How AI Picks Its Sources: What 21,000 ChatGPT Citations Reveal About Content Strategy

A groundbreaking study by Kevin Indig, published on Search Engine Journal (March 24, 2026), analyzes over 21,000 ChatGPT citations from 1.2 million responses. The results reveal, for the first time with data, how AI systems select their sources.

Four Key Findings

1. About 30 Domains Own 67% of All AI Citations Per Topic

AI citations are heavily concentrated: the top 10 domains in a topic account for 46% of all references, the top 30 for 67%. There are effectively only about 30 seats at the citation table. Everything else remains nearly invisible.

Concentration varies by industry:

• Education is the most concentrated: the top 10% of domains capture 59.5% of all citations. tefl.org alone answers 102 different prompts.
• Crypto follows at 43% concentration for the top 10%.
• Healthcare is the least concentrated (13%): no single provider dominates, giving newcomers realistic opportunities.
• CRM/SaaS and HR Tech are similarly distributed (16.1% and 14.4%), as dozens of comparison sites split citations.

2. The Citation Advantage Starts at 10,000 Characters

Longer content correlates with more citations, but there's a ceiling. The jump from 5,000 to 10,000 characters is the largest single step, nearly doubling citations. Pages over 20,000 characters average 10.18 citations versus 2.39 for pages under 500 characters.

Crucially, the rule is industry-specific:

• Finance: Optimal length at 5,000 to 10,000 words (10.9 citations/page). After that, values drop as compact, authoritative summaries are preferred.
• Education, Crypto, Product Analytics: More length consistently pays off.
• SaaS: Weakest length effect. Structure and domain authority matter more than word count.

3. 58% of Cited URLs Are Cited Only Once

Most pages are one-time citations. The strategically valuable "evergreen pages" appear across 10+ different prompts. They share common traits:

• Category-level overview format ("Best X for 2026")
• Broad topic coverage within a single page (What is X, how to choose X, top providers, pricing)
• Explicit year anchoring in URL or title

Key takeaway: A single evergreen page covering 10+ search intents is worth more in AI citation reach than 10 separate single-intent pages.

4. AI Reads Primarily the First 30% of a Page

44.2% of all citations come from the top 30% of a page. The bottom 10% (conclusions, summaries) are virtually invisible to AI, receiving only 2.4 to 4.4% of references.

• Finance pages show the strongest effect: 43.7% of citations land in the first 30%.
• Healthcare and HR Tech have the flattest distribution with more evenly spread content.
• Education content peaks at the 30-40% range, as answers typically come after an introduction.

Recommendations for Businesses

• Topic breadth over domain authority: Covering an entire topic cluster matters more than domain strength.
• Comprehensive pages over many single pages: The traditional "one keyword, one page" model doesn't work for AI citations. Teams that stick to it will be structurally locked out.
• Put key claims in the first 30% of the page: Summaries and conclusions are rarely cited by AI.
• Industry-specific strategy: There is no universal playbook. What works for a SaaS platform could actively harm a finance brand.

Source: Kevin Indig / Search Engine Journal, "The Science Of How AI Picks Its Sources", March 24, 2026. Read the original article