DankRank AI Search Citation Study — Methodology

Study dates
- ChatGPT: September 10, 2026
- Claude: September 11, 2026
- Perplexity: September 11, 2026
- Gemini: September 11, 2026

Design
- 24 fixed, non-branded prompts.
- Four AI search experiences.
- 96 prompt-engine observations.
- Each prompt was evaluated within its engine run and the findings are observational.
- This is not a controlled comparison of underlying foundation models.

Prompt groups
1–6: Provider selection
7–12: AI visibility measurement
13–18: Technical discovery
19–23: Research and experiment credibility
24: Integrated search strategy

What was recorded
- Exact prompt.
- Organizations/brands surfaced.
- Visible cited URLs/sources.
- Engine and run date.
- Available run/provenance notes.

Citation records
citations.csv contains one row per explicit source-list URL record preserved in the run files.
The field raw_url_occurrence_count preserves a formatting asymmetry in the Gemini file: its Markdown source links repeat the same URL in both link text and link target. The governed quantitative QA counted those literal URL occurrences, producing 42 raw URL occurrences from 21 semantic source records. The other three run files use one raw URL occurrence per source-list record.

Headline run totals using comparable visible source records
- ChatGPT: 24/24 prompts with visible citations; 92 visible source records; 64 unique cited domains.
- Claude: 18/24; 56; 21.
- Perplexity: 24/24; 225; 156.
- Gemini: 6/24; 21 visible source records; 6 unique cited domains.

Gemini raw-format note
The preserved Gemini file writes each source as a Markdown link where the URL appears once as link text and once as link target. A literal URL parser therefore sees 42 URL occurrences. Those 42 occurrences represent 21 listed source records. citations.csv stores one row per listed source and preserves the literal multiplier in raw_url_occurrence_count.

Cross-engine overlap
- 211 distinct cited domains across all four runs.
- 180 appeared in one engine.
- 27 appeared in two.
- 3 appeared in three.
- 1 appeared in all four.

Brand / first-party rule
A first-party citation requires the surfaced brand's identifiable owned domain to appear in that same answer's cited sources.
Third-party-hosted brand posts or profiles do not count as first-party.

QA-passed provider-selection first-party results
- ChatGPT: 7/72 = 9.7%
- Claude: 6/53 = 11.3%
- Gemini: 8/30 = 26.7%
- Perplexity: 43/107 = 40.2%

Two explicit mapping notes
- Third (3RD) -> get3rd.com because the raw organization record self-identified the relationship.
- SEMrush agency directory -> agencies.semrush.com, treated as Semrush-owned because it is a Semrush subdomain.

Buyer-source ownership classification
Source ownership was independently QA'd only for prompts 1–6. Accordingly, source_class is populated in citations.csv only for that provider-selection block and intentionally left blank for prompts 7–24.

Owner-normalization rule
Clearly co-owned subdomains were merged to the parent owner for the buyer-source ownership analysis.
Specifically, semrush.com and agencies.semrush.com count as one Semrush-owned source family.
That produces:
- 68 literal hostnames
- 67 owner-normalized source owners

QA-passed buyer-source ownership mix
- Agency / service provider: 42
- Software / platform: 12
- News / editorial: 5
- Community / social: 3
- Directory / marketplace: 2
- Practitioner / consultancy: 2
- Search trade publication: 1

Run-protocol limitations
- One run per engine.
- Search/retrieval protocols were not symmetric.
- ChatGPT states live web research grounded the answers, but its stored raw file lacks the standardized self-declared provenance block used by the other runs.
- Claude used three broad searches and then judgment across the prompt set; some zero-citation answers reflect that run protocol.
- Perplexity operated as a search-native experience and disclosed account personalization.
- Gemini declared search enabled/active but exposed visible citations only on prompts 1–6.
- Results should therefore be described as observed run behaviour, not universal engine-native behaviour.

QA correction
The original ChatGPT tally was 90 citation entries / 71 unique domains.
Independent recount of the preserved raw file corrected this to 92 / 64.
The 211-domain union and 180 / 27 / 3 / 1 overlap distribution were unchanged.

Interpretation limits
- Citation recurrence is not an authority score.
- Commercial ownership does not establish inaccuracy or bias.
- A citation does not prove it caused a recommendation.
- These observations are not population estimates for any engine.
