What Does AI Say About You? A Three-Question Method for Mortgage Professionals

Most mortgage professionals have never checked what AI systems say about them. Not because they don't care, but because nobody has said what checking would even look like. There is no ranking report, no impression count, no dashboard. The recommendation happens inside a private conversation between a borrower and a machine, and by the time it matters, it has already happened.

That does not make it unmeasurable. It makes it measurable in a different way — by asking, systematically, and writing down what comes back. What follows is a method anyone can run. And to show why one measurement is never enough, it ends with a real five-day record: the same three questions, run twice, and what moved between them.

Set the conditions first

Three details decide whether your measurement means anything.

Log out, or use a private window. If you are signed in, the system may be drawing on your own history and stored memory. You will see yourself reflected back and mistake it for discovery.

Use a fresh conversation for every question. Once you have mentioned your own name or company in a thread, everything after it is contaminated. Ask about your firm and then ask "who provides this service" in the same window, and you may well be named — not because you were found, but because you are five lines up.

Ask for sources, every time. Add "cite your sources" to the question. Without it, a model can produce something plausible from your domain name alone, and you will have measured nothing. With it, you find out whether anything of yours was actually retrieved.

The three questions

Ask each one in a clean window, in every system you test — ChatGPT, Gemini and Perplexity at minimum. Keep the wording identical between systems and between rounds. The whole value of this exercise is comparability over time.

1. What is [your company or your name, plus your market]? Cite your sources.

This measures recognition. You are finding out whether anything about you has been retrieved at all, and — more importantly — what was retrieved. Pay attention to which sources appear. If everything traces back to material you published yourself, that is a specific and fixable finding.

2. How can a loan officer become visible when borrowers ask AI for recommendations?

This measures the conversation happening around your subject. You are not looking for your own name here. You are looking at who is being cited as an authority on the question — because those are the sources shaping what borrowers and colleagues are told.

3. Who provides [your service] for [your market]?

This is the one that matters commercially. It is the vendor question, phrased the way someone would ask it who has a problem and no preferred supplier yet. Note every name that appears, and note whether you are among them.

What a single round reveals

Run these across three systems on the same day and the results are not variations on a theme. They are different answers.

On the recognition question, one system may report that your website does not appear in its results at all, and build its answer entirely from press coverage. A second may describe you accurately but attribute everything to "recent press releases" without a single link. A third may cite your site directly, with clickable sources. Same subject, same day, three materially different pictures — which is itself the first useful finding. There is no single "what AI thinks of you." There are as many answers as there are systems, and they do not agree.

The vendor question tends to produce the sharpest result. In one measurement across three systems, twenty-one provider names appeared. Exactly one name appeared twice. Nothing appeared in all three.

That inverts the usual assumption. The instinct on seeing a list of competitors is that the space is taken. The pattern says otherwise: each system was assembling a list from whatever it could find, which is what happens when no source has become the reference. Tracing the citations, one system's entire answer led back to a single page — a comparison article published by one of the companies on the list, reviewing tools in its own category. That company was not named because it dominates the market. It was named because it wrote the article the others appear in.

Why one round is not enough

A single measurement is an anecdote. The reason to run the same questions again, on a fixed schedule, is that the answers move — and the movement is the only thing that tells you whether anything you're doing works.

Here is a concrete example, tracked over five days. On day one, a company running this method on itself found that two of the three systems could not read its website at all — one said the site was not indexed, another knew the company only through press releases and cited no links. On the vendor question, none of the three systems named it.

Five days later, after a handful of technical corrections, the same three questions returned a different picture. All three systems now read the site directly. Two cited individual article pages by name. And on the vendor question — the commercial one — the company now appeared by name in one system's answer where days earlier it had appeared in none.

Two things are worth saying plainly about that example, because they are the honest part. First, the systems moved unevenly: one still had not named the company on the vendor question at all, even after it appeared in the others. You are not "visible" or "invisible" — you are visible in some systems and not others, and only measuring each one tells you which. Second, across both rounds and all three systems, one finding never changed: there was still no independent, third-party coverage. Every system noted, in its own words, that it had only the company's own material to work from. That is not a gap a technical fix closes. It is the one that a single piece of outside validation would.

What to write down

For each question, in each system, note the date, the exact wording you used, the sources cited, every provider or authority named, and whether you appeared. Keep the full answer text, not a summary — the phrasing changes over time in ways a summary will hide. Then repeat on a fixed day each month. Not weekly: the answers vary enough between sessions that short intervals produce noise rather than signal.

Within a quarter you will have something almost nobody in this industry currently holds — a dated record of how AI describes you, and whether that is moving.

A note on what this does and does not tell you

These answers are not stable. Ask the same question twice in one afternoon and the wording will differ, and sometimes the names will too. Location, account history and model updates all move the result. Treat any single answer as one observation, not a verdict — the value is in the pattern across repeated, dated rounds.

What it does tell you is directional, and that is enough to act on. Whether anything of yours is being retrieved. Whether the only sources describing you are sources you published. Whether the question your clients would ask returns your name, someone else's, or a list assembled from nothing in particular. Those are answerable questions. They just have to be asked on purpose.

This article was compiled by autonomousgrowth.io. The method above is offered as-is, for anyone who wants to run it.

Method notes: measurements referenced here were taken in July 2026 across ChatGPT, Gemini and Perplexity, logged out, with a fresh session per question and sources requested in each prompt. Provider counts reflect specific dated runs; AI-generated answers vary by session, location and model version, so readers should run their own measurements rather than rely on figures reported here.