What 7 AI tools cite for real business categories.
Public teardowns of the same hash-locked 7-AI-tool methodology applied across 10 measurements: 9 Hawaii business verticals plus a non-Hawaii cross-geo CPA market (Austin TX). Cohort sizes, source-type distributions, and per-AI-tool breakdowns named. Individual non-customer firms anonymized at the firm level per the non-customer rule. The banking cohort is named in full because these are publicly competing brands.
Start here
Two axes decide what AI cites for your category. Geography is a third.
The headline cross-category finding from all ten published measurements, with a Nashville CPA cross-geo control. Own-site share on web-searching engines lives on one axis (banking and med spas tie at 51%, Nashville CPA sits at 31%). Training-data presence lives on a second axis. The cross-geo CPA measurements surfaced a third dimension: the training-data axis itself can split by engine within the same category depending on geography. Claude's collapse now holds across three unrelated industries (CPA, med spas, and HVAC, all near 2%), and the HVAC result was pre-registered: we forecast it before measuring. Gemma varies by geography (Hawaii 2%, Austin 23%).
Per-category teardowns
The 75-point engine spread no single-engine tool can see.
21-bank cohort named in full. For the same Hawaii consumer banking questions, the same day, five AI tools cite bank-owned websites 60-75% of the time while Microsoft Copilot cites them 0% of the time. The 75-percentage-point spread is invisible to any tool that measures fewer engines.
The Microsoft Copilot first-mover opening.
46-practice cohort. Microsoft Copilot cites zero practice websites across 801 mentions. Every practice in the cohort shares this gap. Three of the top five recurring non-practice sources are insurance carriers (HMSA, Delta Dental, HMSA Dental), a measurable AEO surface most practices likely treat as administrative paperwork.
The top 5 own two-thirds. The opening sits outside.
33-firm cohort. Top 5 firms account for 66% of all firm-owned mentions. The dominant firm alone gets 538 mentions, roughly three times the next firm. For firms outside the top 5, the closable competitive ground sits in two specific places: Microsoft Copilot (cohort-wide 0% firm-owned) and the long tail of buyer-specific questions where AI defaults to legal directories.
The lead-gen middlemen sitting between buyers and the firms.
42-firm cohort. SmartAsset and getwarmer.com (the two lead-gen platforms whose business model is routing buyers to advisors) combined account for 14% of all third-party citations. AI structurally routes wealth-management buyers through matchmaker middlemen before sending them to the firms themselves. Claude’s training-data presence for Hawaii wealth firms is structurally thinner than for adjacent relationship categories.
The training-data engine collapse.
41-firm cohort, 5,809 measured citations. Web-searching engines (OpenAI, Gemini, Perplexity, Google AI Overviews) reach 45-60% firm-owned share. Training-data engines (Claude, Gemma) collapse to 1-2%, the opposite of every other category we have measured. The most distinctive single-category finding in the portfolio. The two-axis framework documented in the cross-category teardown was surfaced by this measurement.
The Claude collapse generalizes. The Gemma collapse does not.
37-firm Austin cohort, 7,738 measured citations. First non-Hawaii measurement, built to test whether the Hawaii CPA training-data findings generalize. Claude's collapse holds across both geographies (Hawaii 1%, Austin 0%). Gemma's does not (Hawaii 2%, Austin 23%). Same query shape, same week, same methodology. The recommendation for an Austin CPA firm differs from the recommendation for a Hawaii CPA firm on Gemma specifically. The cross-geo data surfaced the engine-by-engine split.
The Claude collapse is not just a CPA story.
15-firm cohort. Claude cites med-spa-owned sources 2% of the time while the web-searching engines reach 53-64% on the same questions. Med spa is the second category, after CPA across three geographies, where Claude's training-data citation of local firms collapses to near-zero. Gemma holds at 32%, so the two training-data engines diverge inside one category. The collapse tracks editorially thin local-service categories, not one industry.
The category where AI cites the aggregators, not the agents.
11-firm cohort. Local firms get just 15% of citations while 77% go to portals like Zillow and Realtor.com. The training-data collapse inverts here: Claude cites local firms 18% and Gemma 20%, while the web-searching engines ChatGPT search (1%) and Microsoft Copilot (0%) barely cite an agent at all. The first category where the live-search engines, not Claude, are the blind spot for local firms.
AI sends hotel-seekers to the OTAs. The boutiques it names beat the chains.
13-hotel cohort across all islands. Only 17% of citations go to a hotel's own site; 74% go to Booking, Expedia, and travel editorial. But among hotels AI does name, boutiques out-cite chains 493 to 457, and the most-cited hotel in Hawaii is a boutique. The contest is not boutique versus chain, it is hotel versus OTA, and the training-data engines (Claude 47%, Gemma 42%) are where a distinctive property wins.
We predicted the collapse before we measured it.
12-firm AC-company cohort. We committed a prediction to a timestamped public record before any data existed: Claude would cite these firms under 5%. It landed at 2%, while the web-searching engines reached 38-66% and Gemma sat at 63%, a 61-point split inside the training-data engines. The third unrelated industry where Claude collapses, and the first one we called in advance. That is research, not a claim told to fit the data.
How these get produced
Each teardown is the public anonymized version of the same forensic measurement we run for a paying customer’s engagement. The structural axes (cross-engine, cross-competitor, source-type, position-in-answer) are evidence-grade. The internal pattern-readiness rule (3+ usable runs per category before claiming a pattern) is cleared for every teardown above. The question sets are hash-locked and each run is dated on the /claims/ ledger. Full methodology: /methodology/.
Future teardowns will publish here as new categories are measured. Customers in measured categories see their named, unredacted version of the same data inside the paid engagement deliverable. The public version stays anonymized at the firm level per the non-customer rule.
NeverRanked home · How we measure · Free AI search check · FAQ