GEO and lawyers: the algorithm at the heart of legal advice

GEO (Generative Engine Optimization) is the set of methods that make a law firm citable by AI engines such as ChatGPT, Perplexity and Gemini, now the first place clients turn to find a lawyer.

A company director looking for a lawyer in 2026 rarely starts with Google. They ask ChatGPT. Perplexity. Gemini built into Gmail. And the AI answers with two or three names, presented as a direct recommendation — not a list of links to check, an answer.

The filter has moved. It used to be Google's algorithm that decided which firms appeared on page one. Now it's a language model that decides which firms deserve to be cited. The discipline that has grown up around this new arbitration has a name: GEO, for Generative Engine Optimization. And the first figures available on the French legal market show it rewards nothing like the same players as traditional search ranking.

What the numbers say

The KAIS 2026 barometer ran 30 legal queries — fifteen framed from the point of view of a client looking for a lawyer, fifteen from the point of view of an investor or journalist sizing up the landscape — through ChatGPT, Gemini and Perplexity. 1,870 citations aggregated. A score normalised to 100, the KAIS Visibility Score, rebased to the leader of each sector.

This ranking tells a story no professional directory had told before. Firms that have dominated Chambers and the Legal 500 for twenty years find themselves relegated behind a legaltech built for sole traders. The reason isn't the quality of their work — it never was. It's the way they talk to machines.

How a model chooses whom to cite

An LLM doesn't read a website in isolation to form its answer. It cross-references several sources: the firm's profile, partners' LinkedIn pages, bar directories, press mentions, leading legal reviews, structured data present in the pages. Everything that makes up what engineers call an entity graph.

Three conditions recur in every case where a firm comes out on top.

1. Semantic consistency

If a partner is presented as an M&A specialist on the firm's site but communicates about employment law on LinkedIn, the model detects a mismatch and drops the source. The precaution is built into the architecture of LLMs — they'd rather not cite than cite wrong. For a firm, that means the partner profile, the LinkedIn page, the byline on op-eds and the biography published in the press all have to say exactly the same thing, in the same terms.

2. Precision of vocabulary

"Business lawyer" has become an empty term. Models see that keyword go by on ten thousand identical pages and no longer know what to do with it. The firms that have climbed over the past eighteen months write differently: "cross-border intellectual property litigation in the SaaS software sector," "post-acquisition disputes under German law," "AI Act compliance for developers of generative models." The closer the vocabulary sits to a real use case, the more readily the model finds its answer in that source rather than another.

3. The external authority graph

An op-ed in Les Échos, an analysis in Dalloz Actualité, a contribution to the Revue Lamy Droit des Affaires carry more weight than a post on the firm's own blog. The model recognises a hierarchy among sources. Claiming to be an expert isn't enough. Being named as one by an external node of authority changes the score.

The technical detail that costs dearly

Between 2023 and 2024, many firms blocked OpenAI's crawlers, fearing their analyses would be scraped and served up without any return. The instinct made sense. Two years on, the result is that they're no longer indexed at all by the sources ChatGPT consults in real time through SearchGPT. A miscalibrated robots.txt file now costs more than the intellectual property it was meant to protect.

Allowing the crawl without tagging your pages with the right microdata produces the opposite effect, and one just as costly. The model sees the content but has no way to interpret it. Schema.org's LegalService markup, paired with Attorney and FAQPage, tells the machine explicitly: this is an article written by a lawyer registered with the Paris bar, specialising in business law, analysing a given ruling of the Court of Cassation. The code stays invisible to the human reader. To the model, it's the difference between a page that's legible and a page that's indexable.

The prestige trap

The most unsettling lesson in the barometer concerns the international firms. Linklaters, Clifford Chance, Allen & Overy: objectively among the most powerful in the world. On French-language queries, they come out behind a legaltech that sells service contracts for 49 euros.

The explanation isn't mysterious. Their communications strategy stays Anglo-Saxon. Their editorial output targets Chambers and the Legal 500. Their Schema.org markup often amounts to a bare corporate profile. They're recognised by their peers. They're not indexed by French-language LLMs.

For a director asking Perplexity which firm to consult on a tax dispute, the real quality of Linklaters' expertise makes no difference. What the director sees is an answer the firm doesn't appear in. Over a quarter, that adds up to dozens of prospects who never walk through the door.

What a niche firm can do

The useful lesson in the barometer concerns neither Bredin Prat nor Gide — their positions are locked in for the next two or three years. It concerns every firm ranked between 10 and 30, with real expertise and AI visibility close to zero.

For these firms, producing more generalist legal content is a dead end. The ground is saturated. The lever that pays is the precise vertical. Cybersecurity and data protection. Renewable energy. International succession. Healthcare M&A. AI Act compliance. A firm that publishes two or three times a week on a single one of these niches, with the right technical markup and a consistent presence on LinkedIn and in the specialist press, can move from a score of 5 to a score of 40 on its vertical in nine months. The first results land within thirty to sixty days, the time it takes SearchGPT and the Perplexity index to refresh their reading of the sources.

For a focused firm, showing up in forty to fifty answers a week on high-intent commercial queries represents a flow of qualified prospects no brochure site will ever generate. The profile of the generative-AI user plays fully in your favour here: they're not scanning results, they're asking for a recommendation. The conversion rate of an AI citation regularly beats that of a Google click on the same query.

The profession facing a new arbiter

GEO isn't a marketing fad. It's the technical response to a durable shift in client behaviour. Generative models are settling into the tools directors use every day — the inbox, the browser's search bar, the office suites. A legal question asked out loud to a built-in assistant will never return a page of results. It'll return a firm, or none.

Two or three years from now, the professional rankings will keep coming out. The question is how many clients will still read them, and how many will simply have asked a model to recommend them a lawyer.