The Breed Study dataset: 180 AI answers, 12 states, 3 engines (2 September 2026)

This is the data behind Who decides which breeders an AI names? — every question asked, every engine, every state, and every web source each engine retrieved before it answered, as two plain CSV files anyone can open, check and reuse.

Direct-Me LLC · Mark Griffin · Data collected 2 September 2026 · Dataset page published 9 September 2026
180 answers · 3,884 source citations · 919 registrable domains · every figure on this page recomputed from the raw transcripts on 8 October 2026

---

What it is

On 2 September 2026, between 02:04 and 03:19 UTC, three AI assistants with live web access — Anthropic’s Claude (claude-opus-5), OpenAI’s GPT (gpt-4.1) and Perplexity (sonar) — were each asked five questions a puppy buyer types, once per state, across twelve US states. That is 5 × 12 × 3 = 180 answers. All 180 completed; none failed, none were retried. For each answer the full list of web sources the assistant retrieved was saved. This dataset publishes those source lists, the question text, the engine, the timestamp and the length of each answer. It does not publish the answer text itself.

The five questions

Asked verbatim, with [state] replaced by the two-letter code (for example “Who are the best Golden Retriever breeders in OR?”):

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

Each question was asked 36 times: twelve states, three engines.

The twelve states

Arizona (AZ), California (CA), Colorado (CO), Florida (FL), Georgia (GA), Michigan (MI), North Carolina (NC), Ohio (OH), Oregon (OR), Pennsylvania (PA), Texas (TX) and Washington (WA). Fifteen answers per state.

Per-engine totals

A citation is one entry in an answer’s retrieved-source list, counted with duplicates. Domains are registrable domains (marketplace.akc.org and akc.org count once). The per-engine domain counts overlap, which is why they do not sum to 919.

The fifteen most-cited domains, by engine

The top ten carry 1,352 of 3,884 citations, 34.8%. Row 15 is an artefact of the two-label domain rule, kept here because it is what the downloadable file contains: twelve Australian .com.au hosts collapse into one entry, all of them retrieved in five Washington answers where an engine read “WA” as Western Australia. The next single site is dogbreederlicensing.org at 32 citations in 23 answers.

Download

• breed-study-dataset-2026-10-08.csv — one row per answer, 180 rows, 62 KB. Corrected 8 October 2026.
  Columns: state, engine, model, question, scanned_at_utc, n_sources, domains_cited (semicolon-separated registrable domains), n_unique_domains, answer_word_count, cites_akc_org, cites_gooddog_com.
• breed-study-domains-2026-10-08.csv — one row per registrable domain, 919 rows, 38 KB. Corrected 8 October 2026.
  Columns: rank, domain, citations, answers_cited_in, states_cited_in, then citations and answers broken out per engine (citations_anthropic, citations_openai, citations_perplexity, answers_anthropic, answers_openai, answers_perplexity).

Method notes

• What a citation is. One entry in the retrieved-source list an engine returned alongside its answer. If an answer lists the same URL three times, that is three citations. The exception is Perplexity, which returns its sources in two lists that repeat each other; each Perplexity source is counted once. Thirteen answers — twelve from GPT, one from Claude — returned no sources at all and are kept in the file with n_sources = 0.
• What a domain is. The URL host, lowercased, with a leading www. removed, reduced to its last two dot-separated labels. This is deliberately simple and is disclosed as such: it is the rule under which the corpus has 919 domains, and it mis-groups two-part country suffixes (row 15 above). A public-suffix-aware rule gives 930 domains and an identical top ten.
• What is not in the file. The answer text. Each row carries the answer’s word count and its sources, not its words, so the file is small and does not republish engine output. The dataset also states no count of how many distinct breeding programs were named across the corpus, because that has not been counted.
• Reproducibility. The figures on this page were recomputed from the raw transcripts on 8 October 2026 and match the corrected figures in the published study.
• Correction, 8 October 2026. Our scanner saved every Perplexity source twice, once from each of the two source lists Perplexity returns. The first version of this dataset (files dated 2026-09-02) counted both, so every citation count that included Perplexity was too high: 5,076 citations in total, now 3,884; 2,384 for Perplexity, now 1,192. Answer counts, states and the 919 distinct domains are unchanged, and the domains_cited column was already de-duplicated. The n_sources and citation columns in the new files count each Perplexity source once.

How to cite

Direct-Me LLC (2026). The Breed Study dataset: 180 AI answers about dog breeders in 12 US states, 3 engines. Collected 2 September 2026. https://direct-me.org/research/breeder-ai-study-dataset/

The written analysis is at direct-me.org/research/breeder-ai-study/.

Licence

This dataset is released under Creative Commons Attribution 4.0 International (CC BY 4.0). You may copy, redistribute, transform and build on it for any purpose, including commercially, provided you credit Direct-Me LLC and link to this page.

Direct-Me LLC · Mark Griffin · about the author