AI SEO for law firms: be the answer in ChatGPT, Gemini and AI Overviews

What actually changes when a client asks an assistant instead of Google

A Google results page offers ten links and the reader chooses. An AI answer offers a recommendation, usually naming two or three firms, sometimes one. There is no second page, no scrolling, and no advertising slot to buy your way in.

That single structural difference changes the economics. In Google, being fourth still earns clicks. In an AI answer, being fourth usually means not appearing. For a sector where a client instructs one firm, the gap between named and unnamed is far wider than the gap between position one and position four.

The behaviour is moving quickly. BrightLocal's 2026 consumer survey found 45% of people now use AI tools to find a local business, up from 6% a year earlier. That is US consumers and local businesses generally rather than UK legal specifically, but the direction is not ambiguous and legal enquiries are not exempt from it.

Why ranking in Google does not get you named by AI

This is the finding most firms find hardest to accept, and it is the reason AI visibility needs deliberate work rather than being a by-product of good SEO.

Research across 8,090 keywords in 25 verticals found that Google's AI answers and standalone chatbots cited the same source only 7.2% of the time. They disagree in more than nine cases out of ten. The systems are drawing on different parts of the web and weighing different signals.

Separately, a Visibility 360 study of law firms found only 27% of established firms appear in AI-generated recommendations for prospective clients. That is a US vendor study rather than independent research, so treat the precise figure with caution, but it is consistent with what we see when we test UK firms: good, well-ranked practices that are simply absent from the answer.

The practical upshot is that you cannot infer your AI visibility from your Google rankings. It has to be measured directly, and most firms have never measured it at all.

The four things that decide whether a model names your firm

In order of how often they are the blocker.

• Crawler access. AI crawlers are separate from Googlebot and can be blocked separately, in robots.txt or invisibly at your CDN or firewall. Most AI crawlers also do not execute JavaScript, so a site that builds its content in the browser can be unreadable to them while looking healthy in Google.
• Extractable answers. Models quote passages that stand alone. A paragraph that answers the question completely, without depending on the three above it, gets used. A page that opens with positioning and reaches the point in paragraph four gets skipped in favour of a competitor who led with the answer.
• Entity consistency. Models build an internal picture of your firm from many sources. Contradictions weaken it. An old address on one directory, a former trading name on another, a phone number that changed two years ago: the firm reads as unverifiable, and unverifiable organisations do not get recommended.
• Third-party corroboration. The hardest and most decisive. A model deciding who to name leans on what independent sources say. A firm with no directory presence, no trade press and few recent reviews has given it nothing to work with, however good its own website is.

How we measure AI visibility, given there is no rank tracker

There is no position to track, because the answer differs between users, sessions and phrasings. So we use a fixed prompt panel tested repeatedly, which is the only method that produces a trend rather than an anecdote.

We build 20 to 50 prompts in your clients' words rather than keyword phrasing. For a conveyancer that means "who is the best conveyancing solicitor in Maidenhead", "how do I find a conveyancer for a leasehold flat", "cheapest conveyancing near me". Not "conveyancing solicitor Maidenhead", which is how an SEO would phrase it and not how a client would.

Each prompt is run across ChatGPT, Gemini, Perplexity and Google AI Overviews, and we record four things: whether your firm was named at all, where in the answer, how it was described and whether that description was accurate, and which firms appeared instead.

The single most actionable output is the list of sources the answers cite. If the same three domains keep getting cited for your sector's prompts and you are not on any of them, that is your outreach list, handed to you by the models themselves.

One caution on interpretation: a single test is noise. Month-on-month movement across a fixed panel is signal. Anyone showing you one screenshot of ChatGPT naming a firm is showing you a coincidence.

What is different about this for a law firm

Three things separate legal AI visibility from the generic GEO advice written for e-commerce.

Regulation constrains the copy, and helps. SRA rules require accuracy and prohibit misleading claims, which rules out the aggressive tactics some agencies use. It is also protective, because the attribution and precision compliance forces, named solicitors, real credentials, qualified claims, is close to what the models reward anyway.

Local intent dominates. Most legal queries carry an implicit or explicit location, and Gemini and AI Overviews lean heavily on Google's own local signals. So Google Business Profile quality and review recency feed AI visibility directly, not just the map pack. This surprises firms who assumed AI work was purely about content.

The single-recommendation problem is sharper here. Someone comparing suppliers will read five options. Someone asking an assistant which solicitor to instruct often receives one name. In a sector where the client instructs one firm, being second in the answer is worth far less than being second in Google.

What the work involves, month to month

The first month is diagnostic: confirm crawler access, baseline the prompt panel before anything changes so improvement is provable, and audit entity consistency across the SRA register, Companies House, Google Business Profile and the main legal directories.

From there it is content restructured for extraction, schema deployed across the site, entity contradictions corrected, and citations built from the sources your own prompt panel showed the models already trust. Then the panel is re-tested monthly against the same prompts.

Steps one to three usually produce the largest movement and take weeks rather than months. Citation building is where the time goes and where the durable advantage sits, because it is the part competitors cannot replicate quickly.