The Firm AI Recommends 78 Times Has a Website That Says "Undergoing Maintenance"
The second most-recommended Single Audit CPA firm in our study of the Research Triangle, North Carolina, never had its own website cited — not once in 2,725 citations. When we went to look at that website, we found out why. Then we found out where the machines had learned the firm's name instead, and that is the more useful story.
This case file expands Result 3 of Three Doors and a Brick Wall: How AI Assistants Choose CPA Firms
By Heather Harreld · Published October 7, 2026
There is a small accounting firm in Wake County that five AI assistants recommended 78 times across our two collections — 44 times in August, 34 in September — for Single Audits of nonprofits and governments in the Triangle. That made it the second most-recommended of the ten firms on our panel, on both dates, ahead of firms with two and three times its number of Single Audit clients in the federal record.
Recommendations across both dates, 44 in August and 34 in September. Second most-recommended of the ten panel firms.
Citations to the firm's own website, out of 2,725.
Characters of visible text on its homepage on September 28, all of them "Our Site is Undergoing Maintenance."
Its own website was cited zero times. Zero in 1,443 citations on Aug. 11. Zero in 1,282 on Sept. 28.
On the evening of Sept. 28 we opened the site. The homepage title read "Our Site is Undergoing Maintenance." The page carried 173 characters of visible text. We tried two other addresses from the site's own sitemap, which still lists 119 pages; both returned the same title. The whole site, not just the front door, is one maintenance notice. The server also answers OpenAI's crawler with an HTTP 429 — "too many requests" — while answering a browser normally. Google's ordinary search, asked for the firm by name, describes it in its knowledge panel as a tax preparation service and links to the maintenance page.
Whether the site was in this state in August we cannot say. The Internet Archive's last working capture of the homepage is from July 2025.
So how does a firm with nothing on the web get recommended 78 times?
| Assistant | August (of 50) | September (of 50) | Cites sources? |
|---|---|---|---|
| Claude | 32 | 28 | Never |
| Perplexity | 5 | 3 | Always |
| Google AI Mode | 5 | 0 | Always |
| Gemini | 2 | 2 | Rarely (6–7 of 50) |
| ChatGPT | 0 | 1 | Never |
| Total | 44 | 34 |
The door the firm came through
In the study we sort every recommendation by what the machine was reading when it made it. There are three doors. Door one: the platform read the firm's own page. Door two: the platform already knew the name and cited nothing. Door three: the platform read someone else's document about the firm.
This firm never came through door one. It came through door two, mostly, and door three for the rest.
Door two — 60 of the 78. Claude, which cites no sources on any of these questions, named the firm 32 times out of 50 in August and 28 out of 50 in September. What Claude holds is specific and, as far as the public record goes, accurate: the town, the county, local governments, schools, housing, "mid-size." Two of its descriptions, with the firm's name and town redacted:
[The firm] ([town], NC) — Experience with school audits in the Triangle area
[The firm] ([town]) — Local government and school focus
That sketch matches the firm's engagements in the Federal Audit Clearinghouse. Then Claude applies it to questions the record does not support. On our behavioral-health question, our legal-services question and our Durham child-and-family-health question, Claude recommended this firm in five out of five asks each, on both dates. Nothing in the federal record, the State Treasurer's files or the citation record shows the firm doing that work. A model that knows a firm is a competent local auditor, and knows nothing else, fills the gap with the question it was asked.
Gemini, by contrast, named the firm four times, all on the municipal question, and gave the most accurate description of the firm anywhere in our 500 answers: "a local firm directly in Wake County with extensive experience auditing NC municipalities, small towns, and special districts." Right firm, right work, on the one question that fits — and only there.
Door three — the remaining 18. The platforms that read the live web found the firm too: Perplexity and Google AI Mode 13 times in August, Perplexity 3 times in September, plus Gemini's handful of cited answers. Every one of those readings was of somebody else's document. In August, the most-cited source on responses naming the firm was the State Treasurer's local-government audit files (17 citations). Then a competitor's charter-school page, a PDF on a content-delivery network, a municipal meeting-attachment store, a city council's minutes, two commercial directories.
| Source cited on answers naming the firm | August | September |
|---|---|---|
| State Treasurer's local-government audit files | 17 | 1 |
| A competitor's charter-school page | 20 | 9 |
| A charter school's financial-history archive | 10 | 3 |
| PDF on a content-delivery network | 10 | — |
| Municipal meeting-attachment store | 5 | 3 |
| Commercial directories | 10 | — |
| A nonprofit database built from Form 990s | — | 9 |
| State education department | — | 6 |
| — | 3 | |
| The firm's own website | 0 | 0 |
And one page that both retrieving platforms cited, on the charter-school question, on both dates: a charter school's online archive of its audited financial statements.
The filing cabinet
That archive is a simple page. It lists the school's audit reports by fiscal year, each labeled with the auditor. This firm is the auditor for fiscal 2018 through 2022. A different firm is the auditor for 2023, 2024 and 2025.
The school changed auditors three years ago. The page says so, to anyone who reads down to the bottom.
- FY2018This firm
- FY2019This firm
- FY2020This firm
- FY2021This firm
- FY2022This firm
- FY2023A different firm
- FY2024A different firm
- FY2025A different firm
In September 2026, Perplexity recommended this firm for a charter school Single Audit on the strength of that page. Its sentence, with the firm's name redacted:
[The firm] | They say they have audited many charter schools since 1998 and perform audits under Government Auditing Standards and OMB Circular A-133.
OMB Circular A-133 was superseded by the Uniform Guidance for fiscal years beginning on or after Dec. 26, 2014. Perplexity attributes the sentence to the archive page. The archive page's HTML does not contain "1998," "A-133" or the firm's name as text; the audit reports it links to are scanned images. Where the sentence actually came from we could not establish, and we report it only as Perplexity's attribution. What we can say is that a buyer following the citation would land on a page whose most recent five entries name a different firm.
Google's ordinary search tells the same story from another angle. On Sept. 28, the top eight organic results for the firm's name were two employees' LinkedIn profiles, three directory ranking pages on which the firm is one entry among hundreds, a national business directory, a city council's 2020 minutes and a school board's 2010 minutes. Not one result is a page the firm wrote about itself.
| Rank | What it is | Whose page |
|---|---|---|
| 1–2 | Two employees' LinkedIn profiles | |
| 3–5 | Three directory ranking pages, the firm one entry among hundreds | Commercial directories |
| 6 | A national business directory | Directory |
| 7 | A city council's 2020 minutes | A client |
| 8 | A school board's 2010 minutes | A client |
What it adds up to
A firm can be recommended 78 times while its own site says nothing, because the record of its work is held by its clients: minutes that note the audit was presented, a school's archive of old reports, a state agency's files, a nonprofit database built from Form 990s. That record has three properties a firm should understand.
Accurate about the past. Every document the machines read describes work the firm did. Silent about the present. None of it changed when the firm's site went dark. Not the firm's to edit. A school's archive, a council's minutes and the Treasurer's files belong to someone else.
It is accurate about the past. Every document the machines read about this firm describes work it did.
It is silent about the present. The archive will keep presenting the firm as a charter-school auditor for as long as the 2022 report stays online, and none of it said anything when the firm's own site went dark.
And it is not the firm's to edit. The firm cannot update a school's archive, a council's minutes or the Treasurer's files. It can only publish something of its own that is newer, deeper and reachable — and on Sept. 28 it had not.
The three platforms behave differently with that record. The one that names the firm most holds an accurate sketch of a municipal and school auditor and applies it to nonprofit specialties the record does not show. The one that describes the firm most accurately does so four times, on the one question that fits. The one that cites a source for its description quotes a regulatory reference eleven years out of date and attributes it to a page that does not contain it.
What a firm can take from this
- Your clients' documents describe you whether or not you do. Search your own name in an ordinary search engine and read what comes back. If none of it is yours, that is what the machines have.
- A dark site is not neutral. It does not remove you from the answers — the models already know your name — but it removes the one source that could correct what they know.
- Check what your server tells the crawlers. This site answers OpenAI's crawler with a 429 while answering a browser normally. Most firms have no idea their host does this.
- Ask an assistant about your firm by name, then ask it the question a buyer would ask without your name. This firm passes the first test and, on three specialties, passes the second for work the record does not show it doing. Both results are worth knowing.
What this does not show
Not why any platform named the firm; correspondences between sources and namings are observations on two dates. Not the state of the site on Aug. 11. Not the origin of the "since 1998 / A-133" sentence. Not that the firm does or does not perform the work Claude attributes to it beyond what the federal record shows. Not a rate: 78 of 500, on two dates.
AIWorthy measures and reports. 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 before publication. 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-file hashes are on the record page.
Back to the visual summary