How to Tell If Your Law Firm Is Invisible in ChatGPT, Gemini, and Perplexity

In an earlier post we looked at why AI platforms give different answers about the same law firm. The short version is that each assistant leans on a different mix of signals, so a firm can be recommended constantly on one and never mentioned on another. The practical question that follows is how to find out where your own firm stands.

You can get a useful first read in about twenty minutes with a spreadsheet and three browser tabs, as long as you run the check in a way that makes the results mean something.

Before you type anything

Open ChatGPT, Gemini, and Perplexity in a private or incognito window, and stay logged out where the platform allows it. If you use ChatGPT daily under your own account, it may already know what firm you run, and an answer shaped by that history is not what a stranger who was just rear-ended would see.

Set up a simple sheet before you start. One row per answer, with these columns:

• Date, platform, and whether you were logged in
• The exact prompt you typed
• Whether your firm was named (yes or no)
• Where it appeared in the answer, and how it was described
• Every other firm named
• The sources the answer cited, and whether your own website was among them or the citation went to a directory or bar association instead

It feels like overkill for a quick test, but without it you cannot compare this month to next month.

Which questions to ask

The most common mistake is typing your firm’s name. That tells you almost nothing, because a person who already knows your name is not the person you are trying to reach. The test that matters is unbranded discovery, where someone describes their situation and the assistant decides who to recommend.

Write your prompts the way an injured person would. Use your city, and use the practice areas that actually bring in your cases. These are the ten questions we use in The Answer Share, our study of which PI firms the assistants recommend across 15 markets:

1. best personal injury lawyer in {city}
2. best car accident lawyer in {city}
3. who should I call after a car accident in {city}
4. should I hire a lawyer after a car crash in {state}
5. how do I find a good personal injury attorney in {city}
6. top rated accident and injury law firms in {city}
7. best truck accident lawyer in {city}
8. lawyer to help with an insurance claim after an accident in {city}
9. I was injured in an accident in {city}, which law firm should I use
10. which law firm wins the biggest injury settlements in {city}

For a twenty-minute check, pick five of those and run each one on all three platforms. That gives you fifteen answers. If your firm does most of its work in a narrower lane, such as motorcycle crashes, premises liability, or wrongful death, swap in one or two prompts for that lane. Keep the wording identical across platforms so you are comparing the assistants and not your phrasing.

Once the discovery prompts are done, it is worth asking each platform one branded question, something like “what do you know about [your firm] in [your city]?” Keep that row separate, since it is not a visibility test. It shows you whether the assistant has a coherent picture of your firm at all, whether it has your current name and website, and whether it confuses you with another firm.

What to look for in each answer

Check that “named” really means you. If your firm name is built on a common surname, an answer that seems to mention you may be talking about a different firm with the same name in a different state. Confirm each hit against something only your firm would have, like your office address or your domain. We have seen automated tools get this wrong in both directions, counting a firm that was never named and missing a firm that was.

Check the competitors too. Assistants sometimes name a firm as local when it practices hundreds of miles away. Before you conclude that a rival is beating you in AI answers, confirm the rival is actually in your market.

Look at position, but do not lean on it. Being first in a list of five is better than being fifth, but a single answer’s ordering is a weak signal, so treat it as a secondary detail next to whether you were named.

Read the citations carefully. An answer can use your practice area page and still recommend other firms. On The Answer Share we describe watching an assistant read a firm’s own page, use it to answer the question, describe the author as “one local car-accident law firm,” and then send the reader to a bar association. Being the source and being named are different outcomes. If your site shows up in citations while your name stays out of the answer, write that down, because it points to a different fix than never being cited at all.

Note the wrong details. An old firm name or a retired domain usually signals an identity problem across the web rather than a content gap. In The Answer Share, one Kansas City firm that had changed its name was named once in 30 answers, and that one mention surfaced under the old domain. The new domain was named zero times. The assistants had kept the old identity and never connected it to the new one.

Why one run proves very little

Seeing your firm in one ChatGPT answer is a real result, but a thin one. The platforms disagree with each other, often sharply. In the Jacksonville market of our study, one firm held a 40% answer share across all three assistants while ChatGPT named it zero times. A partner who only checked ChatGPT would have concluded the firm was invisible.

The same platform also disagrees with itself. Ask the identical question tomorrow, or from a different city, and the list can change. The methodology note on our study says as much: AI answers vary with time, location, and phrasing.

The way you reach an assistant matters as well. The consumer ChatGPT app can return a different set of firms than the same model reached another way, because the app layers in things like local map results, live browsing, and personalization. We have run the same Orlando question through the consumer app and through a measurement tool on the same day and gotten substantially different rosters.

The systems also change without notice. We have seen the answers coming back for the same prompts shift between two runs a couple of weeks apart, with nothing announced, in a way that moved firms up and down regardless of anything they did. If this month’s check shows a big swing from last quarter’s, rule out a change on the assistant’s side before you credit or blame your own marketing.

So repeat the check. Same prompts, same platforms, same logged-out setup, once a month. A few months of rows in the same sheet tell you far more than one run.

Signs of a structural problem

Some patterns suggest a problem that more content will not solve.

You are absent everywhere, even on the broad prompts. Missing from one platform is normal. Missing from all three on every discovery question is worth investigating. In The Answer Share, ten household-name advertisers held a 0% answer share in their own markets, and 81% of all scored firms sat at or below a 10% answer share. Advertising reach does not carry over into AI answers on its own.

Your robots.txt blocks AI crawlers. Go to yourfirm.com/robots.txt in a browser. Look for Disallow rules under user agents such as GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot, or ClaudeBot, and for a Google-Extended rule, which controls whether Google may use your content in its AI answers. Rules like these can be added by a web developer or a security plugin, so do not assume nobody at your firm would have blocked them.

Your firewall blocks them, which robots.txt will not show you. A web application firewall can turn away an AI crawler even when robots.txt welcomes it. We have seen a firewall reject any visitor whose user agent contained the word “SearchBot.” That quietly blocked OAI-SearchBot, the crawler ChatGPT uses to fetch pages live, while GPTBot and every other AI crawler got through. The robots.txt file on that site was fully open. Catching it means requesting pages the way each crawler does and comparing the responses, which a normal browser cannot do. The firewall in that case was a WordPress security product of the kind many law firm sites run.

Your content only appears after JavaScript runs. Answer engines read your page source and do not run JavaScript. If your site is built so that the text loads in after the page does, a visitor sees a full page and the crawler sees an empty frame. A rough home check is to right-click a practice area page, choose “View page source,” and search for a sentence you can see on the screen. If it is not in the source, raise it with your web developer.

If you test crawler access with your own scripts, know that some firewalls block the testing tool itself, which looks exactly like a site blocking every AI bot. If an ordinary browser request from the same tool is refused too, suspect the test.

What a manual check can and cannot tell you

Done carefully, this check tells you whether the assistants name your firm for the questions your clients ask, which competitors they name instead, and whether your site is being used as a source.

It also has limits. Fifteen answers from one location on one day is a small sample. You cannot see why an assistant chose one firm over another. You cannot test crawler access the way the crawlers experience it from a browser window. And it is easy to count a false match, or to miss a firewall problem that makes everything else in your sheet irrelevant.

Our snapshot picks up where the manual check leaves off. It records the actual answers for unbranded discovery prompts, lists who the engines named instead of you, and scores the ten readiness signals behind those results, from your owned site and structured data through your Google Business Profile, reviews, directory consistency, and whether you are being cited in answer engines at all. Before any of that means anything, we confirm the crawlers can actually read your site.

If you have run the check above and want to know what is behind the results, see how we work with law firms and book a discovery call. Bring your spreadsheet. If you would rather see what ongoing measurement looks like first, you can test drive the dashboard our clients use.