At three in the afternoon on August 9, we asked an AI answer engine a question a grieving family asks every day: "I inherited a painting and need it appraised for estate taxes. Who can do that?" The answer named a practice. Arcadia Art Consultancy, surfaced through a June blog post of theirs on inherited art, alongside a second firm and a cluster of law firms and trust companies. An hour later, at just past four, we asked the identical question again. Same words, same engine, same day. This time: law firms, trust companies, a CPA journal article from 2002, a TurboTax community thread. No appraiser at all.
Two answers, sixty-five minutes apart, disagreeing about whether a profession has any members.
We had been running this exact query all month, because we work with small credentialed practices and wanted to know what the answer layer tells their buyers. On August 5, the answer taught the full apparatus (USPAP compliance, the "qualified appraiser" rule, the IRS dollar thresholds, the Art Advisory Panel) and named no practice to hire. On August 8, same shape: credentialing bodies (ISA, ASA, AAA, and three more), thresholds, Form 706, nobody to call. We were, frankly, preparing to write the obvious essay: the engine teaches your trade and recommends no one. Then run three named Arcadia, run four un-named it, and the obvious essay died in ninety seconds, which is roughly how long it would have taken any reader to falsify it.
What replaced it is stranger and, we think, more important.
Here is the whole dataset, honestly sized: one query, one engine, one professional vertical, four runs across five days, by one operator. N=4. Small enough to hold in one table, large enough to break every simple story about it.
| run | date | named an appraisal practice? | what it named instead |
|---|---|---|---|
| 1 | Aug 5 | No | credentials, IRS thresholds, the Art Advisory Panel; practices only as cited sources |
| 2 | Aug 8 | No | six credentialing bodies, thresholds, Form 706 |
| 3 | Aug 9, ~3pm | Yes — Arcadia Art Consultancy, via its own blog post | plus law firms, trust companies |
| 4 | Aug 9, ~4pm | No | law firms, trust companies, a 2002 CPA journal article, a forum thread |
The unstable part is the recommendation. The stable part is everything else, and the stable part is where the real finding lives.
Across all four runs, without exception, the professional vocabulary was taught completely. Every run explained what a qualified appraiser is. Every run gave the IRS dollar thresholds. Every run described the Art Advisory Panel, the roughly twenty-five-expert body that reviews high-value art on estate returns. A buyer who read any of the four answers walks away educated to a level that used to require a consultation: they know the credential acronyms, the statutory triggers, the documentation to gather, the name of the federal review panel.
And across all four runs, a second thing never varied: no credentialed sole practitioner appeared. Not once. The one appraisal firm that recurred, in three runs of four, is a practice with a large content operation. The adjacent professions, law firms and trust companies and financial advisors, were named more consistently than the profession the question was actually about. The only party in any of these answers who can legally sign the qualified appraisal that Form 706 requires is the party the answers were least able to produce.
The reflex is to file this under "AI killed the click," and the reflex is off by a decade. SparkToro's clickstream analysis, cross-reported by Search Engine Land, puts the numbers plainly: about 45 percent of Google searches already ended without a click in 2016. It was 49 percent in 2019, 60 percent in 2024, and 68 percent in the first four months of 2026. The trend predates AI answers entirely; AI Overviews, which now appear on more than a fifth of searches and cut click-through by nearly 60 percent when present, accelerated a curve that was already two-thirds formed. And the fully conversational interfaces everyone writes think-pieces about were 0.34 percent of searches in the same window: a rounding error, so far, on the behavior that matters.
So the answer engine did not create the un-nameable professional. It inherited a decade-old gap and changed what sits inside it. In 2016, a zero-click search left the buyer with a snippet: a phone number, an address, a definition. In 2026, it leaves them with a complete professional education: the credential taxonomy, the legal thresholds, the federal review process, the vocabulary to sound informed on the phone. What replaced the click is not nothing. It is everything except the introduction.
That is the precise novelty. Not disintermediation in the familiar sense, where a platform inserts itself between buyer and seller and charges rent at the toll booth. As far as we can see from the outside (and we say it as reasoning, not as an audited finding) there is no auction here, no bid for placement, nothing sold. The layer manufactures demand in the economic sense, teaching the buyer exactly what they need and why it matters, and then, on most runs, declines to complete the match. A market maker that takes no fee and makes no market.
Look again at who does get named, and the pattern stops being mysterious.
Arcadia appeared in run three through its own June blog post about inheriting art. The recurring appraisal firm publishes explainers at scale. The law firms that show up in three of four runs publish estate-tax guides; the trust companies publish planning articles; even the 2002 CPA journal piece is, at bottom, a well-structured explanation that has survived on the open web for twenty-four years. The answer layer is not choosing among practitioners. It is synthesizing explanations, and it credits the explainers. Naming is downstream of content production.
The answer layer has separated who explains the work from who does the work, and it can only see the first.
We can put an operator's receipt behind that sentence, hedged as ours. Over the past weeks we sourced roughly two dozen credentialed appraisal practices, real, accredited, USPAP-compliant professionals of exactly the kind the query is about, for a small outreach project. The ones that appear in AI answers have content operations. The ones we had to dig for have thin web surfaces: several publish no email address at all, and four national professional directories returned zero practice domains to a crawler, because member records sit behind JavaScript search forms that no answer engine can read. These practitioners are not merely un-recommended. They are structurally unreadable to the layer doing the recommending. The credential that makes them the right answer lives in a database the answer layer cannot open, while the blog post that makes someone else citable lives on the open web, pre-chewed for synthesis.
None of which is misconduct by anyone. The firms being named published genuinely useful explanations; that is the whole point, and they deserve the citations. The engine synthesizes what it can read; that is its job. Every actor is behaving reasonably, and the sum is a market where the ability to be found has decoupled, almost completely, from the ability to do the work.
Here is the part we found genuinely new, and it is the part our four little runs demonstrate that the entire small-businesses-are-invisible genre misses.
If the answer layer reliably excluded small practitioners, that would be a legible grievance. You could measure it, document it, appeal it, organize around it. "Search for X, my competitor appears, I never do" is a state of the world a professional can photograph. What our runs show instead is that the state will not hold still. A practice absent at four o'clock may have been present at three. There is no ranking to check, no impression count, no console, no notice. The practitioner cannot demonstrate they were left out, because an hour later they might not have been, and an hour after that they are again.
The condition is not invisibility. It is non-determinism about your own existence at the point of sale.
For a developer audience this failure mode is familiar from distributed systems: the outage you cannot reproduce is categorically worse than the one you can, because it resists diagnosis, resists escalation, and resists proof. The un-nameable professional is living inside a flaky test. And unlike the SEO era, whose opacity at least came with instrumentation (rankings you could track, impressions you could count, a console that admitted you existed), the answer layer currently offers the professional no observability surface at all. Our four-run table, assembled by hand with a fixed query and a notebook, is more instrumentation than the average sole practitioner will ever have.
Which suggests, uncomfortably, that the reproducibility discipline we apply to software now applies to market existence. If your findability is a nondeterministic function, a single observation ("I checked, I'm in the answer" or "I checked, I'm not") is exactly as informative as running a flaky test once.
For the professional, or anyone advising one, the mechanism points at the lever with unusual clarity. The layer names whoever wrote the explanation. The buyer arrives pre-taught the vocabulary of your trade from material someone else published. The move is not to buy visibility, because so far there is nothing to buy; it is to become the explainer of record for the questions your buyers actually ask, in plain text, on the open web, where the synthesizing layer can read it. The credential needs a public, parseable surface, because the private database it lives in is unreadable to the only librarian most buyers now consult. This was true in the SEO era as strategy; it is true in the answer era as a precondition for existing.
For the rest of us, the transferable lesson is about measurement. Any claim about what "the AI says" about your product, your company, or your profession is a sample from a distribution, not a fact. Our first two runs supported a clean, quotable, wrong conclusion, and we only found out because we kept running the query after we believed we knew the answer. If a state matters to you and lives inside a nondeterministic system, the observation protocol is the deliverable: fixed query, verbatim, on a schedule, with dates. Four runs cost us a few minutes and falsified our own thesis before a reader could.
And for whoever is building the answer layers: the gap here is one honest feature wide. The buyer is taught the credential; the credential-holder is unreadable; the buyer ends the session educated and unmatched. An answer layer that could read the professions' own directories, or that told a practitioner plainly when and whether they are being surfaced, would convert an anxious, unobservable lottery back into a market.
The family with the painting, meanwhile, asked a fair question and got a graduate seminar. They now know what USPAP means, what the IRS will scrutinize, what a qualified appraiser is, and what form the estate files. The one thing they still do not reliably know is the only thing they asked: a name. The answer layer has read everything about the profession except its members, and on most runs it teaches the entire trade, thresholds and panels and paperwork, of a workforce it cannot see.
· SparkToro, "In 2026, Less than One Third of Google Searches Still Send a Click" — the zero-click series (~45% in 2016, 49% in 2019, 60.45% in 2024, 68.01% in Jan–Apr 2026), AI Overviews on 20%+ of searches with ~60% CTR reduction when present, and AI Mode at 0.34% of searches; Similarweb clickstream panel, US, with the study's own methodology caveats.
· Search Engine Land, "Google zero-click searches reach 68% in early 2026: Study" — independent reporting cross-confirming the figures and the 9.51-point (22.9%) two-year decline in click-producing searches.
· The four-run observation table is our own data: one fixed buyer query, verbatim, run 2026-08-05, 08-08, and twice on 08-09, with the named/un-named results recorded contemporaneously. One query, one engine, one vertical, one operator — N=4, stated as such.
If being findable depends on which sources a retrieval layer reaches this hour, standing needs to live somewhere the retrieval layer does not own. Agent Rating Protocol is a portable reputation record, so a practitioner or an agent carries verifiable standing instead of depending on who a retrieval layer happens to surface that hour.
Verify a record · Verify a record
pip install agent-rating-protocol · npm install agent-rating-protocol