Clarion Studio  ·  Primary Research  ·  2026

Who AI Names When Clients Look for a Law Firm

200 US firms of 29 to 170 lawyers. 360 questions a client might ask, each put to ChatGPT, Perplexity and Google five times: 5,400 answers. Directories write most of the shortlist. A firm’s own pages earn the citations when the client describes the problem.
Pre-registered at osf.io/3rjx2 before any data came in  ·  Collected 24 September to 7 October 2026  ·  Third wave running through 19 November 2026

Clarion Studio  ·  clarion.studio
54%
of the sources ChatGPT cited, when asked to name law firms, were legal directories. On Perplexity, 60%. In Google’s ordinary results, 13%.
1 in 15
answers named a given firm when the question matched that firm’s own state and practice area. Of 162 firms with a matching question, 62 were never named at all.
The shortlist is written by directories. The citations are earned by pages.

The Short Version

A business owner needs a lawyer. She does what more clients now do and asks an AI assistant: I own a family business and need to plan how it passes to my children. We are in California. Which law firms should I talk to? What comes back is not ten links. It is a short list of firms, with the sources the answer relied on.

We wanted to know who writes that list for firms of 29 to 170 lawyers, the size at which most firms win work by referral and have never measured what the machines say about them. So we built 360 questions across 120 pairs of state and practice area, registered the design publicly, and asked ChatGPT, Perplexity and Google each question five times. Then we counted.

When a client asks an AI tool to name firms, it mostly copies the directories. Legal directories were 54% of ChatGPT’s cited sources and 60% of Perplexity’s, against 13% of Google’s ordinary results. Three times out of four, when one of our 200 firms was named, the same answer was citing a directory page that lists it. Firms with more Best Law Firms rankings were named more often, after allowing for size, links and traffic.

Most firms this size rarely appear. When a question fit a firm’s own state and practice, the AI named that firm in about one answer in fifteen. Ten firms took 43% of every mention. Sixty-two of the 162 firms with a matching question were never named once.

The opening is in a different kind of question. When the client described the problem instead of asking for “the best firms in Texas”, 78% of what ChatGPT cited were law firm websites, and directories fell to 11% of its sources. That finding was found in the data rather than registered in advance, and measured on ChatGPT only. We report it as a lead, not a rule.

And one finding cuts against the work we sell. We scored every firm’s structured data before collecting a single answer. The score showed no detectable link to being named, by AI or by Google. If someone sells you markup as the answer to AI visibility, ask for their evidence. Ours does not support it.


54%
of ChatGPT’s cited sources were legal directories when asked to name firms
1 in 15
matching answers named the firm
78%
of ChatGPT’s sources were law firm websites when the client described the problem

Key Findings at a Glance

MetricFinding
Share of ChatGPT’s cited sources that were legal directories, on “name the firms” questions54%
Same, Perplexity60%
Same, Google’s ordinary results13%
Answers naming one of our firms that also cited a directory page listing it3 in 4
Change in odds of being named per one standard deviation more Best Law Firms rankings, size, links and traffic held constant1.7 to 3.4×
Answers naming a firm when the question matched its state and practiceabout 1 in 15
Firms with at least one matching question162 of 200
Of those, firms never named at all62
Share of all mentions taken by the ten most-named firms43%
ChatGPT sources that were law firm websites, problem-described questions78%
Same, “best firms in [state]” questions28%
Directories’ share of ChatGPT’s sources on problem-described questions11%
Detectable link between a firm’s structured-data score and being namednone
Answers collected5,400

What This Means in Plain Terms

Two questions, two different fixes

The study points to two places a firm can act, and they are not the same place.

“Who are the best firms for…” is decided by your directory profiles. These answers draw most heavily on Chambers, Best Law Firms, Best Lawyers and Super Lawyers. The first check costs an afternoon and no vendor: are your profiles present, current, and listing the practices clients actually ask about? It is the least expensive thing in this report to fix, which is why it comes first.

“I have this problem. Who can help?” can be decided by your own pages. Here the answers cite firm websites that explain the situation in plain terms. A firm with no page on the situation gives the answer nothing of its own to cite. That is the work: in each practice, the pages a client’s own problem leads to, each opening with a short answer a machine can quote, the authority behind it, and a lawyer’s name.

Who is asking

Firms of this size still win much of their work by referral, and nothing in this study measures how many of their clients ask an AI first. What it measures is what those clients get when they do. The people who hire outside counsel are adopting these tools quickly: among in-house lawyers surveyed in 30 countries, the share actively using generative AI went from 23% in 2024 to 52% in 2025, according to the Association of Corporate Counsel and Everlaw.

What it does not say

It does not say that AI answers decide who gets hired. It does not say that markup is useless. Most sites scored low, so the data say little about what a very high score would do. It does not promise that any page, however well built, will be cited. The engines decide. What it shows is where the decisions are being made today, and which of them a firm can reach.


Method

Sample. 200 US law firms of 29 to 170 lawyers. Each headcount was checked on the firm’s own website, not taken from a directory.

Questions. 360 questions a client might ask, across 120 pairs of state and practice area. Three forms: the request for a shortlist (“which firms should I talk to”), the superlative (“the best firms in [state] for…”), and the problem described in the client’s own terms with the state named.

Engines and runs. ChatGPT, Perplexity and Google, each question put five times, between 24 September and 7 October 2026: 5,400 answers. Every named firm and every cited source was recorded.

Registration. The design, the firm list, the questions and the tests were written down and registered at osf.io/3rjx2 before any answer was collected. The exploratory finding in section 3 is labelled as such because it was not among the registered tests.

Markup scoring. Every firm’s site was scored for structured data before collection, so that the score could not be shaped by the results.

What comes next. A third wave runs through 19 November 2026. These figures will be updated, with dates, once it is analysed. Firms in the study receive their own row on request.


Finding 1Ask an AI Tool to Name Firms, and It Mostly Copies the Directories

On the questions that ask for a shortlist, the sources behind the answers were not law firms. They were the directories that list law firms.

Share of cited sources that were legal directories
“Name the firms” questions, 24 September to 7 October 2026
ChatGPT
54%
Perplexity
60%
Google, ordinary results
13%
Three times out of four, when one of the 200 firms was named, the same answer cited a directory page that lists it.

Standing in the directories predicted being named. Firms with more Best Law Firms rankings were named more often, even after allowing for size, links and traffic: one standard deviation more rankings went with roughly 1.7 to 3.4 times the odds of being named. The website did not predict it. That is a submissions question, and it is one a firm’s own marketing team controls.

What this means. For the shortlist question, the directory is the source and the firm is a line in it. A firm that is absent, out of date, or listed under the wrong practices in Chambers, Best Law Firms, Best Lawyers and Super Lawyers is absent from the answer, whatever its website says.


Finding 2Most Firms This Size Rarely Appear

MeasureResult
Firms with at least one question matching their state and practice162 of 200
Answers naming the firm, on its matching questionsabout 1 in 15
Matching firms never named at all62
Share of all mentions taken by the ten most-named firms43%

The distribution is steep. A small group of firms is named again and again. Most are named rarely or never. We could not count every firm practising each area in each state, so these figures show how rarely firms appear, not how that compares with chance.

What this means. Being at zero is common at this size, and it is not a verdict on the firm. It is a description of which sources the answers are built from, and section 3 is about the one place that changes.


Finding 3When the Client Describes the Problem, the Answer Cites Law Firm Websites

Question form, ChatGPTLaw firm websites among cited sourcesDirectories among cited sources
“The best firms in [state] for…”28%majority
The client’s problem, described in plain terms78%11%

Most of the firms cited on the problem-described questions were outside our 200. They were firms, of any size, that had published a page answering that situation in plain terms. The answer had something of theirs to cite, so it cited it.

How to read it. This was found in the data, not registered as a test in advance, and it was measured on ChatGPT only. Treat it as a lead, not a proven rule. It is also the finding the rest of this report rests on, so we have been careful to label it.


Finding 4Technical Markup Showed No Detectable Link to Being Named

We sell structured-data work, and structured data is widely sold as the fix for AI visibility. So we scored every firm’s site for it before collecting a single answer, and then tested whether the score predicted being named. It did not, by AI or by Google, to any detectable degree.

Most sites scored low, so this says little about what a very high score would do. One firm of about 70 lawyers, assessed from public sources, scored 27 of 100 for markup, in the 87th percentile of the set. It was named in none of 420 AI answers about its own states and practices, and it had no Best Law Firms profile we could find.

What this means. Markup is plumbing. It lets a machine verify what a page says, and it does not give the page anything to say. The firms that were cited had pages that answered a question. If a vendor leads with markup, ask for the evidence that it moved an answer.


The Method on One Small Firm

Kraus Law is a two-lawyer Texas business-law practice. In September 2026 we rebuilt its site as a working legal reference: 318 glossary terms in plain English, each with its citation, a coded report on all 131 published opinions of the Texas Business Court, a case tracker kept current, and each judge’s procedures arranged around what a litigant is about to do. The full release went live on 28 September.

Weekly Google clicks, kraus.law
Ten weeks, 10 August to 18 October 2026. Weeks start Monday. Source: Google Search Console.
10 Aug
15
17 Aug
4
24 Aug
8
31 Aug
7
7 Sep
5
14 Sep
12
21 Sep
12
28 Sep
156
5 Oct
135
12 Oct
120
The new pages went live on 29 September. Before launch, weekly clicks ranged from 4 to 15.
Measure29 days before 29 September 202629 days from 29 September 2026
Google clicks38612
Appearances in Google results15,798103,969
ChatGPT runs in which the firm’s own report was the first source cited for “How often does the Texas Business Court grant motions to dismiss?”, tested 1 and 8 October2 of 3

Early results for one small firm. Clicks peaked in launch week and were 120 in the most recent week shown. Inquiries are not yet measured, and none of this proves cause. It is consistent with the exploratory finding in section 3, and the method is designed to extend practice group by practice group in a larger firm.


What a Firm of 30 to 170 Lawyers Can Do

This afternoon, without a vendor. Open Chambers, Best Law Firms, Best Lawyers and Super Lawyers and check every profile: present, current, and listed under the practices clients actually ask about. Finding 1 says this is where the shortlist questions are decided.

This quarter, with your lawyers. In each practice group, write down the ten questions clients bring to a first call. For each, ask whether a page on your site answers it in plain terms, with the authority cited and a lawyer’s name on it. Where the answer is no, that is the page to build. Finding 3 says this is where a firm can be the source the answer cites instead of a line in someone else’s list.

What to be sceptical of. Anyone selling markup, “AEO” or “GEO” as the fix on its own. Finding 4 is our own result against our own interest, and it stands until someone shows better evidence.

The firm owns the asset, and the lawyer owns the byline. Pages built this way live on the firm’s site, in the firm’s repository, with a named owner and a co-owner, so that when a lawyer leaves the history stays with the firm. That design answers the question every managing partner asks about this work, and it is set out on the law firms page, where firms in the study’s size band can request their own row at no charge.


Limits

Five runs per question per engine is a sample, and AI answers vary from run to run. The firm count per state and practice could not be established, so naming rates are not compared with chance. The exploratory finding in section 3 is from ChatGPT only. Markup scores were low across the set, so the null result in section 4 does not describe what a very high score would do. The single-firm case in section 5 is a disclosed family engagement and one firm. It illustrates the method and proves nothing about the population. No firm is named, and individual rows are shared only with the firm concerned.


Companion Reading

The Unverified Profession measured 162 firms of 40 to 250 lawyers in September 2026 against 125 buyer questions, and found that no firm linked a lawyer to a bar admission record in machine-readable form. Read with this study, it explains why the engines fall back on directories: there is little on firm sites they can verify.

The AI Visibility Gap applies a related instrument to 130 independent registered investment advisers. The pattern repeats across professions.

The proposal review is for a firm that has a website proposal in hand and wants to know, before signing, whether anything in it would give an answer engine a reason to cite the firm.


Disclosures

Clarion Studio sells website and search work to law firms, and so has an interest in the answer. The design was registered before collection so that the tests could not be chosen after the fact, and the finding against our interest is reported in full in section 4. The single-firm case is a family engagement, disclosed above and on every page that cites it.

The study was conducted by Clarion Studio using Claude, developed by Anthropic, as the analysis system. Claude was not one of the engines measured, so that no model evaluated outputs from its own class.

Firms in the study who want their own row, and firms of the same size who want the study’s questions run for their states and practices, can request it on the law firms page. Three firms a month, at no charge.


References

This study

Clarion Studio. (2026). Who AI names when clients look for a law firm: pre-registration. Open Science Framework. https://osf.io/3rjx2

Adoption

Association of Corporate Counsel and Everlaw. (2025). Generative AI use among in-house counsel: 23% in 2024 to 52% in 2025.

Companion research

Clarion Studio. (2026, September). The Unverified Profession. https://clarion.studio/research/the-unverified-profession-2026/

Clarion Studio. (2026, May). The AI Visibility Gap. https://clarion.studio/research/the-ai-visibility-gap-2026/

Who AI Names When Clients Look for a Law Firm  ·  Primary research by Clarion Studio  ·  October 2026  ·  200 firms, 360 questions, three engines, 5,400 answers

See your firm’s row.

Each month we run the study’s questions for three firms of 50 to 150 lawyers, for their own states and practices, at no charge. The written assessment is yours whether or not we ever work together.

Request your firm’s assessment