Three Doors and a Brick Wall: How AI Assistants Choose CPA Firms

How firms are referred to. No firm is named in this study. The ten panel firms are listed on the study's record page at aiworthy.ai; their individual results are not attributed to them here. Panel firms are described by what happened to them in the measurement ("the most-named panel firm," "the firm renamed in 2024"), not by name, town or exact client count. Firms outside the panel that appeared in the answers are described the same way. Because the roster is public, some descriptions can be matched to a firm with effort; AIWorthy does not attribute results to firms and will not confirm identifications. Every measured firm has received, or will receive before publication, its own complete record.

---

Why this should matter to a firm that does Single Audits

The buyer's first question is now being answered by a machine, and the machine's answer is checkable.

A finance director whose nonprofit just crossed the federal spending threshold types a question into an AI assistant: we need a Single Audit for the first time — which firms in the Triangle should we talk to? We asked that question, and nine more like it derived from the organizations Triangle firms actually audit, of five AI platforms, five times each, on 11 August and again on 28 September 2026. Two hundred and fifty answers each time. Then we checked every answer against the federal record that names the auditor on every Single Audit engagement in the region.

Here is what a firm in that record should take from the result.

1. Doing the work is not the same as being found for it

Ten firms qualified for the panel by auditing three or more Triangle organizations under the Single Audit requirement; between them they are named as auditor on 75 engagements in the federal files. The four panel firms with the most engagements — 45 of the 75 — were named in 24 of 250 answers in August between them, and 17 in September. The six with the fewest — 30 of the 75 — were named 184 times in August and 163 in September. The firm named most often, 100 times in August and 97 in September, is one of the panel's three smallest filers. The single most prolific nonprofit auditor in the entire record, outside the panel and with more Single Audit clients than any panel firm, was named zero times on both dates and its website was cited zero times in 2,725 citations.

Nothing about the federal filing reached the answer. The Federal Audit Clearinghouse was consulted twice in 1,443 citations in August and twice in 1,282 in September. The state licensing board was consulted never. A firm's engagement history is public, mandatory and complete, and it played no part in whether the firm was recommended.

2. A firm can be recommended under a name it no longer has

A panel firm renamed itself in April 2024; the old name's registration expired 1 May 2024. It was named under the old name 62 times in August and 61 in September, and under its current name zero times; 60 of the 62 and 61 of the 61 present it as a firm to hire today (hand-read on both dates). Through August, all of those came from platforms answering from memory. In September, one of the platforms that reads the live web joined them — and the source it cited for the recommendation was the firm's own website, whose homepage title, service page and team page still carry the old name, and its LinkedIn company page, never renamed. A buyer who follows that recommendation will be told to call a firm that, on paper, ceased to exist seventeen months ago. A second panel firm's Triangle office was named fifteen 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 named once under its current name.

The same mechanism repeated a national firm that merged in 2022. Combined, 79 of 250 answers in August and 85 of 250 in September presented a defunct firm as one to hire today — both figures hand-read.

3. What you tell the machine is what it tells the buyer — including your mistakes

The federal Single Audit threshold rose from $750,000 to $1,000,000 for fiscal years beginning on or after 1 October 2024. In August, 26 answers stated the old figure as current without qualification. In September, 22 answers did (read 29 September; 37 candidates, of which 5 marked the figure as historical without giving the current one and 10 carried it only inside a cited page's snippet). The platform that cites a source for every claim was the largest contributor on both dates — 15 responses in August, 11 in September — and on every one of its September responses that repeats the figure, the sources it cites include a Triangle firm's own blog post that states $750,000 three times and a second Triangle firm's audits page that 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.

4. The web the machines read is not the web you think you're on

Between the two dates, with the underlying models unchanged on every platform, the retrieval layer re-weighted what it reads. Four commercial directories and one business-journal list doubled their share, from 92 citations to 190. The nonprofit sector's own membership roster halved, from 42 to 21. The most-read single page changed from a national firm's Raleigh office page (52 citations) to a local firm's nonprofit industry page (40) — and that firm, which has fewer Single Audit clients than any panel firm, became the most-cited domain in the study, at 103 citations, up from 55. A firm whose site was cited 30 times in August was cited once in September; two others went from 15 and 19 to 1 and 0. The recommendations followed the sources. On the one platform that returns the same answer five times on a date, all ten answers were rebuilt between the dates: not one question kept its text or its source list, and the number of domains behind the ten answers fell from 83 to 39.

The pages that stopped being read did not go anywhere. Thirteen of the fourteen pages those firms had in the August citation record still answered on 28 September, unchanged at their addresses. The retrieval layer did not lose them. It stopped choosing them. Being read in August bought nothing in September.

On the platforms that answer from memory, none of this happened. The same names came back in nearly the same counts — one firm at 22 of 50 on both dates on one platform, 36 of 50 on both dates on another — including the two that no longer exist. What a model remembers about a firm does not change until the model does; what it reads about the firm can change in six weeks.

5. Some firms cannot be found because there is nothing to find

The technical read below was collected and frozen on 28 and 29 September. Each statement is a direct observation from that read.

We read the websites of the firms the study contrasts most sharply. Of the three panel firms named zero times on both dates: one files every federal engagement from a domain that has email and no website at all; one files from a domain whose server refuses every connection while its actual website, at a different address, is open, fast and 371 pages deep — and was never read; one has 1,821 pages, is open to every crawler, and describes itself on its homepage as a "Public Accounting Firm | Professional Services Firm" with no place and no service line. On both of the reachable sites, the nonprofit and audit pages name no place and run under 900 words; the pages that name the place are news items. The most prolific nonprofit auditor in the record, outside the panel, publishes a robots file — a vendor's template — that admits Google and excludes the crawlers behind two of the five platforms. A fifth firm, a panel firm named 78 times across the two dates from models' memory, had a website on 28 September that read "Our Site is Undergoing Maintenance" at every address tried; whether it was in that state in August is not known.

These are five different situations with five different consequences, and a visibility score would report all of them as the same zero.

We then read twenty pages: the ones the machines cited, the ones they cited in August and dropped in September, and the equivalent pages of a firm they never cited. Every cited page names the Triangle, between 11 and 32 times; the never-cited firm's nonprofit page, audit page and Single Audit article name it zero times. But a place name was not enough on its own — one dropped page names Raleigh–Durham thirty times. The pages that were read on both dates share three things: a place, the buyer's own service, and depth. The dropped pages each have at most two. And the cited pages barely use the words "Single Audit" at all: the buyer's question describes a situation, and what was retrieved matched the situation, not the term.

What a partner can check today, without buying anything

• Open your own homepage and read the <title> tag. Is it your current name? Does it say where you are and what you do?
• Open yourdomain.com/robots.txt. Does it name AI crawlers, and if so, what does it say to them?
• Ask any AI assistant the question a buyer would ask about your specialty, without naming your firm. Then ask it to name you. The first test is discovery; the second is recall. Only one resembles what a prospect does.
• Search your own site for "$750,000."

What this study does not show

It does not show that appearing in an AI answer produces clients. It does not show that any page, title or listing causes a recommendation. It is ten questions, five platforms, two dates, one region, reported as counts. AIWorthy measures and reports. The panel, questions and scoring were frozen before any firm was contacted, and results publish regardless. Any measured firm can obtain its own record and 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. Nothing here is a recommendation.

---

Summary

Ten Triangle accounting firms, each named as auditor on three or more Single Audit filings in the federal record, were measured against the questions their buyers ask — once on 11 August 2026 and again on 28 September, with the same ten questions, the same five platforms, and the same five resolved model strings.

What the models remembered did not change. What the web told them did.

• 79 of 250 responses in August, and 85 of 250 in September, named a firm that no longer operates under that name. In August all 79 came from the three platforms that answered without retrieving sources. In September, five came from Google AI Mode — one response returned five times — and the source it cited for the recommendation was the firm's own website, whose homepage title still carries the name the firm dropped in April 2024. A second panel firm was named fifteen times under a predecessor name retired in a 2023 merger and once under its own. All hand-read.
• The platforms answering from stored knowledge replicated to within a handful of responses: summing each panel firm's change on each platform, ChatGPT moved 5 namings, Claude 4, Gemini 7. The platforms that retrieve moved 13 (Perplexity) and 15 (Google AI Mode): one panel firm lost every Perplexity and Google mention it had, another gained a Google recommendation, and the most-read page in the study changed from a national firm's Raleigh office page to a local firm's nonprofit industry page. The pages that stopped being read were still there; the retrieval layer stopped choosing them. Google AI Mode, which answers identically five times on a date, answered every one of the ten questions differently six weeks later, from a different source list.
• The directory layer roughly doubled its share of citations. The mandatory federal record was consulted twice in 1,443 citations in August and twice in 1,282 in September. The state licensing register: never.
• The second most-recommended panel firm — 78 namings across the two dates — was never read at its own site, which on 28 September was a maintenance notice at every address. The platforms that read found it in a former client's archive of old audit reports and in state files; the platform that names it most holds an accurate sketch of a municipal auditor and applies it to nonprofit specialties the record does not show.
• Three panel firms — including the panel's largest filer — were named zero times and read zero times on both dates. A technical read of their sites found one with no website at the domain it files under; one filing from a dead domain whose real site, elsewhere, is open and never read; and one with 1,821 pages and no place name on any page read. On both reachable sites the nonprofit and audit pages name no place.

---

1 · Panel, instrument, collection — what was held constant

Panel. Ten firms from the Federal Audit Clearinghouse AY2023–AY2025 files, screen and roster as published on the record page — every firm meeting the screen after the pre-registered exclusion of five national firms, which are tracked instead. Wherever the study says "complete population," that qualifier applies. The ten firms are named as auditor on 75 Triangle Single Audit engagements between them, ranging from 3 to 13 per firm.

Instrument. Version 1.2: ten questions, one general and nine specialty, frozen 11 August. Five platforms — ChatGPT, Claude, Perplexity, Gemini, Google AI Mode — five asks per question, independent API calls, no conversation history, explicit sampling parameters. 250 events per collection.

Advance notice, and what it may have changed. On 19 August 2026, between the two collections, thirteen measured firms — the ten panel firms and three non-panel firms named in the August findings — were sent their complete records with a right of reply. One address bounced (a partner on the federal filings no longer at the firm). No firm replied. No correction was requested. The web was measured again on 28 September, so the letters are an intervention between the collections, and the study says so: any recipient could have changed its site after reading what had surfaced about it. No recipient's site shows a change the study could attribute to the letter; the firm whose page became the most-cited in September received no letter. Two further courtesy letters, on findings that emerged after the second collection, were sent on 30 September with a reply window closing 14 October.

---

2 · Panel firms named

2.1 · Totals, 250 responses each

Panel namings: 208 in August, 180 in September. Seven of the ten panel firms were named at least once in August; six in September. Sorted from most to least, the ten firms' counts were 100, 44, 21, 18, 16, 8, 1, 0, 0, 0 in August and 97, 34, 20, 12, 10, 7, 0, 0, 0, 0 in September. Every firm named in August was named the same number of times or fewer in September; the largest drop was ten, the smallest zero.

The most-named firm accounted for 100 of 208 panel namings in August and 97 of 180 in September. The three firms named zero times on both dates include the panel's largest filer and its second-largest.

Filed engagements against namings: the four panel firms with the most Single Audit engagements (45 of the panel's 75) were named 24 times in August and 17 in September, combined; the six with the fewest (30 of 75) were named 184 and 163.

Not scored: fifteen mentions of the pre-2023 name of one panel firm's Triangle practice (§4).

2.2 · By platform (each platform 50 responses per date)

The three platforms that answered without retrieving sources moved by 4 to 7 namings across all ten firms; the two that retrieve moved by 13 and 15. On the memory platforms the most-named firm held exactly — 19 → 19, 22 → 22, 36 → 36 of 50. On Google AI Mode, whose five asks return one answer, every change is one response replacing another: the most-named firm lost one question (10 → 5), the second-most-named lost one (5 → 0), a third gained one (0 → 5), and the firm renamed in 2024 gained one under its old name (0 → 5).

2.3 · Own site cited (citations, both dates)

Citations to the ten panel firms' own websites: 102 in August, 65 in September. Seven panel sites were cited at least once across the two dates; three were cited zero times on both, and those three are the three firms never named.

One panel firm was named 78 times across the two dates with its website never cited once. One was read more in September than in August and named zero times under its current name on both dates; §4 explains why.

---

3 · Routes into an answer, on both dates

In August, being read was close to necessary on the retrieving platforms and irrelevant on the memory platforms. September confirms that and adds the second half: the retrieval layer re-weighted its sources wholesale — away from firm homepages and the nonprofit sector's own roster, toward commercial directories and one local firm's specialty page — and the recommendations followed the sources. The memory platforms did not notice, because nothing they rely on changed.

Three routes are used throughout: route one, a firm is named in a response that cites the firm's own site; route two, a firm is named with no citation at all, from the model's stored knowledge; route three, a firm is named in a response whose citations are someone else's pages about it.

Figures for firms outside the instrument's tracked list: citation counts are exact on both dates, from the citation record. Naming counts for these firms were hand counts in August; for September only the most-cited firm's were hand-read; the other ten firms' September naming figures are pattern-match candidates and are labeled as such or omitted where the sentence does not need them.

3.1 · Route one — recommended because the site was read

The September firm. A Raleigh firm with fewer Single Audit clients than any panel firm, not on the panel, saw its site go from 55 citations to 103 — the most-cited domain in the study — and its namings from 45 (hand count) to 65 (hand-read 29 September: 66 responses mention the firm; 65 name it as a firm to consider, one carries the site only in its citation list), with 54 of the 65 in responses that cite the site. Of the eleven that do not, Google's five answers on one question link the site in the body without listing it as a source; four are ChatGPT from memory — under three different renderings of the firm's name — one is Gemini from memory, one is Perplexity via a professional profile. The retrieving platforms, reading the site, use the firm's actual name every time. The page doing the work is the firm's nonprofit industry page (40 citations), which displaced a national firm's Raleigh office page as the most-read page in the corpus. The firm leads the domain list on five of ten questions: general discovery, behavioral health, legal services, small charity and statewide coalition. Between the two dates, Perplexity and Google came to treat that one industry page as the reference for "Triangle nonprofit audit firm."

The page itself did not change; its address did. In August it was cited 25 times at one URL; on 28 September that URL returns a permanent redirect to a new address in an /industries/ section — the same page under the same title, which names the service ("Nonprofit CPA Services"), the city and the state. The site was reorganized into an /industries/ section between the collections, and the retrieval layer read more of it afterward: besides the nonprofit page (25 → 40), the section index, a healthcare page, a private-companies page and a bookkeeping page — none cited in August — drew 26 citations between them. The firm received no notice from this study between the collections. The site's own sitemap dates every section page's last modification to 22–28 September — the nonprofit page to 28 September, the day of the second collection — so the site was being edited through the day it was measured, and the reorganization is bracketed (between 11 August and 28 September) rather than dated; a cluster of dates that tight is consistent with a site-wide republish. Nearly every image on the site carries a place in its alt text: a site managed for place-based search, recently and deliberately. Observation; no claim that the reorganization, the editing or the alt text produced the citations.

A second Raleigh firm outside the panel held the pattern but lost volume: 59 → 45 citations, 39 pattern-match candidates with 34 co-occurring. Same page as August — a Raleigh audits-and-reviews services page — cited 31 times.

The charter-school firm, outside the panel, is the charter-school firm on both dates: 25 of its 46 citations are on the charter-school question, where its site leads the domain list, and 21 of 23 candidate namings came with the site read. A firm named almost exclusively when a specific specialty is asked and its specialty page is retrieved.

The blog-post firm — the one panel firm on this route — was read less (37 → 25) and named less (16 → 10), and all ten September namings co-occur with its site being read. Its blog post listing North Carolina nonprofit audit firms fell from 34 to 20 citations. All ten namings are Google, which returned the same response five times on each of two questions: two Google answers and zero from every other platform. The firm is named when its blog post is retrieved and not otherwise.

The Perplexity-only firm — a panel firm among the four largest filers — is Perplexity-only on both dates (8 → 7). Its site was read more in September than August (15 → 20 citations across the three domains it operates; correction 5), on the charter-school and municipal questions rather than the nonprofit ones. The firm's own site is the sole route in, and being read more did not produce more namings.

Five firms outside the panel were all read less. One went from 30 citations and 21 namings to 1 and 0; another from 15 citations to 1; a third — an outsourced-accounting provider that does not perform audits — from 19 to 0. These firms were in the study only because a page of theirs was retrieved. When the retrieval layer stopped reading them, they vanished from the answers — and their pages did not vanish from the web (§12.6). That is the cleanest demonstration in the corpus of what route one means.

3.2 · Route two — recommended from stored knowledge

The most-named panel firm — 100 → 97, one of the panel's three smallest filers. Its site went from 29 citations to 10, and only 3 of 89 candidate namings co-occurred with the site being read. Claude 22 → 22, Gemini 36 → 36, ChatGPT 19 → 19. The firm is a memory phenomenon: the models know the name, and the web is nearly irrelevant to whether it is recommended.

The memory firm — 44 → 34, zero citations to its site on either date. Claude 32 → 28 accounts for almost all of it, and Claude cites nothing. The picture Claude holds is specific and largely accurate — the town the firm sits in, the county, local government, schools, housing, "mid-size" — and it applies that picture to questions the public record does not support: the firm is named on the behavioral-health, legal-services and child-and-family questions in five of five Claude asks each, on both dates. Gemini's four namings, all on the municipal question, are the most accurate description of the firm anywhere in the 500 answers ("located directly in Wake County … one of the most active auditors for small-to-midsize towns").

The platforms that read found the firm too — 13 times in August (Perplexity 5, Google 5, plus Gemini's cited answers), 3 in September (Perplexity) — and every one of those readings was of somebody else's document. The most-cited source on the firm's responses in August was the State Treasurer's local-government audit files (17 citations). The source Perplexity and Google shared, on the charter-school question on both dates, was a charter school's archive of its audited financial statements, labeled by auditor — the firm for fiscal 2018 through 2022, a different firm for 2023, 2024 and 2025. The school changed auditors three years ago. In September 2026 Perplexity recommended the firm for a charter school Single Audit on the strength of that page, adding that the firm "say[s] they have audited many charter schools since 1998 and perform audits under Government Auditing Standards and OMB Circular A-133" — a circular superseded by the Uniform Guidance in December 2014. The sentence is attributed by Perplexity to the archive page; it does not appear in that page's HTML, and the audit reports linked from it are scanned images. Its origin was not established, and the study reports the sentence as Perplexity's attribution and nothing more.

Google's organic results for the firm's name on 28 September are the same trail: two employees' LinkedIn profiles, a city council's 2020 minutes, a school board's 2010 minutes, three pages of directory rankings. Google's knowledge panel calls the firm a "tax preparation service" and links to the maintenance page (§12.2).

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. That record is accurate about the past and silent about the present. Observation only.

Two Gemini-and-Claude panel firms (21 → 20 and 18 → 12) had sites that were read a handful of times in August and not at all in September (11 → 0 and 5 → 0). Their namings held anyway.

The hybrid national firm — 188 → 185 — is the exception to the routes. Read heavily (67 → 44; its Raleigh office page is still the second most-read page) and also known: ChatGPT 49 → 50 and Gemini 46 → 48 with no citation; Perplexity and Google with. It is the only firm in the study on which the memory platforms and the retrieval platforms agree. On the municipal question it is named in 25 of 25 responses on both dates.

Four regional firms were August's clean route-two examples: named, never read. Three stayed pure — 27 → 16, 19 → 15, 11 → 8, zero citations throughout. The fourth complicates it: 8 citations in September, four to a job listing on an aggregator and the rest on the statewide-coalition question, and Gemini 22 → 34. Gemini cites on 7 of 50 responses, so this move could be either side; it is not reused as the route-two illustration.

The remembered regional firm is the reverse move: 84 → 78 namings held, while its site went from 56 citations (20 co-occurring namings) to 10 (5). Gemini 48 → 50 carries it now. A firm on both routes in August is on the memory route alone in September, and its count barely noticed.

3.3 · Route three — carried by a third party

This is where the six weeks show most. The commercial directory layer roughly doubled: the largest directory 29 → 63 (leading the domain list on the legal-services question and appearing on six others), a second 23 → 43, LinkedIn 30 → 41, and a business journal's list of the region's largest accounting firms 10 → 43. The state nonprofit center's member directory — the one roster that is actually a sector membership list, and August's second most-read page at 42 — halved to 21. The State Treasurer's audit-firm list held at 30, almost entirely on the municipal question.

The regulator: the Federal Audit Clearinghouse was cited twice on each date; the State Board of CPA Examiners never; the AICPA's Single Audit page four times in August and never in September (hand-read 29 September). Across 2,725 citations on two dates, the record that names the auditor on every engagement in the study was consulted four times.

No firm has been traced to the directory growth. The merged national firm — formed in 2022 from two firms, both of whose old names the models still use (§4) — was named 49 times in August: Gemini 36, Claude 8, Perplexity 5, and only 9 of the 49 carried any citation. A memory firm. Its rise to 58 in September is Perplexity 5 → 18, and what Perplexity read was the firm's own site: its Raleigh office page (13 citations, on the general question and two nonprofit questions) and one industry page (5, on the behavioral-health question). The business journal list appears only on that question, beside the firm's own page. A memory firm acquired a route-one presence in six weeks through a place page and a service page. (This firm has eleven Triangle Single Audit clients in the federal record and was excluded from the panel by the pre-registered national-firm rule, not by absence.) A third national firm went the other way, 72 → 48, as Perplexity dropped it 17 → 6 and Google's one response stopped naming it; its Raleigh office page was cited 25 times in August and fell out of the top pages in September.

Recommended on the evidence of a job posting. In August no job posting appeared among the sources (one employer-ratings page and one careers page did). In September two postings were cited nine times, each as the stated basis for naming a firm. On the charter-school question, Perplexity wrote in all five asks that the most-named panel firm was recommended because "a job posting for the firm states experience auditing charter schools and nonprofits," citing a listing for a senior audit associate on a technology jobs board; in two asks it added that the posting "is less specific about North Carolina charter-school clients," and named the firm anyway. On another question, Google AI Mode cited a regional firm's "Tax Manager, Not-for-Profit — Raleigh" listing on a jobs aggregator in four of five asks, naming the firm in each. A help-wanted ad is a firm's own statement about the work it does, published somewhere the retrieval layer reads. Fetched on the evening of 28 September, the first listing returned 404 — the page Perplexity cited five times that afternoon as its evidence no longer existed by the time a buyer could have followed it. The second listing was still live. Whose posting the first listing was cannot now be verified from the page; the attribution to the panel firm is Perplexity's own, in its own words.

3.4 · Recommended from the wrong source

The firm renamed in 2024 — 62 → 61 under its old name — is recommended on both dates by ChatGPT, Gemini and Claude from memory, with no source. In September it is also recommended by Google, and the source is the firm's own current domain, serving pages still titled with the old name, plus a LinkedIn company page never renamed. The firm's actual name was used zero times on either date, though its domain was cited 5 and 10 times. The firm is being read; it is being recommended under the name its own pages still carry. Detail in §4.

The merged national firm's predecessor — 97 → 89 mentions of one of the two 2022 names, roughly a third presenting it as current on both dates, all from memory platforms. Stable because the models are stable.

3.5 · On no route at all

The panel's largest filer, its second-largest, and the firm that merged in 2023 (one ChatGPT naming in August, none in September): zero citations on either date. Their sites were not in the retrievable record on the questions asked, and no model holds them. The same for the most prolific nonprofit auditor in the federal record, outside the panel: zero and zero, twice. §12 reads each of the four sites — the merged firm at its actual domain, not the dead domain in its filings (correction 8). None of the four has a predecessor name in the federal record that a model might be naming instead, other than the 2023 case in §4; the largest filer's history was checked auditee by auditee.

3.6 · What the sources say

On the retrieving platforms, the answer to "which Triangle firms do Single Audits" is assembled from one local firm's nonprofit page, one national firm's Raleigh office page, a Raleigh services page, a blog post, and a growing stack of commercial directories and one business-journal list — with the federal record consulted twice and the licensing register never. When those sources shift, the firms shift with them: one local firm up, two others gone, a national firm read for the first time, another national firm down, a regional firm unread but remembered. On the memory platforms none of that matters: the same names come back in nearly the same counts, including two that no longer exist, because nothing the models rely on changed between 11 August and 28 September.

Reported as observation. No claim is made that a page, a listing, or a directory entry causes a recommendation.

---

4 · The stale name, now with its source cited

Defunct firm presented as current: 79 of 250 (Aug) → 85 of 250 (Sept). ChatGPT 43 → 45, Gemini 31 → 28, Claude 5 → 7, Google AI Mode 0 → 5, Perplexity 0 → 0.

The merged national firm's 2022 predecessor as current: 33 → 35 (of 97 → 89 mentions; the rest note the merger). The 88 September passages were read on 29 September: 35 as current, all ChatGPT; Gemini's 41 and Claude's 10 every one merger-noted. In one ChatGPT answer the two predecessor firms are listed as separate firms to consider, one after the other — the two halves of the merged firm, under the names they gave up in 2022.

The panel firm renamed in 2024 as current: 60 → 61. The 61 were hand-read on 29 September: every one a numbered or bulleted recommendation with no "formerly," "renamed" or "merged" within reach, and the current name absent from all of them; ChatGPT 21, Gemini 28, Claude 7, Google 5. In one Claude answer the firm sits two lines from a correct "(now [merged national firm])" note on the 2022 merger — the model knows one firm's merger and not the other's rename; in one Gemini answer the firm is recommended first, two paragraphs after a correct statement that the threshold rose to $1,000,000. Ten of ten questions, both dates. August's "retrieving platforms: 0 of 100" is "5 of 100" in September.

Google's five are one response on the Durham child-and-family-health question, returned five times. It recommends the firm by its old name, link text pointing at the domain the firm adopted with its new name in April 2024. The sources it cites are the firm's own: an audit-and-assurance page titled with the old name; a team page describing a principal as part of "the [old name] team since 2005"; and a LinkedIn company page still under the old name. The domain was cited 5 times in August and 10 in September; the firm was named under its current name zero times on either date.

Google had cited the domain on the same question in August, when its answer drew on twelve domains and named two other firms. In September the same question's source list narrowed to five — two domains kept, three added, among them the LinkedIn page — and the answer named the old name (§6). The domain was read on both dates; the old-name recommendation arrived with the narrower list and the LinkedIn page.

On 28 September the homepage <title> of the firm's domain read "[Old name] - Trusted Advisors for Over 50 Years." The audit-assurance page Google cited is 690 words, titled with the old name, and names no place; the town in Google's recommendation comes from the LinkedIn snippet it cites, which carries the firm's street address. Read in full on 29 September: the firm is the first of four bullets, offered to engage with a "why consider them" endorsement; the current name appears nowhere in the response. The federal filings show the rename itself: the old domain on the older submissions, the new domain from AY2024.

The domain was renamed. The homepage title, the page titles, the team biography and the LinkedIn page were not. A retrieving platform read the firm's own site and recommended it by the name its own site still uses.

A second predecessor name in the panel (found and read 29 September). A panel firm headquartered outside North Carolina acquired a Durham firm founded in 2001 by merger on 29 June 2023 (reported in the trade press and in the acquiring firm's own announcement); the panel firm's Triangle office is the former Durham firm, and both file under the same email domain. The Durham firm's name appears in 7 August responses and 8 September responses — Gemini, Claude, ChatGPT and, in August, Perplexity — on the charter-school, HUD-housing, behavioral-health and Durham child-and-family questions. The panel firm was named under its own name once in August and never in September. The instrument's alias list carried only the 2024 rename as a former name, so these fifteen mentions were not scored; they were found by text search after the firm's domain correction (correction 8). All fifteen passages were read on 29 September: fifteen of fifteen present the predecessor as a current firm to engage — "Located right in Durham," "Best for," numbered recommendations with specialties attached — and none mentions the successor. Thirteen come from the memory platforms; the two Perplexity mentions (August, charter-school question) cite the State Treasurer's audit-fee data, which still lists the predecessor as auditor of a Durham charter school — the public record carrying the retired name, as a client's archive did in §3.2. ChatGPT's September behavioral-health answer lists both of the panel's retired names in the same numbered list. The predecessor's domain answers only a site-builder placeholder with the domain name as its title; the old firm's pages are gone.

The two stale-name cases in the panel run on different mechanisms: the 2024 name survives on the successor's own current pages and its LinkedIn page, and a retrieving platform read them; the 2023 name survives in the models' training and in the State Treasurer's audit-fee data, with nothing of the firm's own left to read.

---

5 · Retrieval and memory — the table, both dates

---

6 · Google AI Mode: deterministic within a date, rebuilt between them

August: five identical responses on all ten questions. September: five identical on nine; one question returned two. Google's effective sample is eleven responses. Every Google change in §2.2 is one response replacing another. Google also began emitting its own wrapper URLs in reference lists (30 in September, 0 in August); these are excluded from all citation totals.

The stability is within a date only. A per-question comparison of Google's August and September answers found no question that returned the same text on both dates and no question that cited the same set of domains. Counting each domain once per question, August's ten answers drew on 83 domain-question pairs; September's on 39. Eight of the ten September answers are shorter.

Google wrapper URLs excluded. Domains counted once per question.

What survived: the September firm's site kept on five questions and the second Raleigh firm's on four — the two route-one sites of §3.1 — plus one or two holdovers on five other questions. Two questions kept nothing.

What Google stopped reading: the site of the firm that went from 30 citations to 1 (cited on four questions in August, none in September — its route in, and out), the remembered regional firm's site (six questions → one), the hybrid national firm's (four → one), two other firm sites (three → none each), the third national firm's site and the largest directory (two → none each). What it began reading: the misspelled-URL page of §8, on four questions, and, on the charter-school and municipal questions, state sources it had not cited in August — the State Auditor, the Department of Public Instruction, the university's School of Government. On the municipal question Google named no panel firm on either date and cited local governments and out-of-region firms.

A firm's presence in a Google answer on a given day is stable across five asks and can be checked with one. Its presence six weeks later rested on a source list that was rebuilt in the interval. Observation on two dates; no claim that the rebuild followed any rule.

6.2 · AI Mode is not the organic ranking for the question — one day's comparison

On 28 September, the ten instrument prompts were submitted verbatim to Google's classic organic search (Raleigh location, top ten) and the results compared with the URLs AI Mode cited that day for the same prompts. Of 49 URLs AI Mode cited, 3 appear in the organic top ten for the same words, all on one question; at the domain level, 5 of 49. On seven of ten questions the two lists share no domain.

The organic lists explain why. Given a 200-character buyer question, classic search keyword-matched its way to pages with nothing to do with it: a dictionary, a federal health institute, a nuclear regulator, a Midwestern city, Facebook, another state's treasurer. Only on one question did the organic top ten consist of relevant Triangle pages — and that is the question where AI Mode cited three of them. AI Mode's citations were on-topic Triangle audit firms, state agencies and the nonprofit center on every question.

Google describes AI Mode as decomposing a question into multiple searches ("query fan-out") and drawing on their results; the pattern here is consistent with that account and is not a test of it. What the comparison does establish: ranking for the buyer's question as the buyer would ask it is not the route into the AI answer, because the organic ranking for that question is mostly noise. The same day, three buyer-style keyword searches per question — "nonprofit audit firm Raleigh NC," "charter school auditor North Carolina," "municipal audit firm North Carolina" and the like, thirty in all — were compared with the same AI Mode citations. Of 49 cited URLs, 7 rank in the organic top ten for any of the three searches for their question; 11 of 39 domains. Twenty-two of the thirty searches share no domain with AI Mode's citations for the same question. The exception is the general question, where one search ("accounting firms Raleigh nonprofit single audit") returned all four of AI Mode's cited pages; no other search in the set approached that.

Forty-two of the 49 pages AI Mode read were not on page one for any of three plausible searches a buyer might run. Ranking for the keyword is not what puts a page in the AI answer, and being in the AI answer does not indicate a page ranks. What the pages read have in common (§12.6, §9.1) is the kind of page classic search optimization would build — reachable, place and service on the page, on the firm's own site — but which of those pages are read is decided by a process that does not reproduce the ranking. Caveats: the thirty queries are the study's choice; one day; one location setting; the searches Google's fan-out actually generates are not observable, and the general-question result suggests at least one of them resembled a query chosen here. The verbatim organic call for the CDFI question returned an API error and was rerun the next day.

---

7 · The stale threshold

August, hand-read: 26 responses stated $750,000 as the current Single Audit threshold, unqualified; 15 from Google (three distinct responses). September, hand-read 29 September: 37 candidates (responses stating $750,000 and never mentioning $1,000,000), of which 22 state it as current without qualification — Google AI Mode 11 (three distinct responses), Claude 5 (four of them "currently $750,000"), ChatGPT 3, Gemini 2, Perplexity 1 (reported as what a cited page "notes," with "an annual A-133 audit" alongside; counted, flagged as attributed). The remaining 15 are all Google: five, on the general question, say "historically $750,000" — the figure is marked as past, and no current figure is given; ten, on two other questions, carry the figure only in the snippet of a cited page shown beneath the answer, not in the answer itself.

Google's change between the dates is the finding inside the finding. In August its answers stated the old figure flat, fifteen times. In September, on the general question, the same platform writes "historically $750,000" — it knows the number is old and does not say what replaced it — and on two other questions the stale number has left the answer and sits in the source snippets underneath. Fewer stale statements; the same stale sources.

The sources: on all three September Google responses that state the figure as current, the citation list includes the second Raleigh firm's audits page (states $750,000 once; last modified August 2023, §12.8); on two of the three it includes the blog-post firm's post listing North Carolina nonprofit audit firms (states it three times; updated May 2026), and on one question also that firm's post on nonprofit audit requirements. The reference-only cases show a third page carrying it: the misspelled-URL firm's "Does My Nonprofit Need an Audit?" page. The blog post is among the most-read pages in the study on both dates.

Rule applied, as in August: counts when the response states $750,000 as the current trigger without qualification; does not count when the figure is marked as superseded ("historically," "previously," "before October 2024") or appears only in a cited page's snippet.

---

8 · The pages that were read

Most-read pages:

The most-read page in August belonged to a national firm's Raleigh office. In September it is a local firm's nonprofit industry page; that firm — fewer Single Audit clients than any panel firm, not on the panel — went from 55 citations to 103, the most of any domain. The fourth most-read page in September has the city misspelled in its URL. Retrieval did not mind.

Reported as observation. No claim is made that a page shape produces a recommendation.

---

9 · Where the retrieval layer looks

9.1 · Every citation, by source type

All 2,725 citations on the two dates (Google's own wrapper URLs excluded by rule) were classified by domain into fourteen source types — an explicit rule table plus a hand-classified list of residual domains; the rule table and both dates' classified records are part of the frozen artifacts. The residue left unclassified is 1.5% in August and 1.1% in September.

Four observations from the table, none of them a cause.

What was read was overwhelmingly the firms' own pages — two-thirds of August citations, half of September's — and government and client documents about the firms. Between them, firms' sites, government files, client organizations and the nonprofit databases account for 80% of August citations and 67% of September's. The retrieval layer reads the public record and the firms' own descriptions of themselves.

The sources the current AI-visibility advice emphasizes were barely read or not read at all. Not one of 2,725 citations was to Reddit, Quora, or any forum or Q&A site. Not one was to Wikipedia. "Social" is LinkedIn company and profile pages, nothing else. "News" is, on both dates, mostly one business-journal ranking of the largest accounting firms; there is no news coverage of any Triangle audit firm anywhere in the record. Whatever earned media, community discussion and encyclopedic presence do for a brand, they did nothing measurable here.

The directory layer is one platform's habit. Every September directory citation is Perplexity's (203 of 203); Google AI Mode cited a directory 15 times in August and never in September. The doubling in §3.3 is a Perplexity doubling.

Gemini, when it cites at all, cites only official sources — the State Treasurer, the Federal Audit Clearinghouse, a sector association — six and seven citations across fifty responses.

9.2 · The regulators and the rosters

The mandatory federal record naming the auditor on every engagement was consulted twice on each date. The state licensing register, never. The four commercial directories and the business journal went from 92 citations to 190; the one sector membership roster halved.

---

10 · The wrong country, and the wrong continent

August: 49 citations to 14 United Kingdom domains, all on the Durham question (Durham, NC → County Durham), all Perplexity. September: 20 across 4 UK domains, same question, same platform; Google's answer to the same question now cites Triangle sources only. New in September: 44 citations to an Indian business directory and 10 to an Indian business newspaper on the CDFI question — Perplexity went to India. Both confirmed by reading the responses: the UK citations are the same three UK pages on each of the five Durham asks, one of them a UK-government investment directory; the Indian ones are directory listings for accounting services in three Indian cities on each of the five CDFI asks.

---

11 · General versus specialty

Responses naming at least one panel firm:

(August specialty corrected from 152; correction 1.)

11.1 · Tracked firms

Eleven firms outside the panel were tracked by the instrument: the five national firms excluded from the panel by rule, and six regional firms with Triangle Single Audit clients that fell below the screen or outside the three counties.

Moves of ten or more are retrieval-side except the regional firm on Gemini (22 → 34); see §3.2. The merged national firm's own site, unread in August, drew 18 citations in September — its Raleigh office page and one industry page (§3.3). The remembered regional firm's site went from 56 citations to 10 while its naming held at 78 — the "named without being read" state, now on a firm that was read heavily six weeks earlier.

Firms outside the tracked list that appeared in the answers (eleven): citation counts are exact on both dates; naming counts are hand counts in August and, for September, hand-read for the September firm only — the other ten are pattern-match candidates, labeled wherever they appear.

---

12 · What a machine could reach

On 28 September, after the second collection, twelve firms' sites were read against a checklist written down before any site was opened: the six panel firms whose results the comparison contrasts most sharply, five non-panel firms that appear in the citation record, and the one firm outside the panel with the most Single Audit clients in the federal record. A thirteenth site — the actual site of a panel firm whose filed domain is dead — was read on 29 September (correction 8). The checklist follows the conditions the method states must hold before a machine can describe a firm accurately — reach, read, identify, retrieve — and each item is a yes, a no, or a count. Domains were taken from the firms' federal filings.

The read establishes what was retrievable. It does not establish why any firm was or was not recommended, and no causal claim is made.

12.1 · One panel firm cannot be opened at all; another files from a domain that cannot

The no-website firm — the panel's second-largest filer — files every engagement from a domain that has nameservers at Microsoft 365, a mail exchanger at outlook.com, an SPF record, and no web address record of any kind. Email at the domain works; there is no website at it, on any scheme, with or without www. Nothing exists for any platform to read. Zero citations and zero namings on both dates is the only possible result.

The firm that merged in 2023 — headquartered outside North Carolina, with a Durham office — files every engagement from a domain that resolves, redirects to HTTPS, and then fails the TLS handshake for every client that tried: TLS 1.3, TLS 1.2, with and without a server name; the www host answers 403 on plain HTTP. On 28 September the study took the filed domain to be the firm's site. It is not. Google's knowledge panel for the firm points to a different domain, which answers every user agent in 0.18 s, has no AI-crawler rules, carries Organization schema and lists 371 pages. The federal record names a door that does not open; the firm's actual door is under another address, and §12.3 reads it.

The no-website firm is the one panel firm for which "not in the retrievable record" is literal. It is not a positioning problem, a content problem, or a directory problem.

12.2 · One panel firm's homepage is a maintenance page

The memory firm — named 78 times across the two collections, every one from a model's memory. On 28 September its homepage <title> read "Our Site is Undergoing Maintenance" and the server delivered 173 characters of visible text. The server also returned HTTP 429 to the GPTBot user agent. The sitemap lists 119 URLs, none matching a Single Audit or nonprofit term — and two sampled post URLs, fetched separately, returned the same maintenance title. The whole site, not the homepage alone, is one maintenance notice. Whether the site was in this state on 11 August is not known: the Internet Archive has no snapshot near that date, and its last archived working homepage is from July 2025. A firm the models know by name, with nothing of its own on the web for a retriever to confirm it by; what the retrievers found instead is in §3.2.

12.3 · Two panel firms are fully reachable and still never read

The largest filer is the opposite case. The site answers every user agent in 0.19 s, has 1,821 URLs in its sitemap, at least 25 of them matching Single Audit or nonprofit terms, and carries Organization schema. Nothing about reach or read explains zero citations.

What the site says about itself does. All 25 sitemap matches are articles — a sales-tax note for another state, a filing notice for a third, PPP, ARPA, provider relief funds. The firm's nonprofit industry page and its audit-and-assurance page both exist and both answer 200; neither is in the sitemap. The homepage title is "Public Accounting Firm | Professional Services Firm | [firm]." The nonprofit industry page is 862 words, uses "nonprofit" 20 times, and names a place zero times; so does the audit page; so does the Single Audit article. There is no Raleigh location page at the expected address. The content that would answer a Triangle buyer's question exists on the site, and nothing on it says Triangle.

The federal record makes the gap concrete. This firm's Triangle auditees include three Raleigh-area behavioral-health organizations. The instrument asks a behavioral-health question. The firm was named on it zero times in ten responses across the two dates. The record also rules out one explanation: the firm has audited these organizations under its own name in every year the files cover, back to AY2021, so there is no predecessor name for a model to remember instead. (Three auditees moved to other firms in AY2024; none moved to this firm from another name.)

The firm that merged in 2023, at its actual domain (read 29 September), is the second case, and a closer one, because its homepage does mention both "Durham" and "nonprofit." The homepage is 765 words; its seven place mentions come from a news slider announcing a new Durham managing partner. The nonprofit industry page is 383 words and names a place zero times. The nonprofit-and-foundations assurance page is 337 words, uses "Single Audit" five times and "Uniform Guidance" twice, and names a place zero times. The one page that names Durham eleven times is a 516-word news item about a partner appointment. Service pages without a place, a place page without the service, nothing over 800 words. The firm's title tag names the firm and "Assurance, Tax & Advisory Services" — no place. One naming in 500 responses (ChatGPT, August), zero citations, both dates. Its filed domain is dead (§12.1), so the record itself could not have led a machine here; but the site was in the retrievable web, and nothing retrieved it. What the models do hold is the name of the Durham practice before the firm acquired it: that name appears fifteen times across the two dates (§4). Reported as an observation about what was retrieved on two dates; whether that is the reason is not tested here.

12.4 · The most prolific auditor in the record has opted out of two indexes

The most prolific nonprofit auditor in the record — just outside the three counties, with more distinct Single Audit clients than any panel firm — publishes a robots.txt built as an allow-list: a named set of agents may crawl, and everyone else receives Disallow: /. The named set includes Googlebot, Bingbot, Google-Extended and the user-triggered fetchers ChatGPT-User, OAI-SearchBot, Perplexity-User and Claude-SearchBot. It does not include GPTBot, PerplexityBot, ClaudeBot or CCBot — the crawlers that build the indexes those platforms retrieve from. The file also sets a five-second crawl delay and its sitemap lists no pages. The homepage title names the firm's city and state and "Accounting Firm."

The format is a site-builder template. The effect is that Perplexity's index and OpenAI's cannot contain the site, Google's can, and Google never surfaced it for a Triangle question. Zero citations, zero namings, twice.

12.5 · The firms that were read: what their front doors say

The two firms whose titles name a place and a service — the September firm and the blog-post firm — are the two route-one firms whose pages carried the most citations in the study. The most-named panel firm presents itself to a machine by a three-letter abbreviation the models do not use; it is named from memory regardless. When a retrieving platform did read the site, the abbreviation cost the firm its name: one Google AI Mode answer recommends the firm by its abbreviation followed by a parenthetical expansion into three surnames, none of them the firm's — it cites the firm's own governmental-and-affordable-housing page and quotes its street address correctly, and invents the partners (found in the verification spot check, 29 September). Right firm, right page, right address, fabricated name: a third kind of naming failure beside the two stale names in §4. The blog-post firm's post, the page behind every one of its namings, is not linked from the homepage. The firm read 30 times in August and once in September has no sitemap and no structured data. The charter-school firm's structured data declares the firm an Event and an OutletStore.

All twelve sites answered a browser, GPTBot and PerplexityBot identically except the memory firm (429 to GPTBot) and the two unreachable filed domains; the 2023-merger firm's actual site, read the next day, answered all three identically. No site set noindex. No site's robots.txt blocked AI crawlers by rule except the allow-list in §12.4; the Perplexity-only firm's file names thirty AI agents but applies to them only the platform's standard housekeeping exclusions. Two of the twelve sites publish an llms.txt, the file the current advice proposes for AI readers (checked 29 September): the largest filer's is hand-written — it describes the firm and names North Carolina among its states — and the firm was read zero times on both dates; the blog-post firm's was generated by its SEO plugin and lists the sitemap. The renamed firm's site answers unknown paths with a Cloudflare challenge page (403) to scripted readers; its real pages answered 200 on 28 September.

Speed. PageSpeed Insights was run on 28 September against up to three sitemap-sampled pages per reachable domain (plus the homepage where a site has no sitemap), mobile and desktop — site-level figures, not the cited pages themselves. Lab figures are Lighthouse's throttled simulation; field figures are Google's CrUX 75th percentile for real Chrome users, present only where the origin has enough traffic.

Speed does not separate the read from the unread. The most-cited domain in the study has the lowest lab mobile scores in the set; the never-cited firm outside the panel and the most-named panel firm are among the fastest. Every origin with field data is rated FAST on both TTFB and LCP, including the two the lab rates slowest. Whatever the retrieval layer selected on, it was not this.

The memory firm's scores are those of a maintenance page: the sampled post URLs returned LCP equal to FCP and zero blocking time, and a title check on two of them confirmed they serve the same notice as the homepage (§12.2).

12.6 · What was on the pages the machines read, the pages they dropped, and the pages they never read

The citation record names the pages. Twenty were read against the same signals: the seven most-cited pages across the two corpora; six pages that were cited in August and not in September; and the equivalent pages of a panel firm never cited, plus three homepages. Three pages of the 2023-merger firm's actual site were added on 29 September. Signals: visible word count, uses of the study's own terms, Triangle place-name mentions, and whether the page states a federal threshold figure. Server HTML only; no rendering; observation only.

First, the dropped pages are still there. Fourteen URLs from five firms whose citations fell to one or zero were fetched on 28 September; thirteen answered 200 at the same address. The one exception timed out — a reach failure, consistent with that firm's 15 → 1. Nothing was removed. The retrieval layer stopped choosing pages that remained available to it.

Four things are visible in the table and none of them is a cause.

Every page cited on both dates names the Triangle, between 11 and 32 times; every page of the largest filer names it zero times. The like-for-like pair is the September firm's nonprofit industry page against the largest filer's: comparable length, comparable density of "nonprofit," 21 place names against 0, cited 40 times against never.

A place name is not enough on its own. The outsourced-accounting provider's Raleigh–Durham location page names the Triangle 30 times — more than any cited page but one — and went from being read to not being read. The third national firm's Raleigh office page names it 15 times and fell from 25 citations to none. What the dropped pages lack is elsewhere in the table. The location page is about CFO and accounting services, not audit; the provider does not perform audits, so its August citations for audit questions were, if anything, the retrieval layer's error, and its September absence the correction. The national firm's office page is 409 words about tax and wealth. The two local firms' nonprofit pages name the place and the service, at four hundred words each. The 3,571-word audit guide has the service and the depth and no place. The pages that were read on both dates each have three things: a place, the buyer's own service, and depth. Each dropped page has at most two.

The pages read for Single Audit questions barely say "Single Audit." 0, 0, 0, 1, 2, 6 across the six most-cited. The largest filer's article says it 6 times and was never retrieved. Every question in the instrument describes a situation — an organization of a certain kind, in a certain place, needing an audit — and what was retrieved matched the situation. The term itself was close to irrelevant.

The page most responsible for one panel firm's visibility states the superseded threshold three times. The blog-post firm's post is the one page that carries every one of its namings; it is also among the sources cited on every Google response that repeats the $750,000 figure (§7).

Three further signals the current advice names were read from the same saved pages on 29 September: a named author, question-phrased headings, and whether the first eighty words name both the service and the place. A byline appears on one of the seven pages cited on both dates and on four of the never-cited or dropped pages. Question headings appear on almost no cited page (0, 0, 1, 0, 2, 0, 0); the page that follows that advice best — ten question headings, 3,571 words — is the audit guide that was cited in August and dropped in September. Service and place in the opening eighty words: six of seven cited pages, four of six dropped pages, none of the largest filer's four, and only the news item among the 2023-merger firm's four. The opening-words property is shared by what was read and by some of what stopped being read; it is absent from the two reachable firms that were never read.

Twenty pages on one date, and three more the next day. The three-of-three pattern is a description of this table, not a rule; the three rows added 29 September fit it (service without place, under 400 words) but were not part of the original read. One further caveat: the September firm's nonprofit page was last modified, by its own sitemap, on 28 September — the day it was read for this table and the day of the second collection. The words counted here are the page as it stood that evening; the version the retrieval layer indexed before citing it 40 times is not known to be identical.

12.7 · Structured data on every page the machines read

Every distinct URL cited on either date (383, Google wrappers excluded) and the six never-cited control pages were fetched on 28 September and read for structured data — JSON-LD and microdata types, FAQPage schema, and any FAQ-style heading. 265 of the 383 cited pages could be read; the 118 that could not are mostly directory pages that return 403 to scripted readers and are listed in the frozen output.

Structured data is common on the pages that were read and more common on the pages that were not. It varies with the site builder — WordPress themes emit Organization/WebPage/BreadcrumbList; HubSpot sites emit AccountingService; the nonprofit center's directory page (57 citations), the third national firm's office page (31) and the State Treasurer's list (25) carry none — and not with whether the page was read. FAQPage schema, which the current advice singles out, is on 11 of 265 cited pages and on none of the twenty most-cited. Observation on one date; no claim about what schema does or does not do elsewhere.

12.8 · How old the cited pages are

Every cited URL was fetched again on 29 September and read for a content date — JSON-LD dateModified/datePublished and the article meta tags. HTTP Last-Modified headers were collected but excluded from the figures: on dynamic sites they report when the page was generated, not when its content changed, and half of them read "today." 101 of 383 cited pages carry a content date; the rest are homepages, directory listings, PDFs and government pages without one.

The retrieval layer leans toward recently modified pages: three in five dated pages had been changed within six months of the collection, and the most-cited ones more recently still. This is the one item of the current advice — freshness — that the record supports directly.

The oldest pages read are also among the most-read. The second-most-cited page in the study, the second Raleigh firm's audits page (71 citations across the two dates), was last modified on 15 August 2023 and states the superseded $750,000 threshold. The blog-post firm's post, the fourth-most-cited, was updated in May 2026 and states it three times. A page's date and its facts' dates are different things; what the retrieval layer reads is the date. Observation on one fetch; a page's dateModified is what its publisher's system recorded, and 282 cited pages carry no content date at all.

12.9 · The population: every auditor domain in the federal record

The same reach check was run on 29 September against every distinct auditor email domain in the Triangle filings — 57 domains after de-duplicating 91 name variants; three filers gave no usable domain. Two caveats govern the table. The domain is the one the firm typed into its federal filing, not necessarily where its site lives; and the check's "firm name in title" test proved defective (it fails on any firm whose title omits its legal suffix), so that signal is not reported.

Three things the population shows that twelve sites could not.

The federal record's auditor domains are stale. Of the five filed domains with no reachable site, three belong to firms whose site lives at another address — one panel firm, one regional firm and one national firm — and one belongs to a firm with no site at all. The record names the auditor reliably; where to find the auditor, it does not. A retrieval system that trusted the filing would arrive at a closed door for four firms in 57.

About a third of the record's firms exclude at least one AI crawler on their filed domain. Ten of 57 publish an allow-list robots file that disallows every agent it does not name — among them two Big Four firms, two panel firms and the most prolific auditor in the record — and nine answer GPTBot with a 403, 406, 429 or a challenge page while answering a browser normally, among them one panel firm and two national firms. Most of the allow-lists are site-builder templates rather than decisions; the effect on a crawler is the same.

The never-named firms are harder to read, but not cleanly so. Sixteen of 25 never-named domains carry a flag against 14 of 32 named; twelve of 25 have no sitemap against six of 32. But the flags are not a wall: one panel firm with an allow-list robots file and no sitemap was named 21 and 20 times from the memory platforms — and its own-site citations went from 11 to 0 between the dates. Observation on one fetch per domain; no claim that any flag caused any count.

12.10 · What the read does not show

• Not that a title, a place name, a sitemap, a schema type or a recent date produces a recommendation, or that its absence prevents one. The correspondences in §12.5–12.9 are observations on one or two dates.
• Not the state of any site or page on 11 August. The read was taken after the second collection.
• Not rendering, and not the speed of the cited pages themselves: the PageSpeed figures are site-level samples.

---

13 · Verification

Every figure in this study is one of two kinds: reproduced by the frozen scorer and scripts from the frozen corpora, or hand-read. Before publication, a written gate of seventeen items was closed, each with a decision rule stated in advance. The reading items, all closed 29 September:

• The Google response recommending the 2024-renamed firm under its old name, read in full: confirmed on all four points; the town in the recommendation traced to the cited LinkedIn snippet; no mention of the current name.
• All 61 September mentions of the old name: 61 of 61 presented as current, no qualifier, no mention of the current name.
• All 88 September mentions of the merged national firm's predecessor: 35 as current, all ChatGPT; 53 merger-noted.
• All 37 September threshold candidates: 22 unqualified; 5 "historically"; 10 in reference snippets only.
• Three panel-firm mentions per platform, fifteen in all, checked against the raw response: 15 of 15 are namings in the answer body; one names the firm by abbreviation with an invented expansion (§12.5).
• The UK and India citations: confirmed on all five asks of the relevant question on each date.
• September's two regulator citations: both the Federal Audit Clearinghouse.
• The most-cited non-panel firm's 66 September mentions: 65 name it as a firm to consider; 54 cite its site.
• The fifteen predecessor-name mentions of the 2023-merger firm: 15 of 15 present it as current; none mentions the successor.

Items closed as untraced or bracketed rather than resolved: the origin of Perplexity's "since 1998 … A-133" sentence (§3.2), reported as the platform's attribution only; the exact date of the September firm's site reorganization, bracketed between the two collections by the site's own sitemap; and whose job posting the deleted listing was (§3.3), stated as Perplexity's attribution only.

Two items remain open on publication and will be closed on the record page: the reply log for the two courtesy letters sent 30 September (window closes 14 October), and the re-issue of one panel firm's record with a corrected own-site count (correction 5).

13.1 · Corrections made during verification

Dated, in order. Each correction is also recorded on the record page.

1. 18 August specialty count. The August figure for responses naming a panel firm on specialty questions was computed before two scorer corrections were applied; on the corrected scorer it is 145 of 225, not 152. General question 9 of 25 unchanged. (28 September)
2. Regulator citations, August. "FAC + State Board + AICPA combined: 2" undercounted: FAC 2, State Board 0, AICPA 4. (28 September)
3. Place+service rule written down. The August hand figure, 289 citations on 30 pages (20.0%), was replaced by a stated, scripted rule: 304 on 35 pages (21.1%). The same rule is applied to both dates. The most-cited pages are unchanged. (28 September)
4. August defunct-firm classifier. A defect that dropped firm names inside markdown links was corrected; August figures re-derived on the corrected script were unchanged (79 of 250). (28 September)
5. One panel firm's own-site count. The firm operates three domains, all resolving to one site; the August count used one of them. Own site cited is 15 → 20, not 13 → 10; the direction of change reverses. The firm's August record carried 13 and will be re-issued. (28 September)
6. Working-draft misattribution. A working draft described the merged national firm as the beneficiary of the directory growth. Its August namings were from memory; its September rise came from its own site being read for the first time. Corrected before publication. A first pass at this trace merged the two dates' scoring files under colliding keys; caught and rerun before anything was written. (28 September)
7. Structured data on the most-cited page. A working note that the most-cited page carried no structured data was wrong: it carries an AccountingService type in microdata, missed by a JSON-LD-only read. (28 September)
8. One panel firm's website. The study took the firm's filed domain to be its website and reported the site as unreachable. The filed domain is dead; the firm's actual site is at another address, reachable and 371 pages deep. The technical read was corrected and the firm moved from "unreachable" to "reachable and never read." The earlier statement that only one panel firm's domain differed from its filing was also wrong: the check compared the study's list to the filing, and both carried the dead domain. (29 September)
9. Second predecessor name. The domain correction led to the discovery that the same firm acquired a Durham firm in 2023, and that the predecessor's name appears fifteen times in the corpora. The instrument's alias list did not include it; the fifteen mentions are reported as found by text search and were not scored. This finding is post-hoc and is labeled as such. (29 September)
10. Population census signal dropped. The "firm name in title" test in the 57-domain census compares the filed legal name, suffix included, to the page title and fails on any title that omits the suffix. Not reported. (29 September)
11. Page age. HTTP Last-Modified headers were excluded from the page-age figures as server generation times, not content dates. (29 September)

13.2 · What was pre-registered and what was not

Pre-registered before the first collection: the panel screen and roster, the ten questions, the five platforms, the sampling parameters, the scorer's matching rules and the alias list, the hand-verification gate, and the decision to re-collect once with the instrument unchanged. Added after the first collection and labeled as such: the second collection date; the technical read of firm sites; the citation source-type census; the organic and keyword comparisons; the schema and page-age passes; the 57-domain population census; the source traces in §3.1–3.3; the second predecessor name. Everything added after the fact is an observation on the frozen corpora and the live web as it stood on 28–29 September, and is labeled post-hoc on the record page.

---

What this does not show

• Not that any platform improved or declined. Two dates, six weeks apart, five asks each.
• Not that the memory platforms will hold. They held because the models did not change.
• Not that a page shape or a directory listing causes a recommendation.
• Not a rate. Counts over stated denominators on stated dates.
• Not which firm is which. Results are not attributed to firms; the roster is on the record page.

---

Independence and disclosure

The panel, the questions and the scoring were frozen before any firm was contacted, and the results publish regardless of what they show. All thirteen measured firms notified in August, and the two notified in September, received their complete records with a right of reply before publication; none requested a correction. Any measured firm can purchase its own record or technical read from AIWorthy; buying one 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.

---

Frozen artifacts

Every figure in this study derives from the files below or from a hand read recorded in the verification gate. Each directory carries a SHA256 manifest of its files; the hash given is of that manifest.

The frozen artifacts name firms, pages and domains. They are retained for verification and are available to a measured firm for its own record, and to a reviewer on request under the terms on the record page.