An AI visibility gap is the difference between what an AI assistant says when someone asks about your field or your firm, and what is true about you. You find it by asking the assistants the questions a prospective client would ask, saving the answers with the date, and checking each answer against four questions. It takes about an hour and needs no software.
Most guides to this topic are published by companies that sell tracking software, and their method starts with a subscription. This one is written for the owner of a law firm, an advisory practice or an accounting firm who wants to know where they stand before spending anything. It includes what we found when we ran the test on our own studio on 1 October 2026. The result was poor, and it is reported in full below.
What the phrase usually means, and what that leaves out.
The usual definition is competitive. A gap is a question where the AI names your competitors and leaves you out. That gap is real and the test below looks for it. It is also the easiest one for software to count, so the dashboards are built around it.
For a licensed professional, a mention count misses most of the problem. In May 2026 we audited 130 independent investment advisory firms. Perplexity surfaced 89.2 percent of them, and it did so by citing a third-party directory in place of the firm's own website. Only eight of the 130 firms, 6.2 percent, had their own site cited as a primary or significant source. A mention count would have scored nearly the whole sample as visible. The full figures and the method are in The AI Visibility Gap, our 130-firm audit.
That audit covered investment advisers. We have not published equivalent figures for law firms or accounting firms, so read it as one profession's result. The mechanism it shows applies to any practice that sits in a public directory: being named and being described accurately from your own material are separate things, and only the first shows up in a count.
The four gaps.
Being named is one question out of four. Each of the others finds a gap that a mention count scores as a success.
| Gap | What the answer looks like | The question that finds it |
|---|---|---|
| Absent | Other firms are named. Yours is missing. | Is my firm named? |
| Borrowed | You are named, and every source behind the answer is a directory, a review site or a listing you do not control. | Is my own website among the sources? |
| Wrong | You are named, and something said about you is false, out of date, or belongs to someone else with a similar name. | Is every fact about me correct? |
| Unverifiable | You are named and described correctly, and nothing in the answer or on your site ties the description to your license or registration. | Could a stranger confirm this from the regulator's record? |
One answer can show more than one gap. The fourth is hard to see from the answer alone, so the test gives it a separate step.
The test, step by step.
1. Write twelve questions.
Four by name: your firm, yourself with your title, the firm with its city, and "what does [firm] do." Four by category and place, the way a stranger would ask: "estate planning attorney in Tacoma," "fee-only adviser for tech employees in Austin." Four by problem, with no professional label in them: "my business partner wants to buy me out, who should I talk to in Fort Worth." Write them in a client's words. If your intake notes or first emails show how clients described their problem before they met you, copy those.
2. Ask each one on three assistants.
Use ChatGPT, Perplexity, and Google, where the AI answer appears above the search results. Start a new conversation for every question so an earlier answer does not carry into the next one. Note whether you were signed in. If you know your clients use a different assistant, add it.
3. Save every answer.
Copy the full answer and its list of sources into one document, with the date, the assistant and the exact question. Screenshots work as well. Twelve questions on three assistants is thirty-six answers. An undated answer is useless next quarter, because you will have nothing to compare it with.
4. Mark each answer four times.
Named, yes or no. Own website among the sources, yes or no. Every stated fact correct, yes or no. Tied to the public record, yes or no. For the third mark, read the answer the way a prospective client who knows nothing about you would. Check the address, the practice areas, the people named, any fees quoted, and whether the answer is describing your firm at all.
5. Check the public record yourself.
Look up your own entry: your state bar's directory if you are an attorney, adviserinfo.sec.gov if you are an investment adviser, your state board of accountancy if you are a CPA. Then check whether your website links to that entry anywhere, either in visible text or in its structured data. The Schema Inspector on our walkthrough page reads the structured data on any site and reports what is missing. It is free, and it does not tell you how to fix anything. In the adviser audit, none of the 130 firms linked its regulatory record in its structured data.
6. Run it again a week later.
Answers change between runs. In our Northwest Lagotto case study, the order in which the engines listed the breeder was not stable from one run to the next. A single run is a sample. Two runs that agree are the start of a baseline.
What we found when we ran it on ourselves.
On 1 October 2026 we ran a shorter version on clarion.studio: eight questions, on ChatGPT and Perplexity, one run each. Two of the questions used our name. Six did not.
On the six questions without our name, neither assistant named or cited Clarion. Zero of twelve answers. Questions about website design for advisory firms went to other agencies and website builders. On the question of how to get a law firm's site into ChatGPT answers, ChatGPT cited OpenAI's and Google's own documentation, and Perplexity cited legal marketing and software companies. That is the absent gap. For a studio founded this year it is the result we expected.
On our name, ChatGPT was accurate. It gave the right town, the right founder and the right focus, and it cited only our own pages. It also noted that our research and our case study are self-reported, which is fair.
Perplexity got it wrong. Asked for "Clarion Studio Lynden," it answered mostly about a rental studio listed on Peerspace and mentioned us third. Asked what Clarion Studio is, with our web address in the question, it cited our site next to a LinkedIn page that belongs to a different company with the same name, and described us as a branding and content strategy studio. That is the wrong gap. A mention count would have scored both answers as wins.
We build this kind of infrastructure, and our own result is zero of twelve unbranded answers plus one case of mistaken identity. We are publishing it for two reasons. A test that only gets reported when it flatters is worth nothing. And the result shows why the gaps have to be told apart. The absent gap may take months of published work to move, and whether it moves is up to the engines. The wrong gap has a likely cause we can work on now: another company shares our name.
These are single runs on one day. We plan to repeat them monthly, and we expect the numbers to move in both directions.
How to read your own results.
Start with wrong. A false statement about your fees, your practice areas or your identity reaches a person who asked about you by name. Open the sources the assistant listed and find where the error comes from. Places to look first: a stale directory profile, a biography on a former firm's site, a namesake. Each of those can be corrected at the source.
Then borrowed. If the assistants describe you only through directories, the description is whatever the directory holds, and you share its wording with every other listing. In the adviser audit this was the normal condition.
Then unverifiable. This is work on your own site: stating who you are in a form a machine can read, and linking that statement to the regulator's record. How AI Verifies You're Real explains what that involves, layer by layer.
Absent comes last, which reverses the order the software uses. Expect to be absent on broad category questions. We were. Look at the narrow questions, the ones that describe the exact problem you handle in the place you handle it. If you are absent there, check whether your site has a page that answers that question in plain words. If it does not, that page is the first thing to write.
What this test cannot tell you.
It cannot tell you why. You see what the assistant said and which sources it listed. The sources are a clue, and they stop short of an explanation.
It cannot tell you whether anyone acted on the answer. For that you need your site analytics and your intake records. One detail from our own run helps here: every link ChatGPT cited carried the tag utm_source=chatgpt.com, so visits that arrive from its answers can be counted in ordinary analytics.
It cannot be turned into a promise. The engines decide what to cite and they change without notice. Anyone who guarantees you a citation is selling something they do not control. What you do control is whether the facts about you are published, correct and checkable.
It also does not replace software in every case. A firm that needs to track hundreds of questions across many markets every week is better served by a tool than by a person with a document. A single practice with twelve questions can start without one. For the wider picture of what the evidence does and does not support in this field, see AEO vs GEO vs SEO.
Questions people ask.
What is an AI visibility gap?
An AI visibility gap is the difference between what an AI assistant says when someone asks about your field or your firm, and what is true about you. It comes in four kinds: absent (others are named and you are missing), borrowed (you are named only through directories you do not control), wrong (something said about you is false or belongs to someone else), and unverifiable (nothing ties the description to your license or registration).
How do I check whether ChatGPT mentions my firm?
Open a new conversation and ask the question a prospective client would ask, first by your firm's name and then by category and place without your name. Save the full answer and its sources with the date. Repeat on Perplexity and on Google's AI answer, and run it again a week later, because answers change between runs.
Do I need a paid tool to find AI visibility gaps?
No. Twelve questions asked by hand on three assistants will show a single practice where it stands in about an hour. Tracking software earns its cost when you need hundreds of questions checked every week across many markets.
How often should I repeat the test?
Quarterly is enough for most practices, and monthly makes sense while you are changing your site. Use the same questions with the same wording each time so the answers can be compared. Clarion plans to run its own set monthly.
Can ChatGPT run the gap analysis for me?
It can help you draft the twelve questions and sort the answers you have saved. It cannot reliably tell you what it would say to a stranger. For that, ask in a fresh conversation and read the answer yourself.
If I rank well in Google, am I visible in AI answers?
Ranking does not settle it. Ranking and citation are scored differently, and a firm can have one without the other. The evidence is set out in AEO vs GEO vs SEO. The only way to know is to ask the assistants and read what comes back.
Can an AI visibility gap be closed?
Some of it is in your hands. You can correct a wrong directory profile, publish pages that answer the narrow questions you handle, and link your site to your regulatory record. Whether an assistant then names or cites you is the engine's decision, and nobody can guarantee it.