Research
Two studies, in two industries that have nothing to do with each other, run the same way. Both ask the same question — not what an AI assistant says, but what it read before it said it. Both come back with the same answer.
The finding that shows up in both
Buyers now put the question to a machine first. The machine answers out of directories, registries and institutions — and the business being asked about is usually not in the room. Almost nobody has measured what that means for their own market, because almost nobody has looked at the source list instead of the answer.
That sentence is easy to assert. Here is what it looked like when it was actually counted, twice, in two industries picked for how little they have in common:
- Senior living, Oregon. Directories and aggregators supplied 60.9% of every source retrieved, and 95.4% on the questions that actually name communities to a family. Operators’ own websites supplied 1.3% on those same questions. Of the 62 communities the answers named, exactly one had its own website cited anywhere in its market.
- Dog breeders, twelve states. The ten most-quoted domains carry about a third of every citation in the corpus. One directory alone is quoted 571 times, across 84 of the 180 answers. The AKC is named in 146 of 180 answers — roughly four in five.
The two corpora share almost no vocabulary, no geography and no engine mix. What they share is the shape: an institution or an aggregator is the substrate, and the individual business is a detail the assistant may or may not reach.
The breeder study found the more hopeful half of it too, and it is the half nobody publishes. Assistants do name individual businesses, and far more freely than expected — one answer names fourteen breeding programmes in a single sentence. And the cited set is not closed to newcomers: in the Oregon corpus, 100 of the 178 domains were cited only once or twice, and fifteen genuinely new domains — one registered a month before the scan — drew 64 citations between them. New sites do get in. The question is what gets them in, and that is what both studies are trying to answer.
Two industries, one method
Neither study is a survey, an opinion piece, or a vendor’s benchmark. Both do the same four things, and the method is published in full inside each article rather than summarised after it:
- Ask the questions a buyer actually types — in real, named markets, not hypotheticals.
- Keep the entire answer and every source the engine retrieved, word for word, as a saved transcript. The source list is the unit of analysis, not the prose.
- Classify every domain by hand, with nothing left in a default bucket, and join the result to an outside register where one exists — in Oregon, the state’s own list of licensed facilities, including the 121 communities that no answer mentioned.
- Publish the limits with the findings. Both articles state, in their own text, that nothing in them connects an AI citation to a single enquiry, tour, deposit, admission or sale.
| Dog breeders | Senior living | |
|---|---|---|
| Recorded answers | 180 | 40 |
| Markets | 12 states | 6 Oregon markets |
| Questions | 5 per state, per assistant | 5 per market, plus 10 hyper-local |
| Assistants | 3 — ChatGPT, Claude, Perplexity, 60 answers each | 1 — Claude with live web search |
| Sources logged | 5,076 | 1,329 |
| Distinct domains | 919 | 178 |
| Collected | 2 September 2026 | 17–18 August 2026 |
| Outside register | None available | ODHS licensed facilities — 183 communities |
The differences are as instructive as the overlap. The breeder study runs three assistants and finds that each one leans on a different favourite domain, which a single-engine study cannot see at all. The senior-living study runs one engine but joins its results to a government register, which lets it ask a question the breeder study cannot: of every licensed community that exists, which ones never appear? Sixty-two of 183 were named. The other 121 were not.
The two studies
Who decides which breeders an AI names?
Three assistants, the five questions a puppy buyer actually types, in twelve states — and then all 5,076 sources they used to answer. The assistants do name individual breeding programmes, freely. But the shortlist a buyer ends up holding is assembled out of a very short list of websites, and almost none of them belong to a breeder. Mean 28.2 sources per answer; every transcript kept.
The author ran his own two programmes through the same questions. His written work is quoted as a source fifteen times across six of the 180 answers, in Oregon, California and Ohio. His kennel is named in three, all of them Oregon. The writing travels; the business does not.
Which Oregon senior living communities does AI actually recommend?
The questions Oregon families ask when a parent needs care — what are the best assisted living facilities in Eugene, what does it cost per month in Medford, how do I choose a community for my mom in Portland. Every retrieved source recorded, classified by hand, and joined to Oregon’s own licensed-facility register, so the analysis includes the communities that were never named at all.
Two things survive statistical controls as predictors of being named: licensed bed count, and presence on one aggregator’s city roster. Both are associations, not levers. On-site schema markup, technical site score, star rating, photo count, published pricing, review recency and operator footprint all returned null — and the article says so, at length, rather than quietly omitting them.
Both are ungated and neither asks for anything. No form, no email wall, no offer at the end. They are published in full because a study that ends in a sales page is not a study, and because the working is the point — including the parts that went against what the author expected before he ran it.
The data behind the breeder study is open: the Breed Study dataset — 180 answers, 5,076 citations, 919 domains, two CSV files, CC BY 4.0.
From the breeder side of the same data: Why your puppies aren’t selling — six reasons a health-testing breeder’s litter sits, and five things to check tonight.
What neither study shows
Stated here as plainly as it is stated inside both articles, because a research page that hides its limits is advertising.
- No link to revenue. Nothing in either corpus connects an AI citation to a single enquiry, tour, deposit, admission or sale. Not one.
- Retrieval, not attribution. These transcripts record what the engine was shown. They do not record which of those sources the model demonstrably leaned on when it wrote its answer.
- A snapshot, not a trend. The breeder corpus is one day. The senior-living corpus is two. These systems change without notice, and neither study claims to have measured a stable state.
- Known skew, published rather than buried. Half the Oregon corpus is Portland, and the study says so in its own methods section rather than leaving a reader to work it out.
- Correlation is not a lever. Where the Oregon study finds an association, it says the direction of the arrow is not established and that acting on it is untested.
A third industry would work the same way
Nothing in either method is specific to puppies or to care homes. It needs three things: the questions a buyer in that market genuinely types, a list of the real businesses in it to measure the answers against, and the discipline to keep every source list rather than the answers alone. Trades, clinics, equipment dealers, professional services, manufacturers — the procedure does not change. Only the question list and the register do.
The reason there is not yet a third study on this page is time, not method.
If you want this run on your own market
Three ways in, in ascending order of size. None of them is described on this page, on purpose — the research section does not sell, and these are links rather than offers.
- Anyone. The same three assistants, your own questions, your own business, and the transcript sent back to you: the free AI visibility check.
- Breeders. If the breeder study is describing your programme, the work built on it is the Litter Launch.
- Owner-operated companies. If you read the method and thought about your own market rather than about dogs, that conversation starts here.