AI search for law firms: how to get recommended.

Ask an AI assistant which lawyer to hire and it answers with names. In our data it named about seven firms per answer, drawn from about seven websites it had just read, and three in four of those websites were not directories. For a law firm that is the whole opportunity in one line: the assistants are reading firms' own pages more than they are reading rankings. That is unusual. When we asked the same assistants about marketing agencies, ChatGPT cited Clutch in seven of ten answers. This guide is built on 40 recorded answers to ten questions a client would actually ask, in ten cities, and it puts the steps in the order the answers support.

What the assistants do with a law-firm question

In July 2026 we put ten questions to ChatGPT, Claude, Gemini and Perplexity, each with web search or grounding on, and recorded every firm named and every website read. The questions were the kind a client types, not the kind a marketer does: "Best employment lawyer in Dubai?", "Recommend a real estate law firm in Toronto", "Top criminal defense attorney in Delhi?" The full study lists all ten with every answer; this page is about what a firm should do with them.

Ten law-firm questions, four assistants, July 2026. Counted from recorded answers.

Three terms carry this guide. A citation is a website an assistant read and listed under its answer. A named firm is a firm the answer recommended by name, whether or not its own site was cited. A directoryis a ranking, review or listing site such as Chambers or Lawzana, as opposed to a firm's own page or an ordinary news or blog page.

Three things in that table decide the rest of this guide. First, the assistants are not shy. Thirty-nine of the forty answers named specific firms; the one exception was Claude declining to name a startup lawyer in Los Angeles. Second, the shortlist is wide. Across the forty answers, 237 different firms were named, 198 of them exactly once, and no firm was named by all four assistants for the same question. Third, the reading is wide too: the assistants cited 185 different websites, 140 of which were read by only one of the four, and just five were read by all four.

So there is no fixed list of winners to get onto. There is a set of sources each assistant reads for each market, and a shortlist it assembles from them on the day. Both halves of that are things a firm can act on.

Where the citations come from, by market

The table lists, for each question, every website that two or more of the four assistants cited. Firms' own sites are described rather than named here; the dataset names them.

Number in brackets is how many of the four assistants cited that site for that question.

In London and Singapore the assistants defer to the editorial rankings. Legal 500 was cited by all four assistants for the Singapore tax question and by three for London intellectual property; Chambers was cited by three for each. If your firm is not ranked in those guides for the practice area, the assistants have little else to read, and the answer is assembled from firms that are.

In India and the United States the assistants read firms' own pages.For Mumbai immigration, Bangalore family law, Los Angeles startup work and Delhi criminal defence, an individual firm's own site was among the most-cited websites, read by three of the four assistants. Directories still appear, Lawzana across the Indian questions, Best Law Firms and Expertise.com for New York and Toronto, Clutch for corporate and employment work, but they share the answer with practice pages rather than owning it.

That difference is why a generic "get listed everywhere" plan wastes money. A Singapore tax practice and a Bangalore family practice need different first steps, and the data says which.

Step 1: Confirm the assistants can read your site

By the end of this step you will know whether any of the six steps below can work at all.Every assistant in the table reads pages live, and a page it cannot fetch cannot be cited, however good it is. Law-firm sites are unusually likely to have this problem: security plugins and CDN bot rules that block "AI" user agents wholesale are common on WordPress installs, and nobody notices because the site still works in a browser.

1. Run the free crawler check on your domain. It reads your robots.txt and reports which of sixteen AI crawlers are allowed in. No sign-up.
2. Read the OpenAI line carefully. OpenAI runs three separate agents: GPTBot collects training data, OAI-SearchBot powers ChatGPT search citations, and ChatGPT-User fetches a page when a person asks about it. The one that governs whether ChatGPT can cite you is OAI-SearchBot, and it is not the one most robots.txt templates block (OpenAI, Overview of OpenAI crawlers).
3. Do not "fix" a Google-Extended finding by blocking Googlebot. Google-Extended controls Gemini training and grounding only; Google Search and AI Overviews follow Googlebot, and a site that blocks Googlebot disappears from Google entirely (Google, common crawlers).

Verification: the check shows OAI-SearchBot, PerplexityBot, Google-Extended and ClaudeBot allowed, and your practice-area pages return normally when fetched with those user agents.

Step 2: Ask the ten questions your clients ask, three times each

By the end of this step you will have a baseline: which assistants name you, for which questions, and who they name instead.Write the questions the way a client does, practice area plus city, and never your firm's name. "Best divorce lawyer in Bangalore?" tells you something; "Is Sharma and Associates a good firm?" only tells you whether the assistant has heard of you, which is not what a client asks.

1. Ten questions across the two or three practice areas that make money and the cities you actually serve. Include one comparison phrasing a client uses late in the decision, such as "Which firm should a startup use for its first funding round in Delhi?"
2. Put each question to ChatGPT, Gemini, Perplexity and Claude, and repeat each three times on different days. Answers change between runs; across all 50 questions in our wider study, all four assistants converged on the same cited source only three times.
3. Record two things per answer: the firms named, and the websites cited. The second list is the one you will act on.

Verification: a spreadsheet with 120 answers in it, and a count of how many named you. If the count is zero, the next three steps are the work; if it is not, they are how you make it stick.

Step 3: Make each practice-area page the quotable answer

By the end of this step, each practice-area page opens with the answer to the question a client asks about it.In four of the ten questions a single firm's own page was read by three of the four assistants, so this is where they are looking. What they need is a page they can read, verify and quote, which is a different standard from a page that impresses a visitor.

1. Put a heading that matches the client's question, and two plain sentences under it that answer it: what you do, for whom, in which courts or jurisdictions. A hero line and a feature grid give an assistant nothing to quote.
2. Name the lawyers who do the work, with the year they qualified and the matters they handle. Assistants extract facts they can check; "our experienced team" is not one.
3. State the practical facts a client needs and a directory usually carries: address, languages, how fees work, how to book. If those live only on Lawzana or Justdial, the assistant cites Lawzana or Justdial.
4. Add structured data the page can support: an Organization or LocalBusiness block for the firm, Person for each named lawyer, and FAQPage for the questions you answer on the page. Our JSON-LD guide has the markup written out, and the LocalBusiness exampleis the closest starting point for a firm with an office. Google's own guidance is that no special markup is required for its AI features; the point is legibility, not a ticket into the answer.

Verification: paste the page's first paragraph into the client's question and ask whether it answers it. If a stranger could not tell what the firm does from that paragraph, neither can an assistant.

Step 4: Get onto the sources the assistants read in your market

By the end of this step you are listed where the assistants already look, and you have stopped paying for listings they do not. The market table above is the shopping list; here is how to read it.

• London, Singapore, Dubai, and any market where clients hire internationally: Chambers and Legal 500. Both are editorial rankings with a submission process, and both were cited by three or four assistants in our data. Being ranked is slow and it is the only thing those answers were built from.
• United States and Canada:Best Law Firms and Best Lawyers, cited for New York, Toronto and Singapore, and Expertise.com for New York. Firms' own pages appeared alongside them, so this is in addition to Step 3, not instead of it.
• India:Lawzana was cited across the Mumbai and Bangalore questions, and Gemini cited LawRato three times and Perplexity cited Justdial twice across the set. Indian answers leaned on firms' own sites most of all, so the listing supports the page rather than replacing it.
• Corporate and employment work: Clutch appeared for Delhi corporate and Dubai employment questions, which is the same directory that dominated our marketing-agency answers. If your clients are companies, a Clutch profile is being read.
• Wikipedia:ChatGPT cited it in four of ten answers, and Wikipedia only carries pages for firms notable enough to have one. Do not create or edit a page about your own firm; that is a conflict of interest under Wikipedia's rules and it gets removed.

Verification: for each source in the table for your market, you can find your firm on it, and the listing carries the same facts as your practice page.

Step 5: Keep the facts and the dates current

By the end of this step, nothing on your site contradicts the assistants when they check. Because they read live, a partner who left in 2024 and is still on the team page, an office that moved, or a page last updated three years ago is visible to them at the moment of the answer. A dated competitor page is the safer quote. Update the content and the visible date together; changing a date alone on unchanged text is the kind of thing an assistant can compare against its last read.

Step 6: Measure per assistant, every week

By the end of this step you have a trend instead of an anecdote. The four assistants behave differently enough that one number hides the picture. Gemini names 8.1 firms per answer and ChatGPT 10.3, so there is room on their shortlists. Perplexity reads the most websites, 8.1 per answer, and names the fewest firms, 4.7, so being read by it is not the same as being named. A firm that appears in Gemini and not in Perplexity has a specific problem, and the weekly record is what shows it.

Repeat Step 2 every week, or let a workspace do it and keep the raw answers so any number traces back to what an assistant actually said. One check tells you where you stand; ten tell you whether the change worked.

What will not work

• Buying ChatGPT ads to get into the answer. OpenAI places ads below the answer and says they stay separate from it, and it does not show ads to Plus, Pro or Business users at all. What that means for a firm, quoted from OpenAI's own documentation.
• Submitting your site to an assistant. There is nothing to submit to. They read what they can reach.
• A "best law firm in" page for every city you do not serve. The assistants cited firms' own pages when those pages were about real work in a real place. Templated city pages are the pattern Google's spam policies name as scaled content.
• Self-written reviews or rating markup. A firm rating its own services in structured data is a manual-action risk in exactly this category, and it does not appear in anything the assistants cited.

Common questions

Sources

• What AI recommends: law firms, the study this guide is built on, with every question and answer, and the dataset as CSV, CC BY 4.0.
• Overview of OpenAI crawlers, OpenAI, for the three agents and what each does.
• Perplexity crawlers, Perplexity, for PerplexityBot and Perplexity-User.
• Google common crawlers, Google, for Google-Extended and what it does not control.
• AI features and your website, Google Search Central, for the statement that no special markup is required.

Firm names in this article are as the assistants wrote them in July 2026, counted from recorded answers. We have not verified the firms themselves. If something here does not match what you see, tell us and we will correct it.