to rank for. The relevant question for them is narrower: when a user asks a neutral question, such as which sites are licensed in a market, does the answer name licensed brands and describe them correctly?

What the evidence does not support is any claim that a particular tactic gets a brand recommended. Separate research on answer stability, such as SparkToro's 2026 study finding that ChatGPT and Google's AI gave the same brand list less than 1% of the time across repeated runs, suggests that single answers should never be treated as rankings.

What Google has and has not said

Google's documentation on AI features says pages need only be indexed and eligible for a snippet to appear in AI Overviews and AI Mode, and that no special markup is required. It has published nothing specific on how these features treat gambling queries. Since 31 August 2026, sites can opt out of Google's AI features through Search Console without affecting normal rankings, a control that came out of UK competition requirements. Our timeline of Google updates and policy changes tracks these developments.

The gap in the evidence

All three studies were European or British. We found no comparable published test for US states or Canadian provinces, where legality differs by state and province and where an answer such as "online casino is legal in the US" is simply wrong. That is the gap our own protocol for checking AI answers about gambling legality is designed to fill, using our US legal states tracker and Canada guide as ground truth.

Our market comparison shows how different the underlying rules are between countries, which is one reason a single global answer from an assistant is so often wrong. For teams deciding what to do now, the most useful step is modest: ask the main assistants the neutral legality and licensing questions for each market you operate in, record the answers with dates, and compare them with the regulator's register. That gives a factual baseline before any money is spent trying to change the results.