Common AI Visibility Recommendations, Checked Against 2,725 Citations
There are some common best practices for getting a business recommended by AI assistants. We had 2,725 citations from two retrieving platforms, collected six weeks apart, for CPA firms in one region, the Research Triangle, North Carolina. So we checked common recommendations against the record, item by item.
This case file expands Result 7 of Three Doors and a Brick Wall: How AI Assistants Choose CPA Firms
By Heather Harreld · Published October 7, 2026
If you have read anything this year about AI visibility, generative engine optimization or answer engine optimization, you have seen the list. Get mentioned on Reddit. Get a Wikipedia page. Add FAQ sections and FAQ schema markup. Publish on YouTube and LinkedIn. Earn press coverage. Speed up your site. Add an llms.txt file for AI readers. Keep content fresh. Put a named author on everything.
Most of that advice comes from studies of consumer questions asked of consumer chat apps, where Reddit and Wikipedia do appear often. Almost none of it has been tested against a buyer of professional services asking a specific question in a specific place. Our study produced that test as a by-product: two of the five assistants we asked (Perplexity and Google AI Mode) cite their sources on every answer, and across our two collection dates they cited 2,725 web pages. We classified every one of them by type, read the most-cited pages by hand, and ran the standard checks on the sites of the firms that were and were not recommended.
One scoping note first: ChatGPT and Claude in this study answered without searching the web and cited nothing, so everything below concerns what the retrieving assistants read.
| Recommendations | What the record shows | Verdict |
|---|---|---|
| Get mentioned on Reddit, Quora, forums | 0 of 2,725 citations | No support |
| Get a Wikipedia page | 0 citations | No support |
| Publish on YouTube | 0 citations | No support |
| Earn press coverage | No coverage of any Triangle audit firm; "news" is one ranking list | No support |
| Add FAQ sections and FAQ schema | FAQ schema on 4.2% of cited pages, none of the top 20 | No support |
| Add structured data (schema markup) | 69% of cited pages; 100% of never-cited controls | No separation |
| Speed up the site | Most-cited site had the slowest mobile scores | No separation |
| Rank on page one of Google | 3 of 49 AI-cited pages ranked for the question | No separation |
| Add an llms.txt file | On 2 of 12 sites; the hand-written one is on the never-read firm | No separation |
| Put a named author on pages | 1 of 7 cited pages; 4 of the never-cited or dropped | No separation |
| Use question-phrased headings | Almost none on cited pages; the best example was dropped | No separation |
| Keep content fresh | Median cited page 130 days old; 60% under six months | Supported |
Reddit, Quora, forums: 0 of 2,725
Not one citation went to Reddit, Quora, or any forum or question-and-answer site, on either date, on any of the ten questions. The advice to "be part of the conversation on Reddit" may hold for consumer products. For a nonprofit finance director asking who audits charter schools in North Carolina, there is no conversation on Reddit, and the machines did not go looking for one.
Wikipedia: 0. YouTube: 0.
Same result. No Wikipedia citation and no YouTube citation in the record. "Social" was LinkedIn, and only LinkedIn: 30 citations in August and 41 in September, all to company and profile pages. One of those LinkedIn company pages still carries a firm name retired in 2024, and an assistant cited it as evidence for recommending the firm under that name.
Citations to Reddit, Quora or any forum.
Citations to Wikipedia.
Citations to YouTube.
Earned media: one ranking list, no coverage
"News" accounted for 10 citations in August and 73 in September. Of the 73, 43 were a single business journal's ranking of the region's largest accounting firms, and 10 were Indian business press that one assistant wandered into on the community-lender question. There is no news coverage of any Triangle audit firm anywhere in 2,725 citations. Press coverage may do many things for a firm; in this record it did nothing measurable, because there was none to read.
FAQ sections and schema markup: more common on the pages nobody read
We fetched all 383 distinct pages the assistants cited, plus six pages from a firm they never cited, and read each for structured data. Of the 265 cited pages we could read (directories block scripted readers), 69% carried some structured data and 50% carried an organization-type schema. The never-cited control pages: 100% and 83%. The pages nobody read were better marked up than the pages everybody read.
| All cited pages (265) | Cited 10+ times (57) | Never-cited controls (6) | |
|---|---|---|---|
| Any structured data | 69% | 75% | 100% |
| Organization-type schema | 50% | 63% | 83% |
| FAQ schema | 4.2% | 3.5% | 0% |
| Any FAQ heading | 9% | 7% | 0% |
FAQ schema specifically, the item the playbook singles out, was on 11 of 265 cited pages (4.2%) and on none of the twenty most-cited. FAQ-style headings appeared on 9%. The most-cited page in the entire study carries a plain organization schema that its website platform emits by default, nothing more.
What structured data did track was the website builder. WordPress themes emit one set of types; HubSpot sites another; one firm's site declares the firm to be an "Event" and an "OutletStore." None of it tracked whether the page was read.
Page speed: the slowest site was the most cited
We ran Google's PageSpeed test on every reachable firm site, three pages each, mobile and desktop. The most-cited domain in the study had the lowest mobile scores in the set, with a largest-content load time of 14 to 22 seconds in the lab. The firm never cited and the firm most recommended from memory were among the fastest. Every site with real-user field data was rated "fast" by Google's own measure, including the two the lab rated slowest. Whatever the retrieval layer selected on, it was not this.
Ranking on Google: 3 of 49
If AI answers simply reflected search rankings, a firm ranking well would be covered. On September 28 we took the 49 pages Google AI Mode cited that day and checked where each ranked in ordinary Google search, two ways: for the buyer's question typed exactly, and for three shorter keyword searches per question, thirty in all. Three of the 49 were on page one for the exact question. Seven were on page one for any of the thirty keyword searches. Forty-two were on page one for none. Twenty-two of the thirty keyword searches shared no website at all with the AI answer to the same question.
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 gets read is decided by something other than the 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.
llms.txt: present on the least-read firm in the study
The file proposed specifically for AI readers existed on two of the twelve sites we checked. One was generated by an SEO plugin and simply lists the sitemap. The other is hand-written, describes the firm's services and the states it works in, and belongs to the panel firm with the most Single Audit clients, which was cited zero times and recommended zero times on both dates. The one artifact built for the machines sat on the firm the machines never visited.
Named authors and question headings: no separation
A byline appeared on one of the seven pages cited on both dates and on four of the pages that were never cited or were dropped between the dates. Question-phrased headings appeared on almost no cited page; the page that followed that advice most faithfully, with ten question headings in 3,571 words, was cited in August and dropped in September.
Freshness: the one item the record supports
Of the 383 cited pages, 101 carry a publication or modification date in their code. Their median age at collection was 130 days. Three in five had been changed within six months; for the most-cited pages, seven in ten. This is the one piece of the standard playbook the record supports directly, and it comes with a warning. The second-most-cited page in the study was last modified in August 2023 and states a federal threshold that changed in October 2024. The fourth-most-cited was updated in May 2026 and states the same wrong figure three times. The machines read the page's date. They do not read the facts' dates.
| Pages with a date | Median age | Under 6 months | Over 3 years | |
|---|---|---|---|---|
| All cited pages | 101 of 383 | 130 days | 60% | 9% |
| Pages cited 10+ times | 16 | 77 days | 69% | 6% |
Fresh is not the same as right. The second-most-cited page in the study was last modified in August 2023 and states a threshold that changed in October 2024. The fourth-most-cited was updated in May 2026 and states the same wrong figure three times.
What the record shows instead
What the cited pages have in common.
| What the cited pages have in common | The number |
|---|---|
| They are firms' own pages | 67% of August citations, 53% of September's |
| …or government and client documents about the firms | 9–12% government; client pages, 990 databases and the sector roster most of the rest |
| They are reachable | 10 of 57 record domains block unnamed crawlers; 9 answer GPTBot differently; 4 lead nowhere |
| They name the place | Every page cited on both dates names the Triangle, 11 to 32 times |
| They describe the buyer's service, at length | Every page cited on both dates; dropped pages each lack one of place, service or depth |
| They barely use the term "Single Audit" | 0, 0, 0, 1, 2, 6 times across the six most-cited |
| They are recent | 60% of dated pages changed within six months |
They are the firms' own pages, and government and client documents about the firms. Firms' websites were 67% of August citations and 53% of September's. Government files (the State Treasurer's audit data, HUD, council minutes) were another 9% to 12%. Client organizations' own pages, nonprofit databases built from Form 990 filings, and the state nonprofit center's member directory make up most of the rest. Commercial directories doubled from 8% to 16%, entirely on one assistant. The retrieval layer reads the public record and what firms say about themselves.
They are reachable. Of the 57 auditor domains in the federal record for this region, ten publish a robots file that blocks every crawler it does not name, nine answer OpenAI's crawler differently from a browser, and four lead to no site at all. Most of these are website-template defaults nobody chose. The never-recommended firms were flagged more often than the recommended ones, 16 of 25 against 14 of 32, though not cleanly: one firm with an allow-list robots file and no sitemap was still recommended 21 times from memory.
They name the place, describe the buyer's service, and run to length. Every page cited on both dates names the Triangle between 11 and 32 times, describes the buyer's own kind of work, and runs well over a thousand words. The pages dropped between August and September each lacked one of those three. The never-cited firm's nonprofit page, audit page and Single Audit article are comparable in length and name a place zero times. And the pages read for Single Audit questions barely use the term: 0, 0, 0, 1, 2 and 6 times across the six most-cited. The assistants matched the buyer's situation, not the keyword.
They are new, or recently touched. See above, with its warning.
What a firm should take from this
Most of what firms are being sold under the name AI visibility is work the record shows doing nothing here. Before signing a retainer, ask the vendor which items on the list they intend to work on, and what evidence they have that those items move anything for a professional-services buyer asking a specific question.
Then do the unglamorous things the record does support. Check that your site can be reached by every crawler, including the ones your website template may be blocking. Give each service line one page that says where you are, whom you serve and what you do, at length, on your own site. Keep it current, and when you update the date, update the facts. Search your site for "$750,000." None of that is new. It is traditional search work, done properly, which is roughly what the evidence says this is.
What this does not show
Not that any of these properties causes a page to be read; the correspondences are observations on two dates, one region, one service. Not what the consumer chat apps cite for consumer questions, where the Reddit and Wikipedia figures come from. Not Google Business Profile or name-address-phone consistency, which the study did not test beyond one knowledge panel. Not a rate: counts over stated denominators on stated 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