In One Study, 44% of ChatGPT's Verified Citations Came From the First 30% of a Page

In Kevin Indig's analysis of 18,012 verified ChatGPT citations, 44.2% came from the first 30% of a page's content. Separately, AirOps found that only 15% of the 548,534 pages ChatGPT retrieved were cited in the final response. Two findings, one lesson: being retrieved is not being cited, and the answer has to sit where the engine reads first. If your answer comes after three paragraphs of preamble, you are working against the pattern Indig measured.

ChatGPT Retrieves Far More Than It Cites

The retrieval-citation gap is the first thing every founder needs to understand: a page can be fetched during answer generation and still not appear as a citation.

AirOps' analysis found that 85% of the 548,534 pages ChatGPT retrieved across 15,000 prompts were not cited in the final response. Retrieval does not equal citation. The data does not show what happened to those pages inside the model.

And when ChatGPT does cite, the citations cluster at the top of the page: in Indig's analysis 44.2% came from the first 30% of the content, 31.1% from the middle and 24.7% from the final third. If your answer isn't in the opening paragraphs, front-loaded, specific and structured, you are competing against sources that led with theirs.

That is a selection problem as much as a ranking problem.

Each AI Engine Picks Different Winners — and That's the Point

The engines do not cite alike. Qwairy's Q3 2025 study of 118,101 AI answers measured the volume split directly; the Wikipedia concentration comes from Profound's 680-million-citation dataset; the Gemini and Claude points are ZipTie.dev's cross-platform compilation, drawn from sources it links, including other vendors' blogs:

• Perplexity was built citation-first. It runs a live web query for every prompt and averaged 21.87 citations per question in Qwairy's Q3 2025 data (276,139 citations across 12,623 questions). It cites broadly.
• ChatGPT bolted search onto a conversation engine. It averaged 7.92 citations per question in the same Qwairy data (190,894 citations across 24,099 questions) and leans on Wikipedia more than on any other source: 47.9% of the citations within its top 10 most-cited sources in Profound's data (7.8% of all ChatGPT citations there; Qwairy puts Wikipedia at 4.8% of ChatGPT citations in Q3 2025).
• Gemini has the most concentrated citation pattern in ZipTie's table, with a Gini coefficient of 0.351. A small set of elite sources captures most of the citation share.
• Claude applies ethical reranking that can override relevance signals, per ZipTie, favoring peer-reviewed and institutional content.

Four engines, four retrieval architectures, four citation outcomes for the same query. If your entire strategy is optimizing for one engine, you're invisible in three others.

First-Party Content Loses Because It Wasn't Built to Be Cited

Here's the part most brands refuse to hear: your product page, your feature comparison, your corporate blog post — they were built to rank in Google, not to be cited by AI.

Semrush's guide describes the pattern: a Perplexity answer might name your product but cite a G2 review page, a Reddit thread, or a Forbes comparison instead of your own site.

Research from Zhang et al. studied 602 controlled prompts across ChatGPT, Google AI Overview, and Perplexity and measured what they call "citation absorption" — the degree to which a cited source actually shapes the generated answer. Their finding: ChatGPT cites fewer sources than Perplexity, but the sources it does cite have substantially higher influence on the final answer. High-influence pages tend to be longer, more structured, semantically aligned with the query, and rich in extractable evidence: definitions, numbers, comparisons, and procedural steps.

The paper does not compare first-party with third-party sources, so whether research reports, analyst coverage and editorial features fit that profile better than a brand's own product page is our inference, not its finding.

About 16% of Cited Sources in One Audit Were AI-Generated

The citation quality problem runs deeper than first-party vs. third-party. An audit of 712 real-world queries on politics, health and the environment across ChatGPT, Copilot, Gemini, and Perplexity found that roughly 16% of cited sources were themselves AI-generated.

Machines citing machines. And users have no reliable way to tell the difference.

Separate research benchmarking 14 LLMs found that even the strongest frontier models maintain link validity above 94% and relevance above 80%, but achieve only 39–77% factual accuracy in their citations. Worse: as retrieval depth increases from 2 to 150 tool calls, factual accuracy drops by approximately 42%.

In that benchmark of deep-research agents, more retrieval did not produce more accurate citations.

The Fix: Build Evidence Architecture, Not Brand Pages

If engines cite third-party, evidence-dense, front-loaded sources more readily than first-party brand pages (an inference from the studies above, not a measured result), the answer isn't to publish more on your own domain. The answer is to become the source that third-party authorities cite.

This is what Machine Relations is built for. Not PR placements in the traditional sense — earned evidence architecture that compounds across the sources AI engines actually trust.

Three moves based on the research above:

1. Front-load answers. Indig's analysis found citations concentrated at the top of the page (Search Engine Land's summary). If the engine has to read three paragraphs to find your answer, the source that led with theirs has the advantage.

2. Earn third-party citations. Your brand-owned page is a supporting signal. The citation-winning surface is the analyst report, the editorial feature, the research coverage that names you as the evidence. That's earned media designed for machine retrieval.

3. Structure for extraction. Zhang et al. found that high-influence pages are richer in definitions, numerical facts, comparisons and procedural steps. If your content doesn't contain extractable evidence, the engine has nothing to take from it.

The brands winning AI citation share in 2026 aren't the ones publishing the most content. They're the ones whose evidence appears in the sources these engines actually trust.

FAQ

Why does ChatGPT cite Wikipedia more than my brand's website?

ChatGPT cites Wikipedia more than any other source: it was 47.9% of the citations within ChatGPT's top 10 most-cited sources in Profound's dataset of 680 million citations (August 2024 to June 2025), and 7.8% of all ChatGPT citations there; Qwairy's Q3 2025 data puts it at 4.8%. The data does not say why, but Wikipedia pages are structured, front-loaded and evidence-dense, which matches the traits the research above associates with citation. Brand pages that lack that structure start at a disadvantage on those traits.

How many sources does ChatGPT cite per response compared to Perplexity?

In Qwairy's Q3 2025 study of 118,101 AI answers, ChatGPT averaged 7.92 citations per question while Perplexity averaged 21.87, about 2.8 times as many. Perplexity was built as a citation-first search engine; ChatGPT added search to a conversational model. Qwairy measured the gap, not its cause.

What is citation absorption and why does it matter for AI visibility?

Citation absorption measures how much a cited source actually shapes the AI-generated answer — not just whether it's linked, but whether the engine extracted language, evidence, or structure from it. ChatGPT cites fewer sources but absorbs more heavily from each one, meaning the sources it does cite have outsized influence on what users read.

Can AI engines cite AI-generated content?

Yes. An audit of four generative search engines on politics, health and environment queries found that approximately 16% of cited sources are AI-generated, raising questions about source quality and information reliability. Users currently have no standardized way to distinguish AI-generated citations from human-authored ones.