Guide
How AI assistants recommend law firms — and why they recommend someone else
When a prospective client asks ChatGPT, Claude, Gemini or Perplexity which lawyer to hire, the assistant does not consult a fixed list of firms. It searches, reads, and writes an answer that usually names a few firms and often recommends one. This guide explains what is known about that process, what can be observed when the question is asked many times, what might explain why another firm is recommended, and how to test it.
Documented
What happens when a client asks an assistant for a lawyer
An assistant that can search the web usually decides, question by question, whether to. For a local and current question — who handles slip-and-fall cases near me — searching is how it can find firms that exist today. Each company documents the crawler that feeds its search:
- OpenAI uses OAI-SearchBot to surface websites in ChatGPT’s search features (OpenAI, read September 3, 2026).
- Anthropic documents separate crawlers for training, for fetching a page a user asked about, and for search indexing (Anthropic, read September 3, 2026).
- PerplexityBot surfaces and links websites in Perplexity’s results (Perplexity, read September 3, 2026).
- Microsoft’s Copilot sends a query to Bing for more information (Microsoft Learn, read September 3, 2026).
- Google states there are no additional requirements to appear in AI Overviews or AI Mode beyond being eligible for Search (Google Search Central, read September 3, 2026).
What follows from that is plain: a page those crawlers cannot reach is a page the assistant’s search cannot read. What no company documents is how its model weighs the firms it finds when it writes the answer. That part has to be observed.
Observed
What a measurement records when the question is asked many times
Ask one assistant one question and you get an anecdote. Ask a fixed set of client situations across several AI models and every answer ends in one of three outcomes: your firm was recommended first, another firm was, or no single firm was. Along the way, each firm an answer mentions reaches one of four stages:
- Surfaced.
- The answer named the business at all.
- Shortlisted.
- The answer weighed the business as one of the options.
- Recommended.
- The answer put the business in front of the client as something to do.
- Recommended first.
- The answer recommended the business first — its one named pick.
That separates two problems that call for different work. In the fictional example report, the assistants named the firm in 13 of 30 completed answers and recommended it first in 2 of those 13. Of the 28 answers that did not recommend it first, 17 never named it at all and 11 named it and recommended someone else, or no one: that firm's first problem is being found. The two call for different work.
Hypotheses
Why an assistant might recommend another firm
These are the explanations AgentSignal’s own diagnosis works through. Each is a hypothesis: something the recorded answers can point to, never something one scan proves. The assistant’s stated reason — “my recommendation is X because …” — is evidence of what it said, not proof of what caused it.
| Possible reason | What would point to it in the answers | What to change, then re-test |
|---|---|---|
| AI is not finding you | Your firm is missing from most answers for a kind of matter you handle. | Give each kind of matter you actually handle its own crawlable page, and make sure search and AI crawlers may read it. |
| The facts that win the decision are not retrievable from your site | The answer says it could not confirm what the client asked about — fees, a language, evening appointments, who handles the case. | State those facts in plain text on the page they belong to. |
| You read as a generalist | The answer describes the firm it recommended as focused on exactly this kind of matter. | Show the matter-specific experience you can support, on the page for that matter. |
| You are not tied to the places clients ask about | The answer prefers firms it places near the client, and does not place you there. | State your offices and the areas you serve plainly — without near-identical city pages. |
| Nothing independent vouches for you | The answer cites reviews, directories or news for the firm it recommended, and none for yours. | Keep accurate profiles on the directories and review platforms your clients use. Never buy or fabricate reviews. |
| The people who would do the work cannot be checked | The answer names an attorney and their background for the firm it recommended. | Publish attorney pages with admissions and background that can be checked. |
| AI cannot tell which business you are | The answer confuses your firm with another, or gives the wrong address or phone number. | Use one name, address and phone number everywhere your firm is listed. |
Any result you publish — outcomes, awards, testimonials — has to meet your jurisdiction’s advertising rules, which is why no change above asks a firm to claim anything it cannot support (North Carolina State Bar (adopting ABA Model Rule 7.1 and comment), read September 3, 2026).
To test — or to skip
What not to spend money on
Some techniques are sold as ways to be recommended by AI with no evidence behind them: an llms.txt file, special “AI schema”, hundreds of near-identical city pages, bought reviews. Google, for one, documents that no special markup or file is needed for its AI features (Google Search Central, read September 3, 2026). The methodology lists every technique we refuse to recommend, with the reasoning and the dated sources.
To test
How to find out which reason applies to your firm
By hand. Ask several assistants the questions your clients actually ask, in their words, in a fresh session, and note who is named and who is recommended first. It costs nothing and it is worth doing. Its limits: one answer is not a pattern, a signed-in app can personalize what it says, and the answers change from day to day.
As a measurement. Put a fixed, versioned set of client situations to several AI models, record every answer with its sources, and count — found, and recommended first — with the denominators printed. Change one thing you can verify. Then ask the same questions again and read the difference as an observation, not proof. That is what an AI visibility audit from AgentSignal does; the models are called through their APIs with web search available, which is close to, but not the same as, what a person sees in the app.