Two axes decide what AI cites for your category. Geography is a third.

Ten published measurements across nine Hawaii categories plus a cross-geo CPA market (Austin TX), read together with a Nashville CPA control. Same 7-AI-tool methodology, same 18-question discipline (hash-locked per geo). Own-site share is one axis. Training-data presence is a second axis. The Austin and Nashville CPA cross-geo data add a third dimension: the training-data axis itself varies by engine and by geography within the same category.

The cross-category table (web-searching engines)

Own-site share across the five web-searching AI tools (Perplexity, ChatGPT search, Gemini grounded, Microsoft Copilot, Google AI Overviews), pooled by citation volume. Each row is the aggregate of that category’s hash-locked runs, recomputed against current cohorts (per-category as-of dates in the measurement-windows note below). Citations and cohorts evolve, so this is a dated snapshot, not a live feed, which is the whole reason measurement is continuous rather than one-time.

40 percentage points separate the top (51%, where Hawaii banking and Honolulu med spas tie) from Hawaii hotels at the bottom (11%). The bottom of this table is a different kind of bottom than the rest. Where Nashville CPA at 31% still puts a third of its citations on firm-owned sites, Hawaii hotels get only 11% and Honolulu real estate 14%, because the aggregators (Booking and Expedia for hotels, Zillow and Realtor.com for real estate) and the ranking and review platforms have so thoroughly captured those categories that the web-searching engines treat them as the answer. Those two are the most aggregator-dominated categories in the set, where the own-site axis is low not because the businesses have weak sites but because a few platforms own the demand. Banking still leads the way it did in the first measurement: two decades of product pages make a bank’s own website the canonical source AI reaches for. The most informative comparison is the CPA category across geographies. Hawaii CPA (45%) and Austin CPA (47%) land within two points of each other, so on the web-searching surface the category gradient is largely geography-independent. But Nashville CPA sits well below both at 31%. So the web axis is mostly category-driven with real geo variation layered on top, the same shape the training-data axis shows with Gemma. One market does not fully predict the next, even inside a single category.

Why banking leads both axes: two decades of canonical product pages

Banks have done two decades of work making their websites the official source of truth for their own products. A bank’s rates, account types, branch locations, and product comparisons live cleanly on the bank’s own site, in structured product pages, refreshed daily. AI tools have nothing equivalent to cite for a wealth firm or law firm or dental practice or CPA firm. The "best wealth manager in Hawaii" question has no canonical source the way "best bank for small business in Hawaii" does, so AI reaches for a wider mix of editorial third-party content.

The honest finding for an operator on the web-searching surface. If you are a Hawaii bank, your own website is doing roughly half the work of getting AI to mention you. Clean it up. The leverage is on-site. If you are a Hawaii wealth firm, dentist, law firm, or CPA, your own site is doing 35-47% of the work and the rest is happening on third-party surfaces (publications, directories, named-attorney bylines, lead-gen platforms, professional association directories, etc.). The 35-47% on-site number means own-site work matters, and the larger off-site number means it is not enough by itself.

The second axis: training-data engines (and the cross-geo split)

Two of the seven AI tools we measure (Claude and Gemma) answer from training data instead of searching the live web. So when they mention a business by name, it means that business has formed a presence in the AI’s training data: through Wikipedia, news coverage, named-founder content, sustained editorial presence over years. This is the slowest signal in the measurement and the most defensible once it forms. A new competitor cannot show up in Claude’s training data overnight. That presence is earned through years of editorial work that an AI model later learned from.

Across the training-data measurements (the nine Hawaii categories plus two cross-geo CPA markets, Austin TX and Nashville TN), recomputed against current cohorts (per-category as-of dates in the measurement-windows note below):

The cleanest test of two axes: Hawaii hotels rank dead last on web-search (11%) and near the top on training data (Claude 47%). The same category, opposite answers, depending on which surface you ask.

The training-data picture is not the same picture as the web-searching one. Hawaii banks sit at the top of both. Hawaii CPA firms sit mid-table on the web-searching surface (45%) but collapse to 1% / 2% on the training-data surface, the lowest of any measurement by a wide margin. Hawaii law firms are the reverse: near the bottom on web-searching (35%) yet near the top of the training-data table. Honolulu real estate and Hawaii hotels are the most extreme reversals of all. Real estate sits near the bottom of the web-searching axis (14%) yet runs higher on training data (Claude 18%, Gemma 20%). Hawaii hotels go further still: dead last on web-searching (11%) yet Claude 47% and Gemma 42%, the widest inversion in the set, where the training-data engines cite hotels four times as often as the live-search engines do. The two axes do not move together, and the travel categories are the clearest proof of it.

The three CPA geographies are the most informative observations in this teardown. Same category, same query shape (geo-swapped), same methodology. Claude behaves almost identically across all three (Hawaii 1%, Austin 0%, Nashville 3%), so its training-data collapse on local CPA firms is geo-invariant. Gemma does not. It runs 2% in Hawaii, 7% in Nashville, and 23% in Austin. Austin’s spike does not replicate in Nashville, but Nashville sits clearly above Hawaii, so Gemma’s local-firm citation is a geo-dependent continuum, not a fixed behavior. The practical consequence: for one of the two training-data engines, a measurement in one market tells you little about another. You have to measure each geography.

The CPA cross-geo finding: Claude collapses everywhere, Gemma varies by market

The Hawaii CPA teardown originally surfaced a striking pattern: Claude cites Hawaii CPA firms 1% of the time, Gemma cites them 2%, while the web-searching engines reach far higher on the same questions. The original framing called this a "training-data engine collapse" for CPA firms.

Two more CPA geographies, Austin and Nashville, refined that framing into two distinct findings. Claude’s collapse generalizes across all three. Gemma’s behavior is a geo-dependent continuum, not a fixed property of the category.

Claude generalizes. Hawaii CPA Claude own-share is 1% (7 firm-owned cites), Austin 0% (3 cites), Nashville 3% (17 cites). Same engine, same category, three different geographies, near-zero firm-owned citation in all of them. This is a category-level finding about Claude’s training corpus being editorially thin on CPA firms regardless of geography. Three-geography evidence promotes it from a low-confidence single-geo observation to a high-confidence cross-geo pattern.

Gemma is a continuum. Hawaii CPA Gemma own-share is 2% (7 firm-owned cites), Nashville 7% (8 cites), Austin 23% (73 cites). Same engine, same category, three geographies, and the firm-owned citation rate climbs from 2% to 23% across them. Austin’s high number does not replicate in Nashville, but Nashville sits clearly above Hawaii. So Gemma’s representation of local CPA firms is geography-dependent, not a fixed category property. The cause is inferred (likely the geo-density of business-news content in Gemma’s training data). The observation is direct.

The reason this matters: a vendor measuring one geography per category would have published the Hawaii result and called it the category-level pattern. Measuring the same category in three geographies was the only way to discover that one of the two training-data engines (Claude) holds the pattern everywhere while the other (Gemma) varies enough that no single market predicts the next. A precise read of training-data engine behavior emerges only across geographies.

For an operator, the closable competitive surface differs by market. Hawaii CPA firms have both training-data engines as a category-wide blind spot, so competing inside the web-searching engines is the short-term move. Austin CPA firms have Claude as a blind spot but Gemma reachable through sustained web presence over Gemma’s training cycle. Nashville sits between: Claude closed, Gemma a narrower opening than Austin’s. The recommendation depends on which market the operator is in.

A third category, this one predicted in advance, confirms the Claude collapse is not CPA-specific. Honolulu med spas (Claude 2%) were the second category after CPA. Honolulu HVAC and AC-repair companies (teardown 09) are now the third, and this one we forecast before measuring: a prediction that Claude would cite these firms under 5% was committed to a public timestamped record on 2026-06-09, before any HVAC data existed. It landed at 2%, with Gemma at 63% and the web-searching engines at 38% to 66%. So three unrelated local-service industries (accounting, aesthetics, home services) all collapse on Claude. The categories that hold up on Claude (banking 88%, law 51%, wealth 38%, dental 34%) each have a canonical web presence or a deep editorial tradition that puts named firms into the training corpus. The collapse is not a quirk of one category. It is what happens to editorially thin local-service businesses, and it now replicates across three of them in different industries.

HVAC is the widest split in the dataset: Claude 2% against Gemma 63%, a 61-point gap inside one category. That is why "the training-data axis" is really two engines that can diverge completely. And this third instance was a forecast, not a fit: it was a pre-registered prediction rather than an after-the-fact observation, so the pattern is no longer a story told to fit the data. It is a rule that made a risky prediction, and the data agreed.

The honest boundary on that rule is the fourth local-service category we measured: Honolulu real estate. There Claude does not collapse at all. It cites local firms 18% of the time, more than ChatGPT search (1%) or Microsoft Copilot (0%), an outright inversion of the pattern. The reason is the demand side, not the supply side: real-estate queries are so saturated by Zillow and Realtor.com that the live-search engines return portals while Claude, answering from training data, reaches for the established brokerage names it learned. So the collapse is real and replicated across three industries, but it is category-specific, not a law of local business, and real estate marks where it stops.

Hawaii hotels then confirm the exception is a travel-category pattern, not a fluke: hotels invert even harder, Claude 47% and Gemma 42% against 11% on the web engines, because iconic and well-marketed Hawaii hotels carry the editorial footprint the training-data engines learn from while the live-search engines hand the answer to Booking and Expedia. Two travel categories, one shared inversion.

The Microsoft Copilot pattern (holds across every category we measured, including Austin and the Nashville control)

One pattern is consistent across every category we measured (the nine Hawaii categories plus Austin CPA, read with the Nashville CPA control). Microsoft Copilot (which answers using Bing’s organic search results) cites the firms’ own websites 0% to 2% of the time in every measurement.

This means: a buyer who asks Microsoft Copilot for a business recommendation in any of these categories or geographies gets pointed at third-party content (publications, Wikipedia, directories), almost never at the businesses’ own websites. Every business in every category we measured shares this gap. The Austin and Nashville data points confirm the pattern is not Hawaii-specific.

The first-mover read: whichever firm shows up first in Bing’s organic search results for these queries gets the Copilot answer mostly to itself, because every competitor is also invisible there. The gap is open, and it does not appear to be a category-specific or geo-specific problem. It is a Microsoft Copilot pattern across local-service queries.

Where to do the work changes by category and geography

The two-axis pattern, refined by the cross-geo CPA measurements, produces a different recommendation for each category and geography:

• Hawaii consumer bank. Own-site work pays off the most directly on both axes. AI is reading the bank’s website 51% on web-search and 70-88% on training-data. Cleaner structured data, sharper product pages, better FAQ markup. Both axes respond.
• Hawaii wealth firm. Own-site work matters (47%), off-site editorial matters (49%), and training-data presence is mid-pack (38-68%). The on-site cleanup is the necessary baseline. The harder work is getting into the third-party surfaces AI uses for the category (lead-gen platforms, named-planner editorial, professional bylines).
• Honolulu dental practice. Similar shape to wealth on the web-searching surface (43%), middle of the pack on training-data (34-58%). Own-site work and review-directory presence both matter. The Microsoft Copilot first-mover opening is open here as in every category.
• Honolulu med spa. 51% on the web-searching surface, tied with banking at the top, so own-site work pays off directly on the live-web engines where the category already scores well. But Claude collapses to 2% while Gemma holds at 32%, so one of the two training-data engines is a category blind spot and the other is reachable. The med spa playbook: clean on-site infrastructure for the web-searching engines, treat Claude as unreachable in the short term, and build the sustained editorial presence Gemma rewards. The Microsoft Copilot first-mover opening is open here too (0% own-share). See teardown 08.
• Honolulu HVAC / AC company. 40% on the web-searching surface, with ChatGPT search highest at 66%, so own-site work pays off on the live-web engines. Claude collapses to 2% (the pre-registered prediction) while Gemma sits high at 63%, the widest split in the dataset. Same playbook shape as the med spa (on-site for the web-search engines, Claude unreachable short-term, build for Gemma), but with the widest Gemma opening in the set. The Microsoft Copilot opening is wide open (0% own-share). See teardown 09.
• Honolulu real estate firm. The hardest category in the set, and the only inversion. Local firms get just 14% on the web-searching surface because the national portals (Zillow, Realtor.com) and ranking sites own the answer, and ChatGPT search (1%) and Copilot (0%) almost never cite an agent. The one relative bright spot is the training-data engines: Claude (18%) and Gemma (20%) cite local brokerages more than the live-search engines do, the reverse of every other category. The playbook is unlike the others: own-site cleanup alone will not beat Zillow for the live-search slot, so the realistic surface is sustained editorial and brand presence that the training-data engines reward, while treating the portal-dominated web-search engines as a long climb rather than a quick win. See teardown 10.
• Hawaii hotel (boutique or chain). The most aggregator-dominated category measured: 11% on the web-searching surface, with Booking, Expedia, and TripAdvisor owning the rest, and Copilot at 0%. But the training-data engines invert hard (Claude 47%, Gemma 42%), so a hotel with real editorial and brand presence is the named answer there. And among hotels AI does name, boutiques out-cite chains (493 to 457), so a distinctive boutique is not the underdog it fears. The playbook: do not try to beat Booking for the live-search slot, build the editorial and brand footprint the training-data engines reward, which is where boutiques win. See teardown 11.
• Hawaii law firm. 35% on the web-searching surface, but training-data presence is high (51-78%) because attorneys generate long-form editorial content (Chambers profiles, BestLawyers entries, named-attorney bylines). The implication: own-site cleanup plus continued investment in the editorial surfaces that feed into training data over time.
• Hawaii CPA firm. 45% on the web-searching surface, near-zero on BOTH training-data engines (Claude 1%, Gemma 2%). Web-searching engines are where the competitive game plays. Training-data engines are a category-and-geo blind spot that no individual firm can close in the short term. Editorial work would help over years. In the meantime, the web-searching engines are the addressable surface.
• Austin TX CPA firm. 47% on the web-searching surface (within two points of Hawaii CPA’s 45%). Claude is a blind spot (0%) as it is in Hawaii. Gemma is reachable (23%): an Austin CPA firm CAN move the Gemma needle through sustained web presence in Texas-business-news content that Gemma’s training cycle ingests. This is the materially different recommendation between Hawaii and Austin CPAs. The cross-geo data surfaced it.
• Nashville TN CPA firm. 31% on the web-searching surface, the lowest in the set, with two-thirds of citations on third-party surfaces. Claude collapses (3%) as it does everywhere. Gemma sits at 7%, well below Austin’s 23%. The cross-geo lesson: even inside one category, both the addressable web surface and the size of the Gemma opening differ by market. Nashville cannot be read off Austin.

The non-obvious read across measurements: the conventional SEO instinct (clean up your website and AI will cite you) is correct in proportion to where your category sits on the web-searching axis, but it has nothing to say about the training-data axis. And the training-data axis is not one number. It can split by engine and by geography within the same category. A firm in a category like CPA in Hawaii could perfect its on-site infrastructure and not move either Claude or Gemma. The same firm operating in Texas would not move Claude but would have a path to moving Gemma. Both surfaces respond to different work. Both can vary by geography in ways that aren't visible from a single-geo measurement.

The cohort sizes (transparency note)

The cohorts were not the same size. Cohort size is determined by how many distinct businesses AI tools mention with enough frequency to register as recurring, then expanded through cohort-coverage scans across multiple runs. Categories with more independent businesses (dental practices, CPAs, individual law firms) surface larger cohorts. Categories with consolidation (banking, wealth management) surface smaller ones. This is a feature of the measurement, not a bias: the cohort reflects who AI actually cites, not who we picked in advance.

• Hawaii consumer banking: 23 cohort domains across about 21 distinct banks and credit unions (several institutions run multiple domains), named at teardown 01.
• Honolulu dental practices: 46 dental practices (39 surfaced from run #1, 7 more added after run #2 cohort-coverage scan).
• Honolulu med spas: 15 med spas (registered through cohort-coverage scans across 3 runs). See teardown 08 for the per-category breakdown and the Claude-collapse detail.
• Honolulu HVAC: 12 AC companies (registered through cohort-coverage across 3 clean runs). See teardown 09 for the per-category breakdown and the pre-registered prediction.
• Honolulu real estate: 11 local firms (surfaced from run #1 and hard-audited to genuinely local Honolulu/Oahu agents and brokerages, with national portals and ranking sites excluded). The smallest local-firm cohort in the set, because AI concentrates its real-estate citations on a handful of national portals rather than many local firms. See teardown 10 for the per-category breakdown and the inversion finding.
• Hawaii hotels: 13 hotels across all islands (surfaced from run #1 and hard-audited to real properties, tagged boutique vs chain, with OTAs and directories excluded). 7 boutique/independent and 6 chain. See teardown 11 for the per-category breakdown and the boutique-beats-chain finding.
• Hawaii law firms: 33 firms (5 originally registered, 25 surfaced through run #1 cohort-coverage, 3 more after run #3).
• Hawaii wealth management: 42 firms (15 originally registered, 27 surfaced through three rounds of cohort-coverage scans). The expansion lifted the own-site share from 27% on run 1 to 47% after the third expansion, illustrating the cohort-coverage discipline.
• Hawaii CPA firms: 41 firms (5 anchors, 15 surfaced from run #1, 8 from run #2, 11 from run #3, with 2 anchor domain variants corrected). The expansion lifted the own-site share from 27% on run 1 to 45% against the current cohort. See teardown 06 for the per-category breakdown.
• Austin TX CPA firms: 37 firms (5 anchors with 1 domain correction, 16 surfaced from run #1, 8 from run #2, 8 from run #3). The expansion pattern matched the Hawaii CPA category closely. See teardown 07 for the per-category breakdown and the cross-geo finding details.
• Nashville TN CPA firms: 14 firms (5 anchors, 9 surfaced through cohort-coverage across 3 runs). The smallest cohort in the set. Nashville surfaces fewer independent CPA firms at citation frequency than Hawaii or Austin.

All individual firms outside the banking cohort are kept anonymized in this public artifact per the non-customer anonymization rule.

Methodology summary

Each category was measured the same way:

• 7 AI tools: Perplexity, ChatGPT search, Gemini (grounded with live web search), Microsoft Copilot via Bing organic results, Google AI Overviews (the five web-searching engines), plus Claude and Gemma (the two that answer from training data). Gemma is run as an open-weight model, so the model itself is independently inspectable.
• 18 questions per category, locked by hash so every run compares apples to apples. The questions mirror how a real buyer in each category would actually search.
• 3 repetitions per question per AI tool to separate signal from noise.
• 3 usable runs per category across the dataset (Austin CPA, 4 runs). Every measurement cleared the internal pattern-readiness rule. Banking, dental, and law numbers each held within one to two percentage points across their runs (stability signal at clean cohort coverage). Wealth's number rose from 27% on run 1 to 47% after three rounds of cohort expansion. CPA's number rose from 27% on run 1 to 45% against the current cohort. Both are the cohort-coverage discipline working as designed: more runs surface more cohort members, the headline number trends toward true ceiling rather than away from it.

Full methodology, source bucket definitions, and pattern-readiness rules at /methodology/. Every number here traces to a hash-locked question set and a dated run on the claims ledger. Gemma is open-weight, so the model itself is independently inspectable.

Honest scope and what this does not prove

Ten published measurements across nine categories, read with the Nashville control, is enough to establish a two-axis pattern but not enough to declare a universal law, and real estate and hotels already show the pattern has real exceptions. The pattern is consistent in direction (web-searching and training-data engines respond to different mechanisms) and in magnitude (the spread is wide enough to matter) but it lives inside a small sample of categories, mostly in Hawaii with two cross-geo CPA markets (Austin TX, Nashville TN), measured across May and June 2026.

What is not in this teardown:

• Large national categories. The set is nine Hawaii categories plus two mainland CPA markets (Austin, Nashville). Bigger national categories may show different patterns because cohort dynamics differ.
• Categories outside professional services, consumer financial services, local aesthetics, home services, residential real estate, and hotels. The categories measured here cluster narrowly. B2B SaaS, restaurants, healthcare specialties beyond dental, all unmeasured here.
• Whether the pattern holds over time. AI tools refresh their training data and search indices on schedules we do not control. The same categories measured six months from now may produce different numbers.
• Causation. We measured what AI cites. We did not test whether changing what is on a firm’s site or in third-party publications would cause AI to cite differently. That requires pre-registered experiments against named businesses with control for confounds. Different scope, different methodology.