What AI Cites When You Ask What an Expert Says

Where the answers came from

Share of the 1,399 question and assistant pairs that cited at least one site of each type, across both runs.

1. YouTube, podcast apps and show pages

   48%

   ChatGPT 25% · Perplexity 71% · Claude 24% · Gemini 71%

2. The expert's own site or company

   48%

   ChatGPT 49% · Perplexity 58% · Claude 27% · Gemini 57%

3. Transcript, summary and book-note sites

   45%

   ChatGPT 38% · Perplexity 56% · Claude 33% · Gemini 51%

4. LinkedIn, X, Reddit and other social sites

   36%

   ChatGPT 30% · Perplexity 76% · Claude 26% · Gemini 11%

5. News, business press and trade publications

   30%

   ChatGPT 19% · Perplexity 60% · Claude 20% · Gemini 22%

6. Medium and Substack posts by others

   28%

   ChatGPT 5% · Perplexity 30% · Claude 24% · Gemini 53%

7. Gavel

   18%

   ChatGPT 15% · Perplexity 23% · Claude 11% · Gemini 22%

8. Wikipedia

   4%

   ChatGPT 1% · Perplexity 6% · Claude 3% · Gemini 7%

The 12 most cited sites

Out of 1,424 question and assistant pairs.

Each assistant, run to run

Overlap is the share of cited sites the two runs of one question had in common, averaged over the questions.

Searched: share of runs that returned a source. Sources: average sources in one answer. Own site: share of answers citing the expert's own site or company.

Method

1. Questions: we took the 114 people with verified quote cards in Gavel's library. For each, we asked about the three founder problems they speak on most, in the form 'What does <name> say about <topic>?'. We added one question for each of the 83 framework pages Gavel publishes about them. That makes 356 questions, fixed before the first run.
2. Assistants: ChatGPT (gpt-5.6-luna with OpenAI web search), Perplexity (sonar), Claude (claude-haiku-4.5 with Anthropic web search) and Gemini (gemini-3.6-flash with Google grounding). We called them through the OpenRouter API on 23 October 2026. These are API models, so the consumer apps may answer differently.
3. Runs: every question went to every assistant twice, for 2,848 answers and 0 errors. A pair counts a type of source if either run cited it.
4. Sources: we read the citations each answer returned and sorted every site into a type. The expert's own site covers their personal site and newsletter, plus the company or fund they run, such as a16z.com for a16z partners. Most answers also cited other company, consultancy and personal sites, which we did not sort further.
5. Shares count the 1,399 pairs where the assistant returned at least one source. ChatGPT searched in 98% of runs and Claude in 91%. Perplexity and Gemini searched every time.
6. Gemini hides the page behind a search redirect, so its Gavel citations count the site, not the page.