AI search optimisation for law firms: what actually gets a firm recommended

If you take one thing from this page: being named by an AI and being recommended by it are different outcomes, and most law firms are optimising for the wrong one. In Avouch's July 2026 sweep of UK personal injury — ten buyer questions put to ChatGPT, Claude and Gemini — one national firm was named in eight separate answers and recommended in none of them. In the same sweep, Irwin Mitchell was named fifteen times and recommended twelve. Same market, same questions, same week.

That gap is the whole game, and it is invisible to every visibility tool that counts mentions. A dashboard showing eight mentions looks like a firm doing well. What it actually describes is a firm the engine knows about and declines to put forward — which, commercially, is closer to being absent than to being recommended.

Named is not chosen

The two words mean different things and the gap between them is the finding. Named is any answer that mentions the firm at all, however briefly — one line in a list of five options. Recommended is narrower: the engine puts that specific firm forward as its pick, the one it tells the buyer to go with. A firm can rack up mentions for months without ever being the second kind.

Avouch swept two separate UK legal markets on 23 July 2026 — personal injury and commercial debt recovery — with ten buyer-intent questions each, put to three engines with live web search. Forty-two firms surfaced in the personal injury market alone, across 110 distinct source domains. Below is what the conversion from named to recommended actually looked like. The firms shown converted well; the ones that converted poorly are left unnamed, because the pattern is the point and their identity is not.

Irwin Mitchell converted 80% of its mentions into recommendations and Lovetts 88%, while one firm named nearly as often as Irwin Mitchell converted a quarter of its mentions, and another converted none at all. Nothing about the questions changed between those firms. What changed was the evidence available for the engine to point at.

What the engines actually cite as the reason

Because Avouch stores the full answer text, we can read the sentence that does the recommending and see what it leans on. Across both legal markets the same three things appear again and again: an independent directory ranking, review volume, and a specific verifiable number. Here is Claude, unprompted, in our sweep:

“Hugh James is ranked in both Chambers and Legal 500, is the only top-ranked firm in both directories for medical negligence in Wales, and recovers over £30 million a year in compensation for clients.”

Every load-bearing clause in that sentence is externally checkable: two directory rankings, a scope claim, and a figure. That is what a recommendation is built out of. Legal 500 was cited five times in the debt recovery sweep and Chambers four times in personal injury — small numbers against 110 source domains, but they sit on the recommending sentence far more often than their citation count suggests. Trustpilot did similar work where review volume was the differentiator, with one answer citing firms at 5.0 from 600+ reviews and 4.9 from 400+.

The engines also reach for regulatory status as a trust signal without being asked. ChatGPT cited the SRA's own no-win-no-fee guidance directly in one answer, and Claude drew the distinction between regulated firms and claims-management companies using professional indemnity cover — noting that solicitors must carry at least £2 million and are obliged to give best advice. If your firm's regulatory position and cover are not stated plainly somewhere an engine can read them, you are absent from the comparison that decides those questions.

Your own site is cited most, and decides least

The single most-cited domain in the personal injury sweep was a law firm's own website — twelve citations, more than Chambers, Legal 500 and Trustpilot combined. Firms' own sites dominated both markets' source lists. That sounds like good news until you set it against Avouch's larger research: across 1,828 recommendations we re-examined, a recommended firm's own website was never once the only source behind the recommendation, and it is cited about as often when a firm is merely listed as when it is recommended.

So your website is not optional and it is not sufficient. It gets you into the citation list; the third-party layer sitting alongside it decides what the engine does with you. For a law firm that means the directory profile, the review base and the coverage are not marketing extras — they are the evidence the recommendation is built from. The practical order of work is: make your own pages the kind an engine can point at as a reason, then make sure the independent sources it checks agree with them.

What this means for a UK firm, under the SRA

• Put your Legal 500 and Chambers positions where they can be read as text, not baked into an image or a badge. They are the most load-bearing third-party signal in this sector and engines quote them verbatim.
• State one specific, verifiable number about outcomes — the kind of figure a directory or your own published results already support. Vague superiority claims are unusable to an engine and, under the SRA Code, unwise anyway.
• Publish your regulatory facts plainly: SRA number, professional indemnity position, complaints route. Engines reach for these to separate solicitors from claims-management companies, and they cite the regulator's own guidance while doing it.
• Build the review base. Volume did visible work in our sweep — the answers that leaned on reviews quoted counts, not just scores.
• Do not claim to be the best. The SRA restricts unsubstantiated superlatives, and separately, an engine cannot use a claim it has no way to verify. The constraint and the optimisation point the same way here.
• Measure recommendation, not mentions. If your reporting cannot tell you the difference between eight mentions with no recommendations and eight mentions with six, it is not measuring the thing that produces clients.

How this was measured

Two UK legal markets, ten buyer-intent questions each, put to ChatGPT, Claude and Gemini with live web search on 23 July 2026 — 42 firms surfaced in personal injury across 110 source domains, and 74 firms in commercial debt recovery across 98. Every answer was read and normalised into named versus recommended, and the underlying verdicts are stored rather than summarised, so any figure here can be traced back to the answer that produced it. Per-engine results are held separately and never blended into an average, because in Avouch's wider research the same source has ranged from deciding nothing on one engine to deciding roughly half of answers on another.

“There's a difference between showing up and being recommended. AI cannot recommend what it doesn't know.”