A $4 Billion Firm, Measured in Every Market It Operates In

The firm

A wealth management and advisory firm managing approximately $4 billion across roughly 4,900 accounts. Fee-only, fiduciary, headquartered in a coastal metro, with offices ranging from one employee to six outside headquarters.

This research reached the firm through its agency rather than directly. No fee was paid; the firm gave permission to publish in exchange for the findings. What the agency wanted was an independent read it had not produced itself — something it could put in front of its own client that nobody could accuse either party of writing to flatter.

What was measured

2,300 questions across twelve markets, on five AI platforms, in a single collection window. Every claim scored against the firm's Form ADV as filed with the SEC — never against the firm's website, because the website was part of what was being measured.

Two layers. Discovery: does AI name the firm when a buyer describes a need without mentioning it? Accuracy: when AI does describe the firm, is what it says true against the filing?

Finding 1 — The location page is the asset, and one office didn't have one

Sorting every source AI read by what kind of page it was — rather than which site it sat on — showed content types are not interchangeable.

AI recommends firms from pages that answer who and where, and answers questions from pages that explain how. A city page asserts that a firm exists, in a place, serving people there — the three facts a recommendation requires. An article demonstrates expertise but asserts nothing a model can convert into a referral.

The firm's own citation record made it concrete. Across the discovery events where it was named, 94 of its cited pages were location pages. Five were service pages, four were articles, two were the homepage.

The inverse case was sharper. The firm's charitable-giving article was read sixteen times in answers where the firm was not recommended. AI used the firm's own explanation of the strategy to compose its answer — then recommended a community foundation. The expertise informed the response. The firm was not the response.

And the firm's largest office outside headquarters — six staff, a filed street address — had no location page at all. In queries about that city, AI cited a neighboring city's page instead, ten times.

Finding 2 — Two buyer questions route away from advisers entirely

Discovery visibility turned out not to be one number. It was six questions with rates from 20.7% to zero.

The divorce question produced zero recommendations in 186 answered events — for any wealth manager, not only this firm. Asked how to value a share of retirement accounts in a divorce, AI reads family-law firms and recommends attorneys. Legal pages account for 10.4% of what it read there and 0.5% of what it read when any adviser was recommended.

The charitable question behaves the same way at lower intensity — which is the mechanism behind that article being read sixteen times without producing a single recommendation. The firm was competing for a referral the question was never going to generate.

This finding later replicated. In a separate AIWorthy study of fifteen independent advisory firms in a different region, using a different panel and instrument, the same question produced zero recommendations in 375 responses. Two panels, two regions, two dates, same zero.

Finding 3 — There are two routes into an AI answer, and the firm used one

Examining every competitor recommendation alongside the sources cited with it revealed two distinct paths.

Two competitors were recommended in one market without their own websites being read at all. Directory listings carried them entirely.

Across all competitor recommendations, 28% arrived without the firm's own website being read. The equivalent figure for the firm was 22% — and those were platforms answering from memory in the home market, not from a directory.

The firm had one working channel where several competitors had two.

And it was already paying for the second. The filing disclosed participation in an adviser-matching platform in the same directory ecosystem AI reads from. The channel wasn't foreign territory — the participation simply wasn't converting into recommendations the way it did for two competitors.

Finding 4 — Only one platform was reading the firm's website

Of the 135 discovery events citing the firm's own domain, every single one was Perplexity.

Two platforms cited nothing in discovery at all. A fourth cited sources but never once reached the firm's website in a buyer question.

One platform named the firm zero times in 360 answered discovery events. Not once, in any market, on any buyer situation — and it is the platform with the largest consumer base.

That concentrated the firm's entire content investment on a single assistant.

Finding 5 — Two public records the firm controls, telling different stories

The firm's website published location pages for eleven cities. Its SEC filing named ten locations with street addresses. Only nine appeared on both.

AI followed the website.

One marketed city with no named office in the filing was confirmed as an office by one platform on five of five asks — with a different street address on different asks. Two addresses for one office, on the same question, hours apart. Neither matched any address in the filing.

The instability was the finding, not the address. A source being read does not change between asks. A model filling a gap does.

A second marketed city was correctly declined by two platforms, one of which sensibly redirected to a nearby filed office. So marketing a city page is not on its own sufficient to produce a confirmed office in AI's answers — which is what makes the first case worth attention rather than assuming.

The same pattern appeared in leadership. Two platforms named a CEO who stepped down years earlier — a real person, in the wrong role, still listed on the filing as Chairman. The two platforms that read sources got it right. The two answering from memory did not.

And in assets under management: one platform matched the filed figure to four decimal places, citing its source. Another understated the firm by 45%. A prospect could come away believing the firm managed $2.3 billion or $4.2 billion depending on which assistant they opened.

What we got wrong, and corrected

The first delivery reported that Google returned no answer on all 460 financial-services queries — described as platform suppression on this topic class.

That was our error, not the platform's.

Our collection harness requested Google's classic AI Overview, which returns only a token stub for conversational-length prompts and does not render at all beyond a certain length. Testing established the boundary precisely: the same underlying question returned nine text blocks at 32 characters and no result at all at 203. Every prompt in the instrument ran 130–260 characters.

We re-collected all 1,090 Google events using the conversational surface — AI Mode — which is the comparable instrument to the other four platforms.

Google named the firm in 445 of them, and returned 11,195 citations where the original run had recorded none.

In the firm's home market, Google named it in 205 of 460 events. The original reading had shown a blank column, and the correction reversed it in the firm's favor.

On the two marketed-but-unfiled cities, Google's answers diverged in exactly the way the earlier findings predicted. For one, it qualified the answer correctly — noting the office is designated for that city but physically located in the neighboring one, and supplying that address. For the other, it asserted an office with a specific street address and no qualification.

The difference was not AI's carefulness. It was what the firm had published. One city page shows the neighboring address; the other shows a local one. The machine reproduced each faithfully.

That correction was issued unprompted, in writing, with the corrected data. We would rather tell a firm we were wrong than have them act on a finding that isn't.

What the research produced

Not a score. A document the firm could act on, and a set of decisions rather than instructions:

• Which of six buyer questions are winnable, and which two are contests an advisory firm currently cannot win because AI routes them to another profession
• Which content asset carries recommendations and which one is read without producing them
• That the firm's largest unmarketed office is recommended only under another city's name
• That one channel of two is working, and that the second is already being paid for
• That the marketed footprint and the filed footprint diverge — a disclosure-adjacent question as much as a marketing one, which is why it belongs in front of the Chief Compliance Officer as well as the growth lead.

We did not do the work that followed. The firm's agency did. That is the line we hold, and it is why we can measure again afterwards and have the result mean something.

What this study does not claim

• That AI visibility produced clients for this firm. No such link is established and none is asserted.
• That any page shape causes a recommendation. The content-type ratios describe what was read on one date.
• That the firm markets offices it has not filed. Fifteen of the filed locations are private residences with addresses withheld, so the public record cannot resolve it. The defensible statement is narrower: two cities are marketed with commercial addresses and are not named locations in the filing, and AI repeats the marketing.

Corrections and disputes

If we have described a figure or a platform's behavior incorrectly, tell us. We will check it against the source, publish a dated correction if we were wrong, and say so plainly. We correct errors of fact. We do not change findings because someone dislikes them.

AIWorthy is an independent research firm measuring how AI systems find, describe, credit and recommend companies, checked against the public record. We publish what we find, including where our own reference data was wrong. We do not grade a remedy we have been paid to deliver, we do not audit a party we have been paid by, and every commercial relationship is disclosed.

Footer language updated 14 September 2026 to reflect current service scope. No findings, figures, methods or artifacts were changed.

hharreld@aiworthy.ai · aiworthy.ai