What AI Says About Your Firm: A 20-Minute Check

Somewhere today, a client you'd like to have is describing a problem to an AI assistant and getting back a short list of names. You can see what that list says about your firm in about twenty minutes, spread over two days. What you learn depends entirely on how you ask. One generic question, asked once, tells you very little. Five questions in your best client's own words, asked of two engines on two different days, tell you something you can act on.

How do I check if ChatGPT recommends my firm?

Ask it the way your best client would, without your firm's name, in a fresh private chat, and save the answer word for word with the date. Then ask a second engine the same questions, and ask both again the next day. One answer is a reading. The same answer on two engines and two days is a finding.

The check, step by step

• First, write five questions from one ideal client's life. Pick the client you'd most like more of, then five moments when that client has a problem and goes looking for help. Write each one in the client's words, with the situation, the kind of help and the city.
• Second, check each question against your own site: point to the line where you say you do this work. If you can't, the question tests a claim you haven't made, and an absence there tells you nothing.
• Third, keep your firm's name out of every question, and the names of your people too. A question with your name in it is a branded question: it hands the engine the answer, and it flatters you.
• Fourth, give every question its own fresh chat, on two engines, signed out or in a private window. A private window keeps your history out of the answer but not your location, so if your market is somewhere other than where you sit, use a VPN in that market or name the city in the question.
• Fifth, save every answer word for word, with the engine and the date. Don't tidy it, summarize it, or keep only the part about you.
• Sixth, score each answer one of three ways: named (the answer says your firm's name), cited only (your site is among its sources, but the answer never names you), or absent. Where you're named, note your place in the list and how long the list is.
• Seventh, ask the same five questions again the next day, in the same words, on the same two engines, and score them the same way.
• Eighth, believe an absence only when it repeats across three phrasings over two runs: missing from at least three of your questions, on both days. An answer that errored, or one where the engine says it couldn't search, is unknown, never absent.
• Ninth, read who was named, then open their pages and note what they do that yours don't.

Once the five questions are written, the asking and saving come to about ten minutes a day: ten chats, each answer pasted into one document. Writing the questions is the part worth doing slowly, and you only do it once. Next month you ask the same five in the same words, so a change in the answer can't be blamed on the asking.

Never trust a single AI answer about your firm: ask the way your best client would, ask again on another day, and believe an absence only when it repeats.

For a law firm

Ask the way a general counsel or a business owner would describe the problem, with the specialty and the city: a supplier contract headed for arbitration, a commercial litigation firm, Denver. Never put the firm's own name in the question. Score each practice separately, because a firm can be named for one kind of work and absent from another, and a firm weighted toward disputes can be legitimately absent from a transaction question.

For a medical practice

Ask as a patient would, with the specialty and the city: a child's ear infections that keep coming back, a pediatric ENT, Boulder. Then note who the engine names: the physician, the practice or the hospital. Those are three different results, so score them apart. A physician named without the practice, or a hospital named in place of both, isn't the same finding as the practice named.

For a healthcare technology company

Ask as a hospital buyer would describe the problem, in the hospital's own terms: a reporting backlog, a system conversion, a billing process patients complain about. Then note whether the engine names your company or only the category's best-known name. An answer that offers only the biggest name in the category is a different finding from one that names five of your peers and leaves you out, so keep the two apart.

The instrument behind the check

This is the plain-language version of the instrument behind our Mirror: each question written from a line on the firm's own site, and no absence believed until it repeats across three phrasings over two runs. For the firms whose answers we keep on record, the questions stay the same, word for word, from one month to the next. If you'd rather have it run for you, request your own Mirror: the exact answers, who gets named next to you and where the gaps are, prepared by hand and yours to keep.

Questions about running the check

Which engines should I ask?

At least two, because one engine's answer tells you about that engine. ChatGPT is the natural place to start, and a second that shows its sources, such as Perplexity, lets you see which pages an answer drew on. If you know which assistant your clients use, make sure it's one of the two.

Should I ask with my firm's name in the question?

Not in the five. A question with your name in it hands the engine the answer, so it can tell you how you're described but never whether you're found. If you want to know what an engine says about you by name, ask that once per engine and file it apart from the five.

Why did the answer change overnight?

Because an answer is written fresh each time it's asked, and an engine that searches the web can find different pages tomorrow than it found today. That's normal, and it's why the check runs on two days: what holds across both is worth believing, and what appears once is worth watching.

What if a competitor is named every time?

Then you've found the most useful line in your record. No check can tell you why an engine chose them. What you can see is what their pages do that yours don't: a page for each kind of work, named people with real bios, the city stated plainly, structured data in the page source (search it for ld+json). Start there, and say your own true things as plainly as they say theirs.

How often should I run it?

Monthly, with the same five questions in the same words on the same two engines. A record kept that way shows a trend, and a trend tells you far more than any single answer.