Three Doors and a Brick Wall: How AI Assistants Choose CPA Firms
Your next client's first question is now answered by a machine. We read the answer. We asked five AI assistants the questions a nonprofit in the Research Triangle, North Carolina, asks when it needs an audit firm, five hundred times, and checked every answer against the federal record of which CPA firms actually do that work. The firms doing the most of it were barely mentioned. The ones recommended most got in through three different doors. Here is what we found, and what a firm can do about it.
Buyers used to ask a colleague. Now they ask an assistant, and the assistant answers in full sentences, with names.
When a finance director types "we need a Single Audit for the first time, which firms should we talk to?" into ChatGPT, Perplexity or Google, she gets a short list of firms, a sentence about each, and sometimes a link. She did not visit your website. She may never visit it. Whatever that answer said about you, or failed to say, is the first impression.
Unlike a referral conversation, that answer can be read. Anyone can ask the same question today and see which firms are named, in what words, on what evidence. This study did that, systematically, twice, and wrote down everything.
Almost certainly, and this study does not measure how buyers choose. It measures what a buyer is told when they ask. Buyers ask for two reasons: to check a referral before calling ("is this firm any good at charter schools?"), and because a first-time filer often has no one to refer them. The record shows what both of those buyers are told. For one firm in this study it was a name the firm dropped in 2024. For another, a federal rule replaced in 2024.
A Single Audit is the one audit where the government publishes who did the work.
Any organization that spends $1,000,000 or more in federal funds in a year (it was $750,000 before October 2024) must have a Single Audit: one combined audit of its finances and its compliance with federal grant rules. Nonprofits, charter schools, housing providers, community lenders and towns all need one. It is specialized work, and not every CPA firm does it.
Every Single Audit is filed with the Federal Audit Clearinghouse, a public federal database, and the filing names the auditor. That makes this the rare professional service with a complete, public record of who does the work. We did not have to guess which firms want these clients or claim this expertise. The record shows they already did the work, for named organizations, year after year.
Triangle firms named as auditor on three or more Single Audits in the federal record, 2023 to 2025. Every firm that met the screen. Five national firms were tracked separately.
Single Audit engagements those ten firms are named on, from 3 to 13 each. Real clients, real filings.
Written as a buyer would ask, each derived from the kinds of organizations these firms actually audit. No firm is named in any question.
Five assistants, five asks per question, on August 11 and again on September 28, 2026. Same questions, same model versions, both times.
Assistants: ChatGPT, Claude, Perplexity, Gemini and Google AI Mode, through their public interfaces for developers, with no chat history and no personalization. The questions, the panel and the scoring rules were frozen before the first answer was collected.
The firms that do the most of this work were recommended the least.
We split the ten firms by how many Single Audits the federal record shows them doing. The four busiest firms account for 45 of the panel's 75 engagements. Between them, they were recommended 24 times in 250 answers. The six smallest firms, with 30 engagements, were recommended 184 times. The most-recommended firm in the study is one of the three smallest.
The public record of who does the work was almost never consulted.
Two of the five assistants show their sources. Across both dates they cited 2,725 web pages. The Federal Audit Clearinghouse, the database that names the auditor on every one of these engagements, was cited four times. The state board that licenses CPAs: never.
Why this matters more than it sounds. A firm's track record is the strongest evidence it has. In this market that evidence is public, mandatory and complete, and the machines did not look at it. They recommended firms based on what firms say about themselves, what directories list, and what the models happened to learn years ago. If your practice-development story is "we have the most experience," the machines cannot hear it, because they are not reading where that story is written.
citations in 2,725 to the Federal Audit Clearinghouse.
citations to the NC State Board of CPA Examiners.
of August citations; 53% in September. The assistants mostly read what firms wrote about themselves.
of citations, August to September. Paid listings doubled their share in six weeks.
Every recommendation came through one of three doors. The busiest firms stood outside all of them.
For every answer, we looked at what the assistant was reading when it named a firm. Sorted that way, there are only three ways in, and one wall.
It read your website
Perplexity and Google AI Mode search the live web and cite what they find. A firm with fewer Single Audit clients than anyone on the panel became the most-cited website in the study on the strength of one page about nonprofit work that names Raleigh 21 times. This door also closes: another firm's pages were read 30 times in August and once in September, without changing.
It already knew your name
ChatGPT, Claude and Gemini mostly answer from what they learned in training and cite nothing. The most-recommended panel firm came through this door: 100 recommendations, website cited 29 times, then 10. One firm was recommended 78 times while its own website showed only "Our Site is Undergoing Maintenance."
Someone else wrote about you
Directories, a client's online archive of old audit reports, a town council's minutes, a help-wanted ad. One assistant recommended a firm for charter-school work citing a school's archive that shows the firm stopped being that school's auditor three years ago. Another cited a job posting as its evidence; the posting was gone by that evening.
Nothing to find
Three panel firms, including the one with the most Single Audit clients, were recommended zero times and had their websites cited zero times on both dates. Each for a different reason, below.
Three assistants gave the same answers six weeks later. Two changed theirs.
We asked everything again on September 28 with the same model versions. The assistants that answer from memory (ChatGPT, Claude, Gemini) named almost exactly the same firms, the same number of times. The assistants that search the web (Perplexity, Google AI Mode) named different firms, because the web pages they chose to read had changed.
The chart adds up, for each assistant, how much every firm's count moved between the two dates. A small number means the answers held; a large number means they were rebuilt.
What this means for a firm: What a model remembers about you will not change until the model itself is retrained, no matter what you publish this quarter. What the web-searching assistants say about you can change in six weeks, and the only lever is what they can read on your site today.
Google's answers were rebuilt from scratch, using fewer websites.
Google AI Mode gives the same answer every time you ask on a given day, so it is the easiest to compare. For each of the ten questions, we counted how many different websites Google drew on in August and how many in September. Not one question kept the same wording or the same set of sources. Across all ten, the number of websites behind the answers fell from 83 to 39.
In 79 of 250 answers, the buyer was told to call a firm that no longer exists under that name.
One panel firm changed its name in April 2024. The assistants recommended it under the old name 62 times in August and 61 in September, and under its current name never. In September, Google started doing it too, and the evidence Google cited was the firm's own website, whose homepage title, service pages and LinkedIn page all still carry the old name. A second panel firm's Durham office was recommended 15 times under the name it had before a 2023 merger. A national firm that merged in 2022 is still recommended under both of its former names.
of 250 answers, August then September, presented a firm under a name it no longer uses, as a firm to hire today.
times the renamed firm was recommended under its actual name, on either date. Its website was read; the old name is what the site still says.
given as the current Single Audit threshold in 22 September answers. It has been $1,000,000 since October 2024. The cited sources were local firms' own pages.
recommendations for one firm whose entire site was a maintenance notice. The assistants learned about it from its clients' documents.
Four firms scored zero, three on the panel and one tracked separately, for four different reasons. A score would report them all the same way.
We read the websites of the firms that were never recommended, against a checklist written before we opened any of them. Each zero has its own cause, and each cause needs a different fix. A single visibility number, the kind sold by most AI-marketing tools, would have told all four firms the same thing and none of them anything useful.
No website
The panel's second-busiest firm files every engagement from a web domain that has email and no website. Nothing exists for a machine to read.
Wrong address
One firm's domain on its federal filings is dead. Its real website, at another address, is 371 pages, open to every crawler, and was never read once.
No place on any page
The busiest firm on the panel has 1,821 pages, fast and open. Its nonprofit page says "nonprofit" 20 times and never names a city or region. Nothing on the site says Triangle.
Crawlers locked out
The most prolific nonprofit auditor in the regional record uses a website template whose settings admit Google and block the crawlers behind two of the five assistants.
The assistants did not read Reddit, Wikipedia or YouTube. They read firms' own pages, government files and directories.
Most advice on "AI visibility" tells firms to get mentioned on Reddit and Wikipedia, add FAQ markup, publish to YouTube and speed up their site. We checked each against what the assistants actually cited:
- Reddit, Quora, forums, Wikipedia, YouTube: 0 of 2,725 citations. Not one.
- Structured data (schema markup) was on 69% of cited pages and on 100% of the never-cited pages we checked. FAQ markup was on none of the twenty most-cited pages.
- Page speed: the most-cited website in the study had the slowest mobile scores in the set.
- News coverage: none of any Triangle audit firm anywhere in the record. "News" citations were almost all one business-journal ranking list.
- Google rankings: did not put a page into the AI answer. The next section shows that one on its own.
- Freshness is the one piece of the standard advice the record supports: three in five dated pages the assistants read had been changed within six months.
What the cited pages had in common instead: every page cited on both dates named the Triangle, between 11 and 32 times; described the buyer's own service; and ran to real length. The pages that were dropped between August and September each lacked one of those three. And the pages read for Single Audit questions barely used the words "Single Audit." The assistants matched the buyer's situation, not the keyword.
Ranking on page one of Google did not put a page into Google's AI answer.
If AI answers simply repeated the top search results, a firm that ranks well would be covered. We tested that on September 28. Google's AI Mode cited 49 web pages across the ten questions. The same day, we ran each question through ordinary Google search two ways: typed exactly as the buyer asked it, and as three shorter searches a buyer might type instead ("nonprofit audit firm Raleigh NC," "charter school auditor North Carolina," and so on).
Of the 49 pages the AI cited, 3 were on page one of Google for the buyer's exact question. 7 were on page one for any of the thirty shorter searches. 42 were on page one for none of them. Ordinary search, given the buyer's full question, mostly returned pages that had nothing to do with it; the AI's citations were on topic every time. The two systems were choosing different pages.
What this means for a firm: the pages the AI chose look like good search-optimization work, on the firm's own site, naming the place and the service. But which of those pages get read is not decided by the search ranking a firm can see. A firm can rank first for its keyword and be absent from the AI answer, or be in the AI answer with a page that ranks nowhere. Checking your rankings tells you about one of the two systems.
One day, one location setting, Google AI Mode only. Google says AI Mode runs several searches behind the scenes for each question; those searches are not visible, and this comparison does not test them.
Five things to do this month, none of which require buying anything.
The study does not show that any of these produces a recommendation, and it does not show that a recommendation produces a client. It shows what the buyer is told, and these are the places where the record says that answer comes from.
Ask the question yourself, two ways.
Open any AI assistant. Ask the question a prospect would ask about your specialty and your area, without naming your firm. Then ask it about your firm by name. The first test is discovery; the second is recall. Only discovery resembles what a buyer does. Most firms have never read either answer.
Search your own site and your LinkedIn page for any name you no longer use.
Two panel firms were recommended 76 times in September under retired names. One of them is the source of its own problem: the homepage title still carries the old name. If you have renamed or merged since 2020, this is the first thing to fix.
Read your homepage title tag and your robots.txt file as a machine would.
Does the title say your current name, where you are and what you do? Does yourdomain.com/robots.txt block AI crawlers? About a third of the firms in the federal record block at least one, usually through a website template nobody chose. This is a fifteen-minute check for whoever manages your site.
Give each service line one page that names the place, the service and the work, at length.
Not a content program. One page per specialty that says where you are, whom you serve and what you do, and a link to it from the homepage. The busiest firm on the panel has all of that content and names a place on none of it.
Search your site for "$750,000," and check again in six weeks.
Your own pages are the sources the assistants cite, mistakes included. And what the web-searching assistants read is rebuilt continuously, so a single check tells you what one day looked like.
Three Doors and a Brick Wall: How AI Assistants Choose CPA Firms
The full article walks through each door with the evidence behind it, the ten takeaways for a CPA firm, and the questions a skeptic should ask. The complete study, every count, the corrections made during verification, and the frozen-file hashes are on the record page.
No firm is named in this summary; results are not attributed to firms. The panel, questions and scoring were frozen before any firm was contacted, and results publish regardless. Every measured firm received its complete record with a right of reply. Any measured firm can purchase its own record or technical read from AIWorthy; doing so does not change what is published, and paid relationships are disclosed on the record page. AIWorthy does not sell placement, rankings or content.