Who decides which breeders an AI names?

I asked three AI assistants the five questions a puppy buyer actually types, in twelve states, and then read all 5,076 sources they used to answer. They do name breeders — far more freely than I expected. But the shortlist a buyer sees is assembled out of a very short list of websites, and almost none of them belong to a breeder.

Mark Griffin, Direct-Me LLC · Published 3 September 2026
Data collected 2 September 2026 · 180 answers, 12 states, 5 questions, 3 assistants · every figure below recomputed from the raw transcripts

---

Why I ran this

I breed and board dogs. Our family runs a boarding and daycare resort, a Golden Retriever program and a Cavalier program, and like most people reading this I had started noticing that buyers arrive already holding a shortlist — one they got from an AI assistant rather than from a page of search results.

What I could not find anywhere was an honest answer to the obvious question: where does that shortlist come from? Plenty of people were willing to tell me. Nobody was showing their working. So I ran it myself, and I am publishing the working, including the parts that did not go the way I expected.

One of those parts matters enough to say before anything else. I went in believing that AI assistants dodge the question and refuse to name breeders. That belief was wrong, and the data killed it in the first hour. What replaced it is more useful.

What I did

Five questions, phrased the way a buyer phrases them, asked once per state per assistant:

• Who are the best Golden Retriever breeders in [state]?
• Who are the best Cavalier King Charles Spaniel breeders in [state]?
• Where can I find a health-tested Golden Retriever puppy in [state]?
• Which dog breeders in [state] do OFA health testing?
• How do I find a reputable dog breeder in [state] and avoid a puppy mill?

Twelve states: Arizona, California, Colorado, Florida, Georgia, Michigan, North Carolina, Ohio, Oregon, Pennsylvania, Texas and Washington. Three assistants, each with live web access on: Anthropic’s Claude, OpenAI’s GPT, and Perplexity. That is 5 × 12 × 3 = 180 answers, all collected on 2 September 2026 inside a seventy-five minute window.

Every one of the 180 completed. None failed, none timed out, none were dropped or retried, and every answer was saved in full along with the list of web sources the assistant actually retrieved in order to write it. That saved source list is the whole point — it is the only way to see which websites are doing the deciding.

Domains are counted as registrable domains, so marketplace.akc.org and akc.org count once. The three per-assistant domain counts overlap heavily, which is why they do not sum to 919.

First: they do name breeders

Here is a single answer from Perplexity, asked who the best Cavalier King Charles Spaniel breeders in Texas are. It opens by naming fourteen programs in one sentence — Bluebonnet, Valera, Skylynn, Cavernet, Hill Country, Evesham, Westchester, Luxxar, Genesis, Moki Do, PAW’S, Kingbird, AustinCavalier and Casanova — then gives each one a line saying where it is and what it claims.

Asked which Texas breeders do OFA health testing, the same assistant on the same day names ten, each with the specific clearances that breeder publishes. Asked the Cavalier question about Oregon, it names eight. This is not an assistant hedging. This is an assistant building a buyer’s shortlist.

So the anxious version of the question — will an AI even mention a breeder? — is answered, and the answer is yes, routinely. The real question is the harder one underneath it.

They name breeders. The thing worth understanding is how they choose which breeders to name — because that choice is made before the assistant writes a word, out of the handful of websites it retrieved.

The shortlist is decided by a shortlist

Across the 180 answers there were 5,076 source citations, spread over 919 distinct domains. That sounds like a wide, healthy, competitive field. It is not.

The ten most-cited domains carry 1,809 of the 5,076 citations — 35.6%, better than a third of the whole corpus. The other 909 domains share what is left. And 552 of those 919 domains — three in five — appear in exactly one of the 180 answers, which is to say they surfaced once and were never seen again.

Two other ways of holding that number make it land harder than the percentage does. The first is that those ten domains are not merely a large share — they are close to unavoidable. 161 of the 180 answers cite at least one of the ten. Thirteen answers cited nothing at all, so that is 161 of the 167 answers that cited anything. Six answers in the entire study were assembled without touching the top ten.

The second is Good Dog on its own. gooddog.com was cited 571 times, in 84 of the 180 answers, in all twelve states — one website accounting for 11.2% of every source retrieved in the study. Line all 919 domains up by citation count and start adding from the bottom: you have to combine the bottom 431 of them — 47% of the entire field — to reach 572 citations, which is what that single site earned by itself. Inside Perplexity it is starker again. Good Dog supplied 512 of the 2,384 sources Perplexity retrieved, 21.5% of everything it read, and appeared in 54 of its 60 answers.

That is the shape of the field a breeder is actually competing in. It is not 919 websites. It is roughly ten, and one of those ten is doing more work than the bottom half of the rest put together.

“Answers” is the number of the 180 answers in which the domain was cited at least once. The “what it is” column is my own description of each site, not a measurement.

Look at row 14 before anything else. One individual kennel’s own website is cited more often than the listing sites and content sites below it — 45 citations, in 12 different answers, across eight states. A breeder’s own site can absolutely become one of the sources that decides these answers. It is just extraordinarily rare, and it does not happen by accident.

Each assistant has its own gatekeeper

This was the finding I did not see coming, and it is the one with the sharpest practical edge. The three assistants do not read the same web. They lean, very hard, on different sources.

Good Dog is not a gatekeeper. It is Perplexity’s gatekeeper: 512 of its 571 citations came from that one assistant. Claude barely touched it — twelve citations across sixty answers — and leaned instead on the AKC and, remarkably, on a Golden Retriever owners’ forum. GPT’s single most-quoted source in the entire study is a content site most breeders have never heard of. Perplexity is the only one of the three that cited the Better Business Bureau at all, and it did so a hundred times.

The practical consequence: a breeder can be prominently named by one assistant and completely absent from another, on the identical question, on the same day. There is no single “AI” to be visible to. Being listed on Good Dog is close to a prerequisite for existing inside Perplexity’s answers and nearly irrelevant to Claude’s. Being active in the breed club and in a breed forum is the reverse.

Breadth or depth: two different ways in

Divide each domain’s citations by the number of distinct pages on it that were cited, and the top sources split into two clear types.

Good Dog wins on breadth: 275 different pages, roughly one per breeder or per litter, each pulled in when it happens to match. Devoted to Dog wins on depth: thirty-four pages, quoted 227 times, the same handful of state directory pages returned to over and over. The OFA is the extreme case — sixteen pages, 170 citations, better than ten quotations per page.

Both routes work, and they ask different things of a breeder. Breadth means having a listing that exists and stays accurate. Depth means being inside a page that is itself the answer to a question thousands of buyers ask.

What the winning sources have in common

Devoted to Dog is worth stopping on, because it is the most instructive site in the study and the least famous. It is not a breeder, not a registry and not a club. It is a content site — and it out-cited the Golden Retriever Club of America two to one, and was GPT’s single most-quoted source anywhere in the corpus.

What it publishes is a state-by-state breeder directory in which each breeder gets a row, health-clearance and OFA status is set out in a table, and the write-ups carry named authors with stated professional credentials. That is not a better opinion than a breed club’s. It is a more machine-legible one: scoped to a state, structured as a table, attributed to a person, and carrying the specific verification data the buyer’s question is about.

Run down the top ten and the same pattern is there every time. The OFA publishes a queryable database of health-testing results. The AKC publishes structured breeder records. The GRCA and the CKCSC publish state-scoped member and breeder-referral lists. Good Dog publishes one structured profile per breeder with its screening status attached. Every one of them answers “which specific breeders in this specific state meet this specific standard” in a form a machine can lift a name out of.

The winning sources are not the most authoritative or the most trusted. They are the ones that state a specific fact about a specific named breeder in a specific place, in a form that can be quoted without interpretation.

Your own website can be one of those sources

Go back to that Texas Cavalier answer — the one that named fourteen programs in its first sentence. It drew forty citations from sixteen domains. Thirteen of those sixteen were an individual kennel’s own website: bluebonnetcavaliers.com, valeracavaliers.com, skylynncavaliers.com and so on. Only three of the sixteen — Good Dog, Reddit and one content site — were third-party sites at all.

So when an assistant does build a shortlist of individual programs, it is very often quoting those programs’ own pages directly. Eleven of the fourteen named programs were named with their own website open in front of the assistant; the other three — PAW’S, AustinCavalier and Casanova — were named without their own page among the sources at all. That is the good news, and it is the whole argument of this study in one answer: a breeder’s own site is eligible to be a deciding source.

The bad news sits in the same data. Fourteen programs got named out of a state that has hundreds. And 552 of the 919 domains in this study — overwhelmingly individual kennel websites — appeared in exactly one answer out of 180 and never again. Being retrievable once is not the same as being the source a system reaches for.

The question changes which gatekeeper matters

The five questions do not behave alike. Each has its own dominant source, and a breeder legible to one is not automatically legible to the others.

Thirty-six answers per question — 12 states × 3 assistants.

Two things stand out. The Cavalier question draws on the narrowest field in the study — 175 domains, against 307 for the health-tested Golden question. In a smaller breed there are fewer sources doing the deciding, which cuts both ways: harder to be one of them, and worth far more once you are.

And the puppy-mill question — the one the most anxious buyers ask — is the only one whose top source is the Better Business Bureau. Not a club, not a registry, not a health database. A complaints database.

The AKC is the reflex, but it is not the shortlist

The AKC is named in the text of 146 of the 180 answers — about four in five. That is by literal string match on the answer text. By a different method, counting only answers that actually cited a page on akc.org, the figure is 74. Counting either, it is 155.

I am giving all three because they measure three different things, and the honest headline is the spread rather than the flattering end of it. What holds under every method is that the AKC is the reflex reference for “how do I judge a breeder”, and that AKC registration is treated as a floor rather than as a differentiator. In the answers that name individual programs, registration is nearly always mentioned as table stakes, and what separates one named breeder from the next is the specific clearances each publishes.

One kennel’s numbers, measured the same way

Our own programs are in this study on exactly the same footing as everyone else’s. The state-level questions were asked blind, with no breeder names supplied to the assistants, and nothing told them we exist.

• 3 of the 180 answers named one of our programs. All three were Oregon answers.
• 6 answers cited our own pages as a source, quoting them 15 times across 8 distinct URLs.
• Those six answers were in Oregon, California and Ohio.
• Counted as a source rather than as a breeder, greenacresnextgen.com ranks 46th of the 919 domains in this study — the top five per cent — on 12 citations drawn from 6 distinct pages of ours, quoted by Claude and by Perplexity, in three different states.

The out-of-state citations are the finding, and they are uncomfortable to publish. Nothing in California or Ohio recommended our kennel — nobody in either state has any reason to. What happened is that the assistants pulled our written explainers about Cavalier health testing and quoted them to buyers two thousand miles away as reference material, while the shortlist of breeders alongside them was assembled from somewhere else entirely.

Set the two measurements next to each other and the argument of this entire study fits in one line: our written work is quoted to buyers in California and Ohio, and nothing anywhere recommends our kennel. Ranking 46th of 919 as a source and appearing in three answers as a breeder are not the same achievement, and the first does not turn into the second on its own.

Our pages were good enough to be quoted as an authority and not legible enough to be counted as a breeder. That gap is what this whole study is about, and it was open on our own sites first.

What this does not show

Everything above measures what three assistants retrieved on one day. It does not measure buyers, and stating the limits properly matters more to me than the headline does.

• Nothing here connects a citation to an enquiry, a deposit or a placed puppy. Not one. No buyer behaviour was measured at all.
• One capture, one day. All 180 answers were collected on 2 September 2026. These systems change their retrieval behaviour frequently; a re-run in a month may not reproduce these proportions, and I would expect the exact ranks to move.
• Two breeds, twelve states. Goldens and Cavaliers. Other breeds have different clubs, different directories and probably different gatekeepers.
• I have not counted how many distinct breeding programs are named across the corpus. That census does not exist. I have read individual answers and counted the names inside them, which is why this page says “fourteen in this answer” and never “N across the study.” If you see that number quoted anywhere, it was invented.
• Thirteen of the 180 answers cited no sources at all — twelve of the thirteen from GPT, and all thirteen on the puppy-mill question, which it tends to answer from general knowledge rather than by searching. Those answers are counted in the 180 and contribute nothing to the 5,076. Removing them would raise the mean, so I have left them in.
• Two-letter state codes are ambiguous, and it shows. Five of the fifteen Washington answers pulled in Australian sources — 53 citations to .au domains — because “WA” is also Western Australia. I have not corrected for it. It is a real property of how these questions get interpreted, and a breeder in Washington state should know about it.
• The domain descriptions in the tables are mine. The counts are machine-derived; the “what it is” column is a human judgement and should be treated as one.

The sources the assistants leaned on, and the data itself

Seven sites account for most of what the three assistants quoted when they answered these questions. They are worth knowing by name, because they are where a buyer’s answer is actually being assembled: Good Dog (571 citations across 84 answers, almost all of them Perplexity’s), Devoted to Dog (227, ChatGPT’s favourite), the American Kennel Club (204, Claude’s first source), the Orthopedic Foundation for Animals (170), the Golden Retriever Forum (143), the Golden Retriever Club of America (104) and the Cavalier King Charles Spaniel Club (69). Every figure above is a count of citations in the 180 recorded answers, not a judgement about any of those sites.

The counts are open. The full dataset — one row per answer, one row per cited domain, no answer text — is published with a CC BY 4.0 licence at the Breed Study dataset page, with the two CSV files to download. If you count something different in it, I want to hear about it.

How this was counted

All 180 responses were captured with their retrieved source lists intact and stored as raw JSON, one file per state, question and assistant. Every figure on this page was recomputed from those files on 3 September 2026, and the arithmetic behind each one is written down.

• Citations are source entries as the assistant returned them, counted with duplicates: if an answer cited the same page three times, that is three citations. 5,076 in total.
• Domains are registrable domains — www. stripped, subdomains merged, so marketplace.akc.org and apps.akc.org count as akc.org. This yields 919. A stricter method that also handles two-part country suffixes such as .com.au yields 930. The top ten and the 1,809 are identical under both, as is 35.6%. Two figures on this page are not, and both are quoted here under the 919 rule for consistency with the headline: the number of domains appearing in exactly one answer is 552 under the 919 rule and 561 under the 930 rule, and the fifteenth row of the domain table is puppyfinder.com under the 930 rule but a meaningless com.au under the 919 one, so that table row uses the 930 rule and says so here. Quoting a count from one rule against a domain total from the other is exactly the mistake corrected on this page on 4 September.
• “Named” means a literal string match in the answer text. Literal matching under-counts: a program referred to only by a nickname would be missed. An absence in these counts is not proof of absence in the answer.
• Nothing was excluded. All 180 answers are in every total, including the 13 that cited no sources. A failed call would have been recorded as a failed call and reported as one. None failed.

---

Corrections, 4 September 2026. Every figure on this page was recounted from the transcripts on 4 September. Three did not survive, and are fixed above. One. The count of domains appearing in exactly one answer was given as “561 of those 919”. Both numbers are real but they come from different domain-grouping rules, and mixing them was the error: under the 919-domain rule this page headlines, the figure is 552; 561 is the answer under the 930-domain rule. The page now uses 552 throughout and states both in the method notes. Two. The Texas Cavalier answer was described as citing eleven kennel websites out of sixteen domains, which cannot be right alongside the three third-party sites named in the same sentence. It is thirteen kennel sites of sixteen; eleven of those thirteen belong to programs named in the answer’s opening sentence, which is the sentence that was miscarried over. Three. The thirteen answers that cited no sources were described as twelve-of-thirteen on the puppy-mill question. It is all thirteen. No other figure on the page changed. The sections on citation concentration and on our own kennel were expanded the same day with further counts from the same transcripts.

Mark Griffin runs Direct-Me LLC and, with his family, a boarding and daycare resort, a Golden Retriever program and a Cavalier program in Oregon. He also sells marketing services to breeders, which is a plain conflict of interest and the reason every method and limit in this study is stated in full rather than summarised. What he sells is described on the breeder services page; there is no offer, price or sign-up anywhere in this article, deliberately. An earlier study applying the same method to a different industry is here. The author’s own program, Green Acres Nextgen Cavaliers in Boring, Oregon, was one of the kennels whose pages the engines cited.