ChatGPT cites only half the pages it retrieves

What happened

Per the Ahrefs blog, ChatGPT retrieves dozens of URLs when answering a single prompt but only cites roughly 50% of them. The study analyzed 1.4 million prompts to understand why one retrieved page earns a citation while another, equally available to the model, gets nothing.

Ahrefs reports that the gap between retrieval and citation is not random. Pages that get cited tend to match the specific intent of the prompt more cleanly, present claims in self-contained passages the model can lift without ambiguity, and carry signals of authority the model trusts at synthesis time. Retrieval gets you into the room. Citation requires a different set of qualities.

The headline number to internalize: being indexed and retrievable buys you a coin flip, at best. The other half of the work happens at the moment of answer construction, where the model decides which sources to name.

Why it matters for your brand

Most enterprise GEO programs are still optimizing for retrieval. Schema, crawlability, content freshness, embeddings hygiene. That work matters, but Ahrefs has now quantified what we have suspected for months: retrieval is necessary and insufficient. If your bank, foundation, or industrial group is already showing up in ChatGPT's retrieval set and not getting cited, the fix is not more SEO. It is rewriting the passages the model is choosing between.

For financial services brands, this is a sharp problem. When a CFO asks ChatGPT about counterparty risk frameworks or transition finance taxonomies, the model is pulling from regulators, consultancies, banks, and trade press simultaneously. The page that gets named in the answer is the one that stated the definition cleanly in two sentences, not the one that buried it in a 3,000 word thought leadership essay. McKinsey and BIS win these citations because their prose is built for extraction. Most bank marketing copy is not.

For multilaterals and UN agencies, the stakes are different but the mechanic is the same. UNDRR, CGAP, IMF, and WHO all produce dense PDF reports that ChatGPT can retrieve but struggles to quote. The institutions that translate their findings into clean web-native passages, with the headline statistic in the first sentence and the methodology in plain English, will be cited as the authority on disaster risk, financial inclusion, or pandemic preparedness. The institutions that keep their best work locked in 80 page PDFs will be retrieved and ignored.

For major industrial groups, the citation gap maps directly to category authority. When a procurement lead asks ChatGPT about low-carbon cement, embodied carbon disclosures, or ISO 14067, the answer will name two or three companies. Holcim, Heidelberg, CRH, or none of them. That selection is not determined by who has the biggest brand. It is determined by who wrote the most quotable sentence on the topic. Industrial marketers who think their job ends at publishing a sustainability report are going to lose share of voice in the answer layer.

The strategic implication: brand teams need to start auditing not just whether their content is retrieved, but whether it is structured to be quoted. That is a content design problem, not a distribution problem, and it sits with the editorial function, not the SEO function.

The signal in context

This Ahrefs finding sharpens a pattern we have been tracking across recent Pulse coverage: the AI answer layer rewards a different kind of writing than the open web rewarded. showed that models prefer self-contained chunks. showed how multilaterals lose ground when their best evidence is locked in long-form documents. The 50% citation rate is the quantitative spine that ties those observations together.

The brands that will dominate ChatGPT, Perplexity, and Gemini answers in 2025 are not the brands with the biggest content libraries. They are the brands whose sentences are easiest to quote. That reframes the editorial brief for every CMO whose audience now starts research inside an LLM.