PricingContactGet Started Free

Blog / AI Search

How Do You Check Whether ChatGPT Recommends Your Law Firm?

You can test whether AI tools name your firm with a repeatable prompt panel, not a single query. This post walks through building the panel, logging results, handling answer variability, and knowing when manual checking stops being enough.

GrowthAgents AINovember 17, 202610 min read

Check if ChatGPT recommends your law firm: one lit brass nameplate on a lobby directory

The short answer: you check whether ChatGPT recommends your firm by running the same fixed set of prompts, several times each, and logging what comes back: whether your firm is named, which URL is cited, which competitors appear, which engine you used and the date. One screenshot proves nothing. A repeated, logged panel gives you a baseline you can compare month to month.

This is the core of Law Firm AI Search measurement, and you can start it today with a spreadsheet. This post covers how to build the panel, what to record, how to read the results, and where the method stops being useful. It is marketing information, not legal advice.

How do you test whether ChatGPT recommends your firm?

Treat it like a small, repeatable experiment rather than a curiosity search. Searching your own firm name once and seeing a good answer tells you very little, because partners tend to test the one prompt most likely to flatter them.

The method has four steps:

  1. Write a fixed prompt panel. Aim for 15 to 25 prompts that mirror how a real prospect would ask, covering the prompt types in the next section. Do not edit them between rounds. Changing the wording changes the experiment.
  2. Run each prompt more than once. Three to five runs per prompt is a practical minimum, in a fresh conversation each time.
  3. Log every run. Use one row per run, with the fields described below.
  4. Repeat on a schedule. Monthly is enough for most firms. The value is in the trend, not any single round.

If you want a quick read before building anything, a free readiness check of your site is available, and the manual panel below is the measurement layer that sits on top of it.

Which prompts should you run?

Real prospects do not ask “who is the best law firm.” They ask about a problem, a place and a decision. Build your panel across four prompt types so you see different parts of your visibility.

Prompt type Example pattern What it tells you
Practice area + city “Who are good [practice area] lawyers in [city]?” Whether you are in the consideration set for a high-intent search
Comparison “[Your firm] vs [competitor] for [practice area]” How you are described next to a named rival
Question “What should I do after [situation] in [state]?” Whether your content is cited as a source for informational answers
Brand “What is [your firm] known for?” or “Is [your firm] reputable?” Whether the description of you is accurate and current

Include prompts for every practice area you actively want to grow, not only the one you are famous for. Vary city and neighborhood wording, since prospects do. Keep the prompts neutral. Do not put your firm name into the non-brand prompts, or you will inflate your own results.

One drafting rule matters for legal marketing: write prompts as a prospect would, not as a lawyer would. “My landlord won’t return my deposit, who can help in Austin” behaves differently from “security deposit litigation attorney Austin.”

Why do AI answers change from run to run?

Because generative models are built to vary. OpenAI’s own help documentation notes that, with randomness enabled, “seeing different completions is expected” and that even tiny differences in a prompt, such as an extra space or newline, can change the output. That page is written for developers using the API, but the behavior it describes is why a consumer chat answer can differ between two identical-looking questions.

For ChatGPT’s search mode there is a second source of variation: the web itself. OpenAI says responses that use search contain inline citations, and what gets retrieved and cited depends on what is available at that moment.

The practical consequence is simple. A single run is an anecdote. If your firm appears in one of five runs of the same prompt, that is a very different finding from appearing in five of five, and you cannot tell the difference without repeating the prompt. This is also why we avoid declaring winners or rankings from small samples. We explain our general stance in What We Won’t Claim.

What should you log?

Keep the log boring and consistent. One row per run, these columns:

  • Date and run number. Lets you separate repeats from different months.
  • Engine and mode. ChatGPT with or without web search, or another assistant. Do not mix them in one column of results.
  • Prompt ID. A short code that maps back to your fixed panel.
  • Mention (yes/no). Was your firm named in the answer?
  • Cited URL. Which page of yours, if any, appeared as a source. Paste the exact URL.
  • Competitors named. Every other firm that appeared, in the order shown.
  • Description accuracy. A simple flag if the firm was described wrongly: wrong practice area, wrong city, outdated attorney, or a merged or closed office.
  • Notes. Anything odd, such as the answer declining to name firms at all.

Record whether the answer used web search. When a response is search-backed, ChatGPT lists its sources, and that source list is some of the most useful data you will collect, because it shows which third-party sites (directories, news, review platforms) the engine leans on for your market.

How do you read the results?

Separate three things that people often blur together.

Mention

The firm is named in the answer text. This is the broadest signal. A mention can be neutral, or even unflattering, so read the sentence around it.

Citation

One of your own pages is linked as a source. A citation means the engine used your content to support an answer. You can be cited without being named as a recommendation, and named without being cited, for example when the engine pulled your name from a directory.

Accurate description

The answer describes you correctly. A mention with the wrong practice area or a departed partner is a problem to fix, not a win to count.

From the log you can compute a simple rate: the share of runs in which your firm was mentioned, and the share in which your own URL was cited, compared with the competitors who appeared. That comparison is what we call citation share. Keep it descriptive. “We were cited in 4 of 20 runs for our panel in March” is a fact. “We rank third in ChatGPT” is not, because there is no fixed ranking to be third in.

For broader context on law firm AI visibility, see No Law Firm Cracks the Top 20 AI Visibility List. Our approach to measurement and its caveats is on the methodology page.

When does manual checking stop scaling?

A 20-prompt panel run four times across one engine is 80 runs per round. That is manageable once. It stops being manageable when:

  • You add a second and third engine, which multiplies the runs.
  • You have several practice areas or offices, each needing its own city-specific prompts.
  • You want to track competitors across rounds, not just yourself.
  • You need consistent logging by more than one person, and the data starts to drift because each person words and records things differently.
  • You want to tie changes to specific content or listing updates, which needs a clean history.

At that point the work is repetition and bookkeeping, which is what software handles well. GrowthAgents is a multi-agent AI marketing platform for law firms and professional services, and measuring citation share is one of the things it does. The method above is the same method, run at a larger scale.

What can this method not tell you?

Be clear about the limits before you show a partner a spreadsheet.

  • Personalization. ChatGPT can remember things across chats, and OpenAI describes memory as something you can switch off, with temporary chats that do not use memory. If you test from an account where you have asked a lot of law-firm questions, your results may not match a stranger’s. Use a clean or temporary session and note which one you used.
  • Location. OpenAI states that ChatGPT search collects general location information based on your IP address. A prospect in another part of your metro area, or on a phone, may see something different from your office desktop. Log where you ran the test.
  • No ranking guarantee. There is no stable ranked list of law firms inside ChatGPT. Appearing in your panel today does not mean appearing tomorrow, and no tactic can guarantee a mention or a citation.
  • No attribution. A mention does not tell you how many consultations it produced. Do not convert a mention rate into a revenue figure.
  • Engines change. Model updates and interface changes can move your numbers without anything changing on your side. Our post on ChatGPT, Claude and Gemini is a reminder that the landscape is not fixed to one engine.

The method gives you a defensible, repeatable view of whether your firm is showing up and how it is described. It does not give you an answer to “why,” and it does not give you a forecast.

FAQ

How many prompts should I run?

Start with 15 to 25 prompts across your core practice areas and cities, run three to five times each. That is a practical starting point rather than a statistical standard. A smaller panel gives noisier results, and more prompts help if you cover several practice areas.

Should I use a logged-out browser?

A clean session reduces the effect of your own history. OpenAI says temporary chats do not use memory, so that is one way to keep personalization out of the test. Whichever you use, use the same setup every round and write it in the log. Do not assume a logged-out or temporary session matches what every prospect sees, since location and other settings still apply.

Does appearing once mean my firm is visible?

No. Because answers vary between runs, one appearance may be luck. Look at the share of runs in which you appear across repeats and across prompts, and whether that share holds up in the next round.

Can I pay to be recommended by ChatGPT?

Not in the answer itself, according to OpenAI. When it began testing ads in the U.S. in February 2026, it said that “Ads do not influence the answers ChatGPT gives you” and that ads are labeled as sponsored and kept separate from organic responses. Paid ads are a different channel from being recommended, and legal advertising is subject to your jurisdiction’s rules, so check with your own ethics counsel before running any campaign.

Does GrowthAgents guarantee my firm will be cited?

No. Nobody can. We measure and report where you stand and help you improve the things within your control.

Run the check on your firm

If you want a starting point before building your own panel, request the Free AI Search Analysis. It is a readiness check of your site, and it will show you where to begin.

A city seen from high above

Can AI search read your site?

Free readiness check. About a minute.