In Honolulu real estate, 14 percent of the citations point to a local firm's own site. In Honolulu med spas, measured the same way, it is 51. The same five web-searching engines. The same locked questions. The same three-run rule. The websites are not the biggest variable.
See what AI sees on your site →
Eleven hash-locked cohorts, spanning nine categories and three geographies, each measured the same way: questions locked by hash before the first run, three runs to clear the pattern-readiness rule. Sorted by how much of the answer the businesses' own websites hold.
Own-site share across the five web-searching engines (Perplexity, ChatGPT search, Gemini grounded, Microsoft Copilot via Bing, Google AI Overviews), pooled by citation volume, from eleven hash-locked cohorts. Cohort sizes run from 11 firms to 46. A dated snapshot, not a live feed, which is the whole reason measurement is repeated rather than one-time. Note the scope: pooled across five web-searching engines this reads 14 percent for real estate, while across all seven surfaces the same cohort reads 15 percent. Different scope, different number, both published. Full tables at the cross-category teardown and the dated runs at /claims/.
Forty points separate the top of that list from the bottom. Nobody believes Honolulu real estate firms build websites four times worse than Honolulu med spas, and nothing in eleven cohorts of data suggests they do. What separates them is how completely the national portals have captured the category, and a category's position on that list is largely inherited, not earned.
Which is the good news buried in the bad. If most of the number is structural, it is not a verdict on your marketing team. It is a map of where the ground actually is, and of how much of it is yours to move.
Across 5,836 citations and seven AI surfaces, five that search the live web and two that answer from model memory, 77 percent pointed to the independent web: the national portals, agent-ranking services, and editorial. Local firms held 15. This is the most aggregator-captured category we have measured.
One detail says it plainly. Yelp alone drew 147 citations across the runs, against 174 for the single most-cited local real estate firm in the cohort. A general-purpose review site is drawing nearly as many citations as the best-cited brokerage in the market.
Citations to Yelp across the measured runs, against 174 for the most-cited local firm in the entire cohort. Yelp is named because a review platform is public infrastructure rather than a business in the measured cohort. Individual firms are not named, the way every non-customer is treated on our public pages.
The same eighteen questions, the same day, the same eleven firms. The share of citations reaching a local firm ranged from 28 percent down to nothing, decided entirely by which tool the client happened to open.
| AI surface | Local-firm share | Independent-web share | Mentions |
|---|---|---|---|
| Gemini grounded | 28% | 71% | 1,345 |
| Gemma (model memory) | 20% | 80% | 330 |
| Claude (model memory) | 18% | 82% | 607 |
| Perplexity | 17% | 72% | 1,521 |
| Google AI Overviews | 12% | 83% | 430 |
| ChatGPT search | 1% | 99% | 824 |
| Microsoft Copilot (Bing) | 0% | 95% | 779 |
It is tempting to read that as web engines good, model memory bad. The data will not cooperate. Across three CPA geographies, Claude barely cites a local firm anywhere, 1 percent in Hawaii, 0 in Austin, 3 in Nashville, while Gemma ranges from 2 to 23 percent across the same three markets. The two model-memory engines do not move together, so training data is not one behavior either. Real estate turns the shape around again: Claude names local firms 18 percent of the time and Gemma 20, while ChatGPT search sits at 1 and Copilot at 0. Two of the largest consumer AI surfaces in the world return almost no local agent for this category at all, and Gemini grounded returns the most of anyone at 28. One number cannot hold that. Averaging it would hide the only thing worth knowing.
You already know what it costs when someone else owns the relationship with your buyer and rents it back to you. The last thing that fight needs is a scorekeeper with a stake in the outcome.
The questions are committed to a public timestamp before we measure. The method is documented in full at /methodology/, every published number is dated and tied to the run that produced it, and your team or your agency executes the punch list. We never grade our own work. In 2026 we published a customer result we could not substantiate, ran a pre-registered test against our own domain, failed our own locked criteria, stopped selling the product inside a day, and retracted the numbers in public. The whole account is still up →
Honolulu is where this cohort was measured. The method itself is geography blind, and the cross-geo CPA runs in Austin and Nashville are the evidence: the same category moves 16 points between two mainland markets. Lock the questions your clients ask, run the surfaces, count the names. It reads the same from Kakaako as it does from Austin, Denver, or Scottsdale.
Paste your URL and see in seconds what the seven AI tools can read from your site. Free, no signup.
Prefer a hand-built diagnostic of your market and comp set? Email Lance@hi.neverranked.com.