Winning clients via ChatGPT: the first contact law firms never see

Your first contact moved, and no dashboard reported it

Hiring a lawyer used to start with a search box, ten blue links and an afternoon comparing three firm websites. Now the first move is to describe the whole mess to ChatGPT: 'I was fired with no notice yesterday — what are my options and who do I call?' The model lays out the legal position, flags the deadline that matters, and then names places to start. That exchange is the new first contact, and it is over before anyone has seen your website.

The shortlisting happens inside that conversation. If the model names three family-law firms in your city and yours is not one of them, you lost the pitch without hearing about it. There is no position-11 report and no impression count to check: OpenAI, Anthropic and Perplexity publish neither a ranking formula nor query data for what they cite. Ahrefs found that roughly 80 percent of the URLs ChatGPT cites do not rank anywhere in Google's top 100, which is how a firm can look healthy in every dashboard it owns and still be absent here. That is the blind spot for firms still working only on classic SEO.

What actually decides which firm ChatGPT names

Answer engines like ChatGPT, Gemini or Perplexity answer from two places: what the model absorbed during training, and pages it fetches live at question time. The split matters more than most firms assume. Profound looked at roughly 730,000 US English ChatGPT conversations from late 2025 and found only about 18 percent trigger a web search at all — the rest are answered from training data alone. So you are playing two games at once: being retrievable today, and having been described often enough, in enough places, that the model already knows what you do. A road-traffic defence lawyer whose own site, register entry, directory listings and by-lined articles all say the same thing wins both.

What the model weighs is not one keyword on one tuned landing page but whether your sources agree about you. Ahrefs, across about 75,000 brands, found that how often a brand is mentioned around the web correlates with its AI citation rate at 0.664 — roughly three times the correlation for backlinks, at 0.218. Mentions, and the agreement between them, are the most evidence-backed lever anyone has actually measured. So when your homepage says you advise on everything, your Google Business Profile says criminal defence and a directory profile says medical negligence, no picture forms, and the firm with one unmistakable answer gets named instead. A narrow focus stopped being a marketing sacrifice and became the price of being legible at all.

There is no ad account for this. Nobody sells you a place among the sources an answer cites, and no engine publishes the weighting behind that choice, so every confident claim about a specific formula is somebody's inference. What is left is the substance and structure of what you have already published, and that asymmetry favours you: a three-partner firm with one clear specialism can beat a fifty-lawyer practice that describes itself in generalities.

Run the queries your clients run, then read what comes back

Do this before you read anything else. Open ChatGPT and type the sentences your clients actually type: 'I got a speeding notice, do I need a traffic lawyer?' or 'which firm here handles contested wills and forced-share claims?' or 'who defends against a copyright infringement letter?' Ask each one in a fresh chat, because citation rates fall as a conversation runs on — Profound measured citations in 12.6 percent of first-turn exchanges and 3.0 percent by turn 20. What comes back is the least flattering and most useful description of your firm you will read this year.

You will get one of three results. You are absent, and the work has not started. You are named for the wrong thing — tenancy disputes when construction claims are the whole practice — which is worse than absence, because a wrong description travels. Or the model dodges to a generic referral: the bar association, some lawyer-finder service. Take the middle case seriously. Columbia's Tow Center asked eight AI search tools to identify 1,600 news articles and more than 60 percent of the answers were wrong, while ChatGPT signalled any uncertainty in only 15 of 200 attempts and never once declined to answer. These systems state wrong things about you with complete composure.

The test costs nothing and no consultant can spin it. Write down ten questions a real client would ask in each of your practice areas, run them through ChatGPT, Gemini and Perplexity, and save the answers with the date attached. That file is your GEO baseline.

Accreditation and a narrow focus: the strongest lever you own

Few professions hand the machines cleaner signals for AI visibility than law. Formal specialist accreditation — an accredited specialist in family law, a board-certified employment lawyer, whatever your jurisdiction calls it — is a protected, checkable label tied to exactly one area of practice. That is the shape of fact a retrieval system can match with confidence. Repeat that title identically on your own site, in the public register that granted it and in every directory that lists you, and it becomes the single strongest thing known about you in machine-readable form.

Most firms then bury it. 'Lawyers for every stage of life' sounds welcoming and gives a model nothing to work with. Answer three questions instead, in plain words, high on the page: which two or three areas of law you actually work in, which region you cover, and who you act for — individuals, founders, mid-market companies, insurers. The tighter that triangle, the more often you are the obvious name for a matching question.

Then describe situations rather than statute numbers. 'We act for employees in unfair dismissal claims, settlement agreements and severance negotiations' is retrievable; 'advice in individual employment law' is not. Write down the moments in which somebody actually reaches for the phone, in the words they would use to describe them.

The pages that get cited: real questions, dated answers

A retrieval system is hunting for a passage it can lift and attribute. Question as the heading, answer in the first sentence, reasoning underneath — that shape survives extraction. 'How long do I have to challenge a dismissal?' answered in one line, with the deadline stated and dated, is worth more than a 2,000-word essay on employment law. Ahrefs, across 1.4 million real ChatGPT prompts, found the model cites roughly half the URLs it retrieves, so being fetched is not the finish line. Being the cleanest available answer is.

The conduct rules push you in the same direction. You may inform and explain; you may not promise outcomes or advertise in a way that creates unjustified expectations. Factual, sourced, precisely written legal information is what your regulator wants and what a model can quote without exposure. Writing to the standard your bar already imposes is most of the optimization.

One counterintuitive detail from the same Ahrefs data: cited pages had a median age of about 500 days. Freshness on its own does not win citations; a page that has existed, been linked to and been kept correct for a year and a half does. So publish the specific thing — 'speeding penalties in 2026, with the current bands' rather than a page called 'traffic law' — put a visible last-reviewed date on it, and update that page when the law moves instead of writing a new one beside it.

One set of facts, identical in every listing

The picture of your firm gets assembled from fragments: your own site, your Google Business Profile, legal directories and review sites, the bar or law-society register, local business listings, press mentions. When those fragments disagree about your name, your address, your opening hours or which areas of law you cover, the model has no settled fact to state, so it hedges or fills the gap itself. Schema.org markup earns its keep here, because it makes details like hours and location unambiguous to a parser. Google has said since 2018 that it is not a ranking factor, so treat it as extraction hygiene rather than leverage.

Work through the listings one at a time with a single reference record open beside you. Same firm name, same street address, same phone number, same hours everywhere. The same two or three practice areas, in the same words. The accreditation spelled the way the register spells it. This is unglamorous administrative work and it is usually the highest-return hour in the whole project, because it removes the contradictions that were quietly disqualifying you.

Reviews and third-party mentions do the same job from outside your control. When client after client writes 'she got my divorce through without a hearing,' that repetition is the mention signal Ahrefs measured, arriving in language you did not write. You cannot manufacture it. You can ask the clients you served well to describe, honestly and in their own words, what you actually did.

What to do in the next 30 days

Week one is measurement. Ten client questions, three engines, a fresh chat for each, every answer pasted into one document with the date. Then sharpen the site: cut the practice-area list back to what you genuinely do, put the accreditation where a skim-reader hits it first, and replace abstract area-of-law headings with the situations clients actually arrive with.

Next, reconcile everything published about you elsewhere — Google Business Profile, directories, the register — against that one reference record. In parallel, start a small answers section: real client questions, one clear answer each. Three to five genuinely useful pieces beat twenty thin ones, and Google's own guidance is to write them for readers rather than producing a parallel set of pages aimed at AI, which it treats as scaled content abuse.

Then keep watching. Re-run the same ten questions every few weeks and read the drift: named or not, named for what, named beside whom. Nothing here is stable enough to set and forget — ChatGPT's share of citations to one large source reportedly swung by tens of percentage points within a month after a platform-side change. The firms that start measuring now are building a record their competitors do not have. The first conversation is already happening; the only open question is whether your name is in it.

Confidentiality and advertising rules: the hard limits

Before you make the firm more visible, be clear about what is being made visible: you, not your clients. Everything you publish — answers, worked examples, case notes — has to be anonymised so that no reader can work backwards to a real person. The duty of confidentiality does not soften because the audience is a language model, and published pages get indexed, quoted and recombined in ways you will never see. "Unfair dismissal after a site closure" is a safe example. Names, file numbers, dates that pin a matter down, or details a local reader would recognise are not.

The advertising rules land in the same place. Conduct rules permit factual information about your practice and prohibit claims that create unjustified expectations about results. "Guaranteed win" and "best lawyer in the city" are exposed under those rules and useless anyway: an unverifiable superlative gives a retrieval system nothing to attach to. Say what you do, who you do it for, and what accreditation backs it, then let the checkable facts carry the weight.

Four weeks to a profile a machine can read

Week one is an audit with no writing in it. Search your own firm name. Ask ChatGPT for a specialist in your field in your city and note whether you appear, where, and with what description. Then open every listing that mentions you — directories, the register, your Google profile, your own practice pages — and check firm name, practice areas, accreditation, address, hours and phone against one another. Contradictions found in that hour are the commonest reason a model has nothing confident to say about you.

Weeks two and three are writing. Five to eight real client questions per practice area, each with the answer you would actually give on the phone: plain, complete, specific about deadlines and the next step. Not a memo to another lawyer. Week four assembles it — a structured practice-area page, the answers section, one consistent firm profile, a visible review date on anything time-sensitive. Then re-run week one's queries and compare them line by line against what you saved.

Questions firms ask us about AI visibility

"Do I have to buy ads for this?" No, and you could not if you wanted to: nobody sells a slot among the sources an answer cites. The spend that matters is attention — time on the answers themselves, and time making every listing agree. What the citation data correlates with is mentions across the web, not media budget.

"Is this a fad?" Look at the volumes and at what people do afterwards. ChatGPT reported 900 million weekly users in February 2026, Gemini passed a billion monthly users in August 2026, and Google puts AI Overviews above two billion people a month. Over the same stretch, SparkToro and Similarweb found 68 percent of US Google searches ending with no click to any website, up from 60 percent in 2024. The answer is increasingly the destination. It is also the natural first stop for the questions people are least comfortable asking a stranger — a divorce, a dismissal, a contested will.

"How do I measure it?" With a fixed question set and a calendar reminder: same prompts, same engines, every few weeks, saved with dates so you can see movement instead of remembering it. Your SEO reporting will not cover this — Ahrefs found only about 6 to 8 percent of ChatGPT-cited URLs also sit in Google's top ten for the same query. And keep asking on the phone: "How did you find us?" When "ChatGPT told me to call you" starts showing up in that answer, the work is landing.