Three doors into an AI answer — and the audit firms standing outside all of them

The accounting firms that perform the most Single Audits in the Research Triangle are, for the most part, invisible to the AI assistants a nonprofit finance director now asks for a recommendation — and the firms those assistants do recommend got in through three different doors, none of which was the federal record of who actually does the work. The study measured CPA firms in the Research Triangle, North Carolina.

That is the finding of a two-collection study by AIWorthy, an independent research lab in Apex, N.C., which asked five AI platforms — ChatGPT, Claude, Perplexity, Gemini and Google AI Mode — the ten questions a Triangle nonprofit asks when it needs its first Single Audit, five times each, on Aug. 11 and again on Sept. 28, 2026. Five hundred answers, checked one by one against the Federal Audit Clearinghouse, the mandatory public record that names the auditor on every Single Audit engagement in the region.

The ten firms on the study's panel are named as auditor on 75 Triangle Single Audit engagements between them. The four largest filers among them — 45 of those 75 engagements — were recommended 24 times in August and 17 in September, combined. The six smallest — 30 engagements — were recommended 184 times, then 163. The most-recommended firm in the panel is one of its three smallest. The single most prolific nonprofit auditor in the whole regional record, a firm just outside the panel's three counties with more Single Audit clients than any panel firm, was recommended zero times and its website was cited zero times in 2,725 citations.

The federal record itself was consulted twice in 1,443 citations in August and twice in 1,282 in September. The state licensing board: never.

The doors

The study's method is to read not just who was recommended but what the machine was reading when it made the recommendation. Sorted that way, every recommendation in 500 answers came through one of three doors.

Door one: the machine read the firm's own page. This is the door the two platforms that search the live web — Perplexity and Google AI Mode — use most. It is also the door that swings. The most-cited website in the entire study belongs to a Raleigh firm with fewer Single Audit clients than any firm on the panel. One page did it: a 1,436-word nonprofit industry page that names Raleigh and the Triangle 21 times. Between the two collections, both retrieving platforms came to treat that page as the reference for "Triangle nonprofit audit firm." It was cited 40 times in September, more than any other page, and the firm's site went from 55 citations to 103.

Door one closes as easily as it opens. Another local firm was cited 30 times in August and recommended 21 times; in September it was cited once and recommended zero times. Its pages had not changed. They still answered at the same addresses on Sept. 28. Google had read them on four of ten questions in August and on none in September. Being read in August bought nothing in September. Between the two dates — with the underlying AI models unchanged on every platform — Google AI Mode rebuilt every one of its ten answers from a different source list, and the number of domains behind those answers fell from 83 to 39.

Door two: the machine already knew the name. ChatGPT, Claude and Gemini answered these questions largely from stored knowledge, citing nothing, and what they know did not move in six weeks. Summing every panel firm's change on each platform, ChatGPT moved five recommendations, Claude four, Gemini seven — against 13 and 15 on the two platforms that read. The most-recommended panel firm, named 100 times in August and 97 in September, is a door-two firm: its website was cited 29 times, then 10, and only three of its recommendations came in an answer that cited the site at all. Claude named it 22 times out of 50 on both dates; Gemini 36 out of 50 on both.

The purest door-two case is a panel firm recommended 78 times across both dates — 60 of them by Claude — with its own website cited zero times. On Sept. 28 that website was a single "Our Site is Undergoing Maintenance" notice at every address tried. So where did the machines learn the name? From the firm's clients. The platforms that read found it in the State Treasurer's local-government audit files, in a city council's minutes and, on both dates on both retrieving platforms, in one charter school's online archive of its audited financial statements. That archive labels each year's report by auditor: this firm through fiscal 2022, a different firm for 2023, 2024 and 2025. The school changed auditors three years ago. In September, Perplexity recommended the firm for a charter school Single Audit citing that archive, and added that the firm performs audits under "OMB Circular A-133" — a rule replaced in 2014. Claude, which cites nothing, holds an accurate sketch of the firm as a municipal auditor and recommends it, five times out of five on both dates, for behavioral-health, legal-aid and child-and-family audits the public record never shows it doing.

A firm's reputation with the machines, in other words, can be held entirely in its clients' filing cabinets — and those cabinets are accurate about the past and silent about the present.

Door three: someone else's list. Commercial directories roughly doubled their share of what the machines read in six weeks, from 92 citations to 190 across four directories and one business journal's ranking of the region's largest firms, while the nonprofit sector's own membership roster halved, from 42 to 21. Every one of the September directory citations came from Perplexity. No firm's rise has been traced to the directory growth.

The third-party document that did get firms recommended was a help-wanted ad. On the charter-school question, Perplexity told the buyer in all five asks that its evidence for recommending one panel firm was "a job posting for the firm [that] states experience auditing charter schools and nonprofits." In two of the five it added that the posting was "less specific about North Carolina charter-school clients," then recommended the firm anyway. By that evening, the posting it cited returned "404. Page Not Found." Google AI Mode did the same on another question with a regional firm's tax-manager listing on a jobs aggregator, still live.

The wrong door

One panel firm renamed itself in April 2024; the old name's registration expired that May. The platforms recommended it under the old name 62 times in August and 61 in September, and under its current name never. Sixty-one of the 61 September mentions, read by hand, present it as a firm to hire today. Through August those came only from the door-two platforms, answering from memory. In September, Google AI Mode joined them, and the source it cited was the firm's own website — homepage title, service page and team biography all still under the old name — plus a LinkedIn company page never renamed. The domain was changed. The words on it were not. A retrieving platform read the firm's own site and recommended it by the name its own site still uses.

A second panel firm's Durham office was recommended 15 times across the two dates under the name it had before a June 2023 merger — every time as a firm to hire today — and the firm was recommended once under its current name. Add a national firm that merged in 2022 and is still recommended under both of its former names, and 79 of 250 answers in August and 85 of 250 in September told a buyer to call a firm that no longer exists.

No door at all

Three panel firms — including the panel's largest filer and its second-largest — were recommended zero times and cited zero times on both dates. AIWorthy read their websites after the second collection.

The largest filer's site is fast, open to every crawler and 1,821 pages deep. Its nonprofit page uses "nonprofit" 20 times and names a place zero times; so does its audit page; so does its Single Audit article. Its homepage title is "Public Accounting Firm | Professional Services Firm." The federal record shows the firm audits three Raleigh-area behavioral-health organizations; the study asks a behavioral-health question; the firm was named on it zero times in ten answers. The content that would answer a Triangle buyer's question is on the site, and nothing on it says Triangle.

The second-largest filer files every federal engagement from a domain that has email and no website. The third firm files from a domain whose server refuses every connection, while its actual website — at a different address, 371 pages, open to every crawler — was never read once; its nonprofit pages, too, name no place. And the most prolific auditor in the record publishes a robots file, built from a vendor's template, that admits Google's crawler and shuts out the crawlers behind two of the five platforms.

Four firms, four different reasons, and a visibility score would report all of them as the same zero.

What the pages behind door one have in common

Twenty pages were read side by side: the seven most-cited, six that were cited in August and dropped in September, and the never-cited pages of the largest filer. Every page cited on both dates names the Triangle between 11 and 32 times, describes the buyer's own service, and runs to depth. Each dropped page has two of those three — a place without the service, the service without a place, or both without depth. The never-cited firm has the service and the depth and no place, on every page.

And the pages that got read for Single Audit questions barely use the words "Single Audit" — 0, 0, 0, 1, 2 and 6 times across the six most-cited. The buyer's question describes a situation; what was retrieved matched the situation, not the term.

What showed no association at all: page speed (the most-cited site has the worst mobile lab scores in the set); ranking on page one of Google for the buyer's question (3 of 49 cited pages) or for an obvious keyword search (7 of 49); structured data (more common on the never-cited pages than the cited ones); and every source the current AI-visibility advice emphasizes — Reddit, Quora, forums, Wikipedia and YouTube drew zero of 2,725 citations, and there was no news coverage of any Triangle audit firm anywhere in the record. The one piece of that advice the record supports is freshness: three in five dated pages the machines read had been changed within six months.

Including your mistakes

The federal Single Audit threshold rose from $750,000 to $1,000,000 for fiscal years beginning on or after Oct. 1, 2024. Twenty-six answers in August and 22 in September stated the old figure as current, without qualification. Google AI Mode was the largest contributor on both dates, and on every September Google answer that repeats the figure, the cited sources include one Triangle firm's own blog post — the page responsible for every recommendation that firm received — which states $750,000 three times, and a second Triangle firm's audits page, last modified in August 2023, which states it once. It is not one stale post. It is the local professional web, reproduced with citations attached, in front of a prospective client.

What is in it for a CPA firm

None of this shows that any page, title or listing causes a recommendation, or that a recommendation produces a client. The study is ten questions, five platforms, two dates, one region, reported as counts. But a firm can act on a risk without a proof of cause, and the record puts ten things in front of a managing partner.

1. Your prospects' first question is now answered without you in the room — and you can read the answer. Ten questions, five platforms, 500 answers: this is what a finance director sees before she calls anyone. Unlike a referral conversation, it is checkable. A firm can ask the same question today and read exactly what a buyer is told, including which competitor is recommended and on what evidence.

2. Your Single Audit track record does not count. The federal record naming you as auditor on every engagement was consulted four times in 2,725 citations. If the practice-development story is "we have the most experience," the machines cannot hear it; they are not reading where that story is written.

3. There are two games, and most firms are in only one. Three platforms answer from memory and moved by four to seven recommendations across all ten panel firms in six weeks. Two platforms read the live web and moved by 13 and 15. Ask an assistant about your firm by name and you test recall. Ask it the buyer's question without your name and you test discovery. A well-known, unread firm wins recall and loses discovery; a firm read on one good page wins discovery and, if the models do not know it, nothing else. Know which game you are in before spending on either.

4. What the models "know" about you is old, and you cannot edit it — but you can find out what it is. Two panel firms are recommended under retired names 76 times in September; one is recommended 78 times from clients' minutes and a school's archive of reports through 2022. The models' picture of a firm is the web as it stood when they were trained. The practical question is not "how do I fix it" but "what is it saying, and what newer page of ours corrects it?" If the answer is nothing of yours, that is the gap.

5. If you renamed or merged, the old name is still working — for you and against you. The renamed firm's homepage title, service pages, team biographies and LinkedIn page all still carry the old name; when Google read them, it recommended the old name. A buyer told to call a firm that no longer exists may not find you. Any firm that has changed its name since about 2020 should search its own site and its LinkedIn page for the old one.

6. The pages that get read have a shape, and it is cheap. A place, named often; the buyer's own service; depth. The largest filer on the panel has all the content and names a place zero times on any page, and was never read. A firm does not need a content program. It needs one page per service line that says where it is, whom it serves and what it does, at length — and a link to it from the homepage, which several firms in this study lack.

7. Reach is a precondition, and a third of the record's firms have a problem they do not know about. Of 57 auditor domains in the federal record, 10 publish a vendor-template robots file that blocks every crawler it does not name, nine answer OpenAI's crawler differently from a browser, four filed domains lead to no site, one firm has no website and one panel firm's whole site was a maintenance notice. None of these is a marketing decision. All of them are a fifteen-minute check by whoever manages the site, and the one thing on this list a firm can fix by Friday.

8. Your own pages are telling prospects the wrong threshold. Search your site for "$750,000." Every Google answer that repeated the superseded figure cited a Triangle firm's own page. This is a competence signal in front of a prospective client, not a marketing one.

9. Most of what firms are being sold to fix this showed nothing. Zero citations to Reddit, Quora, forums, Wikipedia or YouTube. Structured data as common on never-read pages as on read ones, and FAQ schema on none of the top twenty. The slowest site in the set was the most cited. Page-one ranking for the buyer's question: 3 of 49 cited pages. Freshness is the one piece of the standard advice the record supports. Before signing an AI-visibility retainer, ask the vendor which of these they intend to work on, and what evidence they have for it.

10. It changes in six weeks, so one check is not a check. The page that earned a firm 21 recommendations in August earned it zero in September without changing. What the retrieval layer reads is rebuilt continuously; what the models remember is frozen until the next model. A firm that checks once has learned what one day looked like.

The bottom line for a managing partner: the machines are answering your prospects' first question from an old picture of the market, a handful of place-and-service pages and other people's lists — not from the record of who does the work. That is bad for the firms doing the most of it, and, for the moment, checkable and partly fixable by anyone who looks. The study does not say that fixing it wins clients. It says what the buyer is being told, and that most firms have never read it.

About the study

The panel, questions and scoring were frozen before any firm was contacted, and the results publish regardless. All measured firms received their complete records with a right of reply before publication; none requested a correction. No firm is named; results are not attributed to firms. Any measured firm can purchase its own record or technical read from AIWorthy; doing so does not change what is published, and paid relationships with measured firms are disclosed on the record page. AIWorthy does not sell placement, rankings or content. Methods, corrections and frozen-artifact hashes are on the record page at aiworthy.ai.