Clarion Studio  ·  Primary Research  ·  2026

The Unverified Profession

162 US law firms, 125 buyer questions, three AI and search surfaces. Two in five firms never appeared once. And where the assistants did answer, nothing the firms publish let them check what they were saying.
A 162-firm audit of independent and mid-market US law firms  ·  40–250 attorneys  ·  Three platforms, six instruments  ·  September 2026

Clarion Studio  ·  clarion.studio
40.7%
of firms never appeared once across 125 questions of the form a buyer actually asks, on any of the three surfaces measured.
0
of 153 firms link an attorney to a state bar admission record. The one fact the profession rests on is not machine-readable anywhere.
Missing from the answer, or wrong in it.

The Short Version

Type your own firm's name into an AI assistant. You will get an answer, and it will be fluent, immediate and specific. It will name partners. It will say where each of them is admitted to practise. It will not hedge.

Whether that answer is reliable is one question, and this study measured it. But there is a prior problem with the test you have just run. Your buyers are not asking it. They do not know your name, and not knowing it is the reason they are asking at all.

That second question is the one nobody has measured. Not tell me about this firm, but which are the best environmental law firms in Alabama. We built 125 of them, one for each state and practice combination in our sample, and ran every one on ChatGPT, on Perplexity and on Google.

The two questions turn out to be almost unrelated. On branded questions, the answers are assembled from the firm's own website, Wikipedia, Reddit, wire copy, and a legal-explainer aggregator most firms have never heard of; the well-known ranking directories appear for 8.4 per cent of firms. On buyer questions, the directories are close to the whole answer, present in 96.7 per cent of ChatGPT responses, 99.2 per cent of Perplexity's, and 97.6 per cent of Google's organic top ten.

Asked the question that describes exactly what it does and where it does it, the average firm in this study surfaces roughly one time in three. Sixty-six firms of 162, 40.7 per cent, did not appear once, on any platform, across all 125 questions.

The obvious response is to buy more directory listings. We tested that too, on all 162 firms, and it does not work: 92.6 per cent of the sample already holds a Super Lawyers profile, the average firm holds 3.3 of the seven directory profiles we could measure, and the number a firm holds predicts its visibility at a correlation of +0.02, which is another way of writing zero.

Underneath all of it sits the finding that gives this study its name. Every US state bar publishes a free, searchable, authoritative record of who is licensed to practise. Not one firm in this study connects a lawyer to that record in a form a machine can read.

So when an assistant states where a partner is admitted, it is not reading a record. There is no record for it to read. It navigates the way a pilot does once the instruments that fix position are gone: take the last known heading, apply the time elapsed, state where you are. Dead reckoning is disciplined and it is usually close. It also drifts, and it gives no sign of the moment it began to.

Asked the same question three times, the assistant named rosters of attorneys that overlapped at a median of 0.46. For 5.9 per cent of the lawyers it named more than once, it stated different bar admissions between runs. One was placed in one state, then a second, then the first again, inside a few minutes, in answer to the same words.


0%
link an attorney to a bar admission record
97.6%
of buyer questions answered through a directory
+0.02
predictive value of holding those directory listings

Key Findings at a Glance

MetricFinding
Buyer questions whose ChatGPT answer cites a ranking directory96.7%
Same, Perplexity99.2%
Same, Google organic top ten97.6%
Buyer questions where the first organic result is a directory67.2%
Median rank of the first directory / first law firm's own site1 / 2
Firms appearing on none of the three surfaces, across 125 questions40.7%
Firms surfacing on their own state-and-practice question, Google28.6%
Firms already holding a Super Lawyers profile92.6%
Correlation, number of directory profiles to discovery visibility+0.02
Correlation, breadth of directory ranking to AI visibility+0.28
The same correlation measured against Google+0.06
Correlation, machine-readable identity score to visibility (AI / Google)−0.00 / +0.04
Buyer questions returning a Google map pack87.2%
Map-pack positions held by firms in this sample0.6%
Firms linking an attorney to a state bar admission record0%
Mean machine-readable identity score (of 5)0.82
Firms scoring 0 or 1 of 590.2%
Firms whose own domain is absent from their top ten sources29.0%
Share of own-site citations earned by insights, alerts and posts45.1%
Share earned by attorney bios and practice-area pages combined8.0%
Median agreement between three identical live queries0.46
Attorneys whose stated bar admissions changed between runs5.9%

What This Means in Plain Terms

Two different assets, with two different gatekeepers

A firm's branded visibility is what an assistant says when somebody types the firm's name. Its discovery visibility is whether the firm appears at all when somebody describes the problem instead of the firm. They are built from different sources, they respond to different things, and a firm can be strong on one and absent from the other.

Almost every report a law firm receives about AI search measures the first kind, because the first kind is easy to measure: you know the firm's name, so you ask about it. The second kind requires building the buyer's question from scratch, in every state and practice area the firm operates in, and running it. That is what this study did.

What a machine can and cannot confirm about your firm

Every US state bar publishes a searchable directory of admitted attorneys with licence status. It is the authoritative record of who may practise and where. It is public, it is free, and it is the single fact on which the profession's authority rests.

A machine cannot get from your website to that record. Not because the record is hidden, but because nothing on your site points at it in a form a machine reads. Structured data is how a website states a fact in a way software can act on rather than merely display: this organisation, this person, this credential, this jurisdiction, and here is where you can confirm it. Of the 153 firms whose sites we could read, three point at any authoritative legal source at all, and in every case it is a commercial ranking directory rather than a regulator.

So when an assistant is asked where your partner is admitted, it does what it can. It reads prose, infers, and states a conclusion in the same confident register it uses for everything else. Sometimes it is right. Sometimes it is right on Tuesday and wrong on Wednesday.

Why the obvious fixes are not fixes

Once you know the directories mediate the buyer's question, buying more directory listings looks like the answer. It is not, and the reason is arithmetic rather than opinion. Listings are already close to universal in this market, so they cannot be what separates one firm from another. We measured that across all 162 firms and found no relationship at all.

Structured data looks like the other obvious answer, and this study is in part about its absence. We tested it as a predictor on both surfaces and found nothing. We publish that below, in full, because a study that only reports the results its author would like is not a study.


Methodology

The sample frame is 162 independent and mid-market US law firms of 40 to 250 attorneys, constructed across eleven regional and practice segments and fourteen normalised practice groups. Firms operating the distributed or non-equity model form a deliberate eleventh segment. Each firm's attorney count is sourced to that firm's own published page; one count could not be verified and is marked as such in the dataset. Am Law 100 firms are out of scope. The frame targets the size band in which an independent firm actually competes.

Six instruments were run, all producing observations rather than estimates.

The discovery instrument is 125 unbranded questions of the form best [practice group] law firms in [state], one for each state-and-practice cell represented in the sample, plus three national questions for the distributed-model segment. The set covers 161 of the 162 firms. Every question was put to ChatGPT, to Perplexity with web search, and to Google, and every cited source, organic result and map-pack entry was recorded.

Robustness was measured on a 30-cell subsample: three phrasings of each question to test whether the results are an artefact of wording, and three repeats of each to measure run-to-run variance. 180 runs in total.

Branded citation counts were drawn from a commercial AI-answer corpus covering ChatGPT and Google AI Overview, measured on identical domains in the same market and language. The figure counts distinct AI answers in which a firm's domain appears as a cited source. It measures citation selection, not whether the cited page shaped the answer.

Directory presence and standing were measured per firm. Presence: the linking profile each firm holds on each of the major legal directories. Standing: the complete published ranking table for every firm on the largest of them: every tier, in every practice area, in every metropolitan market.

Structured data was extracted by loading each firm's homepage in a JavaScript-rendering crawler and reading JSON-LD out of the live document object model rather than the raw HTML, so that client-side-injected schema was captured. Nine sites returned HTTP 403 to an automated crawler and could not be audited; the schema figures therefore describe 153 firms. A second pass followed the people index through to an individual attorney biography on a stratified subsample of 33 firms, of which 22 yielded a readable bio page.

Fidelity was measured by live prompting. Twelve firms, spanning the full visibility range, were each put to a live assistant with web search enabled, using one question per firm and running that identical question three times.

Controls. Organic traffic, link authority and attorney headcount were pulled for all 162 domains so that every reported relationship could be tested against firm size and search strength rather than reported raw.


Finding 1Branded Questions and Buyer Questions Run Through Different Worlds

The position repeated across the legal marketing industry through 2026 is that a handful of directories govern which firms AI assistants cite. Measured against branded questions, that is not what happens. Across the 155 firms with branded citation data, the six commercial ranking directories appear in the source set of thirteen firms - 8.4 per cent. What sits between those firms and the answers given about them is something else entirely:

SourceFirms it appears forShare
Wikipedia9963.9%
legalclarity.org4931.6%
Reddit4025.8%
Reuters3220.6%
JD Supra1912.3%
Cornell LII / Justia case law16 each10.3%

legalclarity.org is a legal-explainer aggregator. It sits in the source set of nearly a third of the firms in this study, and it is not a publication any of them has a relationship with.

Then we asked the other question. Not tell me about this firm, but the question a client with a problem actually types, best insurance defence law firms in Missouri, across 125 state-and-practice combinations, on three surfaces.

SurfaceAnswers citing at least one ranking directory
ChatGPT96.7%
Perplexity99.2%
Google, organic top ten97.6%

On Google, the first organic result is a directory on 67.2 per cent of these questions. The median rank of the first directory result is 1. The median rank of the first actual law firm's own website is 2. Not one of the 125 result sets was free of directories, aggregators or legal-recruitment sites; the median question gave them four of the ten organic positions.

Two of those intermediaries appear only on Google and are invisible to both AI platforms: a legal-recruitment site present on 58 per cent of these questions, and a careers-ranking publisher on 24 per cent. More than half of all buyer-discovery results put a lawyer-recruitment property in the top ten.

What this means. Branded visibility and discovery visibility are separate assets with separate intermediaries. Both figures above are true, and they are not about the same thing. Any measurement that does not say which question it asked is not reporting a position - it is reporting an average of two unrelated ones.


Finding 2Two in Five Firms Are Missing From the Question That Describes Them

For every firm in the frame we built the single most relevant question that exists for it: its own state, its own primary practice group. If a firm is going to be found by description rather than by name, this is the question that finds it.

Result
Firms surfacing on their own state-and-practice question (Google)28.6%
Questions on which any sampled firm appears in Google organic58.4%
Firms surfacing at least once across all 125 questions, ChatGPT46 of 162
Same, Perplexity66 of 162
Same, Google61 of 162
Firms surfacing on all three surfaces20 of 162
Firms surfacing on none of the three, across all 125 questions66 of 162 (40.7%)

The three surfaces also disagree about which firms exist. Eleven firms appear on Google and neither AI platform. Twelve appear on ChatGPT alone. Sixteen on Perplexity alone.

There is a further surface almost nobody in this market is looking at. 87.2 per cent of these buyer questions return a Google map pack: 321 positions across the 125 questions, held by 292 distinct firms. The 162 firms in this sample hold two of them: 0.6 per cent. That surface is fed by Google Business Profile rather than by anything on a firm's website, and it is a different population of firms entirely.

What this means. A firm can hold a strong brand, a well-reviewed website and a healthy search footprint, and still not be in the room when a buyer describes the matter rather than naming the firm. This is a description of a state, measured on one set of questions at one point in time. It is not a claim about what any individual firm's clients do.


Finding 3Directory Listings Are Near-Universal, and Predict Nothing

Finding 1 produces an obvious conclusion: if the directories mediate the buyer's question, buy your way onto the directories. We tested it directly, on all 162 firms.

Almost every firm is already there.

DirectoryFirms in the sample holding a profile
Super Lawyers92.6%
Justia67.9%
Chambers61.7%
Best Lawyers56.2%
Best Law Firms29.0%
Avvo9.9%
Legal 5009.3%

The average firm holds profiles on 3.3 of the seven directories this method can measure. The median is three. Exactly one firm of 162 holds none.

And the number a firm holds tells you nothing about whether it gets found. The rank correlation between directory profile count and appearing in any discovery answer is +0.02 across 162 firms, which is indistinguishable from no relationship at all. Firms holding three or more profiles fail to surface anywhere 41.4 per cent of the time. Firms holding fewer than three fail 39.1 per cent of the time. Twenty-five firms hold four or more directory profiles and appear nowhere; fifty-seven of the sixty-six wholly invisible firms hold a Super Lawyers profile.

What this means. Being listed is a threshold this market crossed some time ago. It cannot be the differentiator, because almost everybody has it. Any proposal that treats directory listings as the lever on AI visibility is selling something this dataset does not support, and the arithmetic is easy for anyone to repeat.

A related observation, offered as such: eighteen firms of 162 are ranked by a major directory and have never claimed the profile that ranking entitles them to, which leaves the directory holding a page about them that they do not control and have never seen.


Finding 4The AI Platforms and Google Are Playing by Different Rules

If presence does not move, the next question is what does. We tested every variable in the study against discovery visibility on each surface separately. The result is a clean split.

Variablevs AI platformsvs Google
Breadth of directory ranking+0.28+0.06
Link authority+0.28+0.28
Organic traffic value+0.30+0.23
Attorney headcount+0.21+0.04
Machine-readable identity score−0.00+0.04
Directory profile count+0.02+0.02

Google's discovery results are predicted by classical link authority and search traffic, and by nothing else we measured. The AI platforms respond to those too, and additionally to how broadly a firm is ranked inside the directories - how many practice areas and markets it holds a tier in, rather than whether it holds a profile.

That relationship is not simply firm size. Holding attorney headcount constant it remains at +0.22. Run the test the other way, holding ranking breadth constant instead of headcount, and the size effect disappears entirely, at +0.10 and not statistically distinguishable from zero. Sorted into quartiles by ranking breadth, 28.9 per cent of firms in the bottom quartile surface on an AI platform at all, against 75.0 per cent of the top quartile.

The limits of this one, stated plainly. Once organic traffic and link authority are also held constant the relationship weakens to +0.16, and a bootstrap confidence interval around that figure touches zero. Twelve relationships were screened in this analysis, so that figure does not survive a strict multiple-comparison correction. The form of the result that does hold up under full controls is the simpler one, whether a firm appears at all rather than how often, at +0.21. We report the direction and the gap between the quartiles. We do not report an effect size, and nothing here is causal.

What this means. A firm's AI discovery position is not a by-product of its search position, and the two respond to different inputs. A provider reporting on search health is not reporting on this, and would have no reason to know it was missing.


Finding 5The Profession Is Not Machine-Readable

Across 153 crawlable firms:

SignalFirmsShare
Organization, LegalService or LocalBusiness type present10669.3%
No structured data of any kind on the homepage2214.4%
Jurisdiction bound in machine-readable form95.9%
Any Person or Attorney entity63.9%
A link to any authoritative legal record32.0%
A question-and-answer page marked up as one10.7%
A link to a state bar admission record00.0%

Scored across five dimensions - organisation identity, attorney identity, jurisdiction binding, question-and-answer structure, and a verifiable external link - the mean is 0.82 of 5. The median is 1. The maximum reached by any firm is 3. Forty-six firms score zero and ninety-two score one, so ninety per cent of the profession sits in two adjacent buckets.

Three firms in the entire sample link themselves to any authoritative legal source, and in every case it is a commercial ranking directory, never a regulator.

A second pass went through to individual attorney biographies on 22 firms. Four carry attorney-level structured data. One links to a bar association, which is a membership body rather than a licence record. None connects a named lawyer to their own admission record.

And it does not predict anything. We tested the rubric score against discovery visibility on both surfaces: −0.00 against the AI platforms, +0.04 against Google. Two ways of writing zero. The honest explanation is at least partly statistical. Where ninety per cent of a sample sits in two adjacent buckets, there is almost no variation for an effect to show up in. But the measurement is what it is, and this is the result.

What this means. The absence is close to total, which makes it unclaimed ground rather than a competitive disadvantage. No firm in this sample is being beaten on this dimension, because no firm is playing on it. That is an observation about position, not a lever on citation, and this study does not support treating it as one.


Finding 6The Pages That Earn Citations Are Not the Pages Firms Build

Eighty-four of 155 firms have any of their own pages among the ten pages most cited alongside them. Across those firms, 112 own-site pages were identified and classified.

Page typePagesShare
Insight, publication or client alert2623.2%
Blog or news post2522.3%
Other content page2118.8%
Homepage or About1715.2%
PDF document119.8%
Attorney biography54.5%
Practice-area page43.6%
Case study or matter page32.7%

Dated substantive writing accounts for 46 per cent of every own-site citation in the sample. Attorney biographies and practice-area pages, which absorb most of a law firm's website budget and nearly all of its partner review time, account for 8 per cent between them. Eleven of the cited pages are PDFs, which suggests the format matters less than whether the document says something specific and dated.

A related figure, counted by domain rather than by page: for 45 firms, 29 per cent of the sample, the firm's own domain does not appear in its top ten sources at all. Everything a machine reads when it discusses those firms was written by somebody else.

What this means. This is an observed distribution across one sample, not a demonstrated mechanism. It is offered as what the cited pages look like, not as a promise about what publishing more of them produces.


Finding 7Three Identical Questions, Three Different Answers

The corpus instrument is deterministic. A byte-identical re-query of ten firms returned identical figures on all ten. So the reproducibility question belongs to live prompting, and that is where it was measured.

Twelve firms. One question each: which attorneys handle this practice area, and in which states is each admitted. Put to a live assistant with web search, three times, identically.

Median overlap of the named-attorney set across the three runs: 0.46. Three firms of twelve returned the same roster all three times. At the other end, one firm's third run named ten attorneys of whom eight had appeared in neither earlier run.

Of 202 attorneys named in at least two of the three runs, twelve (5.9 per cent) had their bar admissions stated differently between runs. In the clearest case a single lawyer was placed in one state, then a second state, then the first state again, within minutes, in response to the same words.

Separately, of 22 attorney-biography URLs the assistant cited as its source for an admission, twenty-one resolve and one does not. The dead one is a fabricated path, presented in exactly the form the twenty-one real citations took.

What this means. A single run of a query like this measures nothing durable, and any report built on one is reporting noise. Beyond that: every US jurisdiction regulates communications about a lawyer's services that are false or misleading. An answer that places a firm's lawyer in a state where she is not admitted is a statement the firm did not write, cannot see, and is not monitoring, and it differs between one reader and the next. This study establishes that it happens at a measurable rate. It does not establish who bears responsibility for it, and nothing here should be read as advice on that question.

This portion rests on twelve firms. The rate has wide error bars and is offered as a preliminary result rather than a settled one.


Finding 8What a Zero Does Not Mean

Seven firms have zero ChatGPT citations in the branded corpus. The tempting reading is that an assistant has never heard of them. That reading is wrong, and this study says so before anyone else has to.

Four of those firms were among the twelve put to a live assistant. It answered about all four, naming attorneys and their admissions, citing bio pages on the firms' own sites that resolve correctly. One of them returned thirty-six attorneys with admissions, identically in all three runs. That was the most reproducible answer produced by any firm in the study, including the two market leaders.

What this means. A zero in this study means a firm is not being surfaced to somebody who does not already know its name. It does not mean the assistant cannot answer when a name is typed in. The finding is about discovery, not existence, and the distinction is the whole of it.


Cross-Cutting Themes

The same result in a second profession

Clarion ran a structurally similar audit of 130 independent registered investment advisers in May 2026. That study found mean schema coverage of 1.93 of 5, and not one firm linking its SEC registration in a form an AI could verify.

The law firm sample scores 0.82 of 5, and not one firm links a bar admission record. Two regulated professions, one instrument, 292 firms, and in both the regulatory link is absent everywhere. That makes it a finding about professional services rather than about either profession.

A citation figure without a platform attached is not a measurement

On branded questions, rank correlation between the two platforms across 162 firms is 0.79. Strong, and incomplete: twelve firms moving two or more quartiles between them. The median firm is cited 8.4 times more often on one than the other, and that multiple runs from 3.2× at the tenth percentile to 26.5× at the ninetieth. Ten firms exceed 40×. Seven are cited on one platform and never on the other, which has no ratio at all.

This reproduces, on an independent 162-firm corpus, the structural conclusion of Zhang, He and Yao (arXiv preprint 2604.25707, April 2026), who measured 602 prompts and 21,143 citations and found platform-level citation behaviour different enough that cross-platform totals are not comparable. That work is a preprint and has not been peer reviewed. Its method is published, which is why it is usable here.

Google's AI Overview is present and unreadable

An AI Overview appeared on 54.4 per cent of the 125 buyer questions. Its contents are not synchronously retrievable through any vendor available to this study, so we can report that it appears and not what it cites. We would rather state the limit than model around it.

Practice area discriminates more than geography

Plaintiff and trial firms are cited at roughly sixteen times the rate of general-litigation defence firms on one platform. Insurance defence and general litigation sit at the bottom of both platforms and account for five of the seven zeros. Intellectual property firms show the widest gap between the two platforms of any sizeable practice group. Regionally, the Mid-Atlantic and the Northeast lead; Texas is the weakest state block on both platforms despite holding several of the largest firms in the frame by headcount.


Conclusions and Outlook

Five things follow from the data, and two do not.

There are two questions, and they have different gatekeepers. On branded questions the intermediaries are reference sites, forums and wire copy; the ranking directories appear for 8.4 per cent of firms. On the buyer's question the directories are present in 96.7, 99.2 and 97.6 per cent of answers on three independent surfaces. A measurement that does not say which question it asked is not describing a position.

Two in five firms are absent from the buyer's question altogether. Sixty-six of 162, across 125 questions, on every surface measured.

Directory listings are not the lever. They are near-universal in this market and their count predicts discovery visibility at +0.02. What travels with AI visibility is the breadth of a firm's ranking inside the directories, and even that is a modest, uncontrolled-for-causation association we report as a direction rather than a number.

The two surfaces respond to different inputs. Google to link authority, the AI platforms additionally to directory standing. A firm cannot read one position off the other.

The machine-readable layer of this profession is empty. Not weak. Empty, on the one dimension - the licence record - that the profession's authority actually rests on. And what an assistant says about a firm's lawyers is unstable: identical questions produced rosters agreeing at a median of 0.46, and one attorney in twenty had their admissions restated differently between runs.

The two things that do not follow. None of this establishes causation, anywhere, in any direction. Every relationship reported here is an association measured at one point in time. And building the missing structured layer is not shown to produce citations. We tested exactly that and found no relationship on either surface.

What can be said is narrower and more useful than either. The discovery layer of this profession is mediated by a small number of properties that firms do not own and mostly do not manage. The verification layer, which would let a machine confirm the one claim a law firm makes that matters, does not exist at all, and a firm that builds it would currently be the only firm in its sample that has.

Until one exists, every statement made about a lawyer in this profession is a position estimated rather than fixed, by a system with no way to take a reading and no reason to doubt itself.


Companion Reading

The companion study, The AI Visibility Gap, applies the same instrument to 130 independent registered investment advisers. Read together, the two documents make the cross-profession case that neither makes alone.

How AI Verifies You're Real teaches the entity graph these findings describe, from first principles, with a live inspector that scores any site against the same rubric used here.

The Mythbusting Register records claims about AI search that do not survive measurement, and gains two entries from this study: that the directories govern AI citation generally, and that holding directory listings improves discovery visibility.


About Clarion Studio

Clarion Studio is a digital infrastructure practice serving high-trust professional services firms. Clarion designs and builds entity graphs, answer-engine architecture and editorial systems that make a firm's actual expertise legible to the layer now mediating discovery.

Clarion's practice is anchored in primary research of the kind documented here, applied to the positioning and infrastructure decisions of individual firms. Clarion is based in the Pacific Northwest and works with a small number of firms each year.


Methodology Note and Disclosures

This study measured 162 independent and mid-market US law firms of 40 to 250 attorneys across eleven regional and practice segments. Attorney counts are sourced to each firm's own published page. One count could not be verified from the firm's own materials and is marked unverified in the dataset.

Branded citation figures are drawn from a commercial AI-answer corpus covering ChatGPT and Google AI Overview, measured in the United States in English in September 2026. The figure is a sampled corpus observation, not a census of what any individual user sees, and AI answers are non-deterministic. It counts citation selection - whether a domain was retrieved and cited - and not citation absorption, meaning whether the cited page shaped the answer. Claude and Copilot were not measured.

Discovery figures rest on 125 unbranded questions run on ChatGPT, on Perplexity with web search, and on Google, in the United States in English in September 2026. One question per state-and-practice cell is a single sample of a large possible question space; a different phrasing set would produce different individual results, which is why a 30-cell phrasing-sensitivity and run-variance subsample was run alongside it. Two ChatGPT questions returned no sources at all and are excluded from that platform's denominator.

Directory presence is measured through the linking relationship between each directory and each firm's site. Three directories (Martindale, Lawyers.com and FindLaw) emit no crawlable links to firm sites and are therefore invisible to this method; we confirmed this by testing two national firms that certainly hold profiles on them. Directory presence figures describe seven directories, not ten, and a firm may hold profiles this method cannot see.

Nine firm websites returned HTTP 403 to an automated crawler and could not be audited for structured data. Those nine skew larger than the sample, so the structured-data figures describe the 153 crawlable firms and should not be read as describing the largest firms in the frame. Attorney-level figures rest on 22 biography pages, one per firm, drawn from a stratified subsample. The fidelity and reproducibility findings rest on twelve firms and 36 live runs.

Every correlation reported here is an association. No causal claim is made or implied anywhere in this document. Twelve relationships were screened in the predictor analysis; where a result does not survive a multiple-comparison correction, that is stated at the point the result is reported.

No firm is named. Aggregate figures are published; individual firm positions are not, and are shared only with the firm concerned.

Verification: three figures were re-read outside their original batch and matched exactly, one of them confirmed by a second independent vendor; the citation corpus was re-queried byte-identically and returned identical figures on all ten firms tested; and every statistic in this report was re-derived a second time from the raw files by an independent code path. Two instrument faults were found during the work and are disclosed here rather than quietly corrected: an earlier version of the directory-standing extractor exceeded a browser memory limit on large firms' pages and truncated their ranking tables, which would have under-counted precisely the largest firms. It was rewritten and all 162 firms re-measured. Separately, an earlier draft of this study reported the branded-question directory figure as a general result; that error was caught before publication and is the reason Finding 1 is written as it is.

The audit was conducted by Clarion Studio using Claude, developed by Anthropic, as the analysis system. Claude was not used as a measurement target, so that no model class evaluated outputs from its own class.

Clarion Studio provides digital infrastructure services to professional services firms. That commercial relationship creates a potential interest in findings that support the value of such investment. No adjustment was made to any score or finding to produce a more dramatic result. Three findings in this report cut against that interest and are published unchanged: that structured identity predicts neither AI nor Google discovery visibility, that directory listings predict nothing, and the demonstration in Finding 8 that uncited firms remain answerable.


References

Primary research on AI citation behaviour

Zhang, K., He, X., & Yao, J. (2026, April 29). From citation selection to citation absorption: a measurement framework for generative engine optimization across AI search platforms. arXiv preprint 2604.25707. Not peer reviewed.

Clarion Studio. (2026, May 27). The AI Visibility Gap: how independent RIAs are being found by AI. https://clarion.studio/research/the-ai-visibility-gap-2026/

Structured data and search documentation

Google Search Central. Introduction to structured data markup in Google Search. https://developers.google.com/search/docs/appearance/structured-data/intro-structured-data

Schema.org. LegalService schema specification. https://schema.org/LegalService

Schema.org. Person schema specification. https://schema.org/Person

Schema.org. EducationalOccupationalCredential schema specification. https://schema.org/EducationalOccupationalCredential

Regulatory records referenced

State bar attorney directories, all fifty states and the District of Columbia. Each publishes a searchable record of admitted attorneys with licence status.

The Unverified Profession  ·  Primary research by Clarion Studio  ·  September 2026  ·  162 firms, 125 buyer questions, three platforms, six instruments

Where does your firm stand?

Every engagement begins with a diagnostic conversation. No pitch. No slides. A structured assessment of where you are today and whether there's a fit.

Start a Conversation