DankRank AI Search Citation Study — Public Dataset

Files
- prompts.csv — the fixed 24-prompt panel and prompt classes.
- citations.csv — explicit source-list records by engine and prompt.
- brands.csv — surfaced-brand records, with first-party citation coding for prompts 1–6.
- methodology.txt — definitions, run conditions, QA corrections and limitations.
- README.txt — this guide.

Important scope notes
1. source_class is intentionally populated only for prompts 1–6. That ownership classification was independently QA'd. We did not backfill unreviewed source classes for prompts 7–24.
2. first_party_cited is populated for the provider-selection prompts 1–6, where the metric was independently checked.
3. owned_domain is included when a governed mapping was needed for the first-party analysis. A blank owned_domain does not mean the organization lacks a website; it means no public mapping was required for this release.
4. Gemini's raw file writes source URLs in Markdown link syntax, causing each source URL to occur twice textually. The public comparison therefore uses 21 visible source records, not 42 literal URL occurrences. citations.csv contains one row per listed source and preserves the literal-count multiplier in raw_url_occurrence_count.
5. The study is one documented run per engine and should not be used to make universal claims about any platform.

Headline checks
- Prompt rows: 24
- Prompt-engine observations: 96
- Distinct cited domains: 211
- Cross-engine overlap: 180 / 27 / 3 / 1
- First-party rates in prompts 1–6:
  ChatGPT 7/72
  Claude 6/53
  Gemini 8/30
  Perplexity 43/107
- Buyer-source ownership: 42 / 12 / 5 / 3 / 2 / 2 / 1 across 67 owner-normalized sources

Reuse
You are welcome to rerun the prompt panel, challenge the coding, or compare later observations. Please preserve the documented limitations when citing the findings.
