Stop Retrieving Documents. Start Retrieving Claims.

One of the most interesting AI papers published this week is not about ChatGPT, Google, or search engines. It is about chemistry. Researchers behind AskChem argue that we have been retrieving the wrong thing for decades. Instead of searching for papers, they propose searching for claims. Every paper is decomposed into small, verifiable, evidence-backed claims, each linked directly to its source, then organised into taxonomies and evidence graphs for retrieval and synthesis. At first glance, that sounds like a niche problem for scientists. It is not. It may represent the direction AI search is heading.

<h2>Documents Were Built for Humans</h2>

Traditional search engines were designed around documents. You searched for information, the search engine returned webpages, and you synthesised the answer yourself. Large language models changed that. They no longer retrieve documents for people — they synthesise information from many sources into a single response. That means the document is no longer the most useful unit of knowledge. The claim is.

<h2>Claims Are the Building Blocks of AI</h2>

When ChatGPT answers a question, it does not simply repeat webpages. It constructs a response from hundreds of smaller ideas. Consider the question: 'Why is Volvo considered one of the safest car manufacturers?' The answer is not one document. It is a collection of claims: Volvo invented the three-point seatbelt; Volvo consistently performs well in independent crash testing; Volvo has invested heavily in automotive safety research; Volvo's vehicles receive high safety ratings from independent organisations. Each of those is an individual claim. Together, they create the broader association: Volvo equals safety. AskChem treats those claims as the primary object of retrieval rather than the documents that contain them.

<h2>This Has Implications Far Beyond Science</h2>

Marketing has traditionally focused on keywords. SEO shifted attention toward webpages. AI optimisation may shift attention again — toward claims. Brands do not really want AI to remember every page they have published. They want AI to consistently remember the right facts. Those facts are claims. AskChem indexes 2.4 million structured claims from 147,000 papers, links each claim to source evidence, organises them into multiple taxonomies, and even identifies contradictions across the literature. As AI search evolves, brands may need to think the same way.

<h2>The Future May Be Claim-Centric</h2>

This raises a different question for marketers. Instead of asking 'How many times did AI mention my brand?', they may soon ask: Which claims does AI consistently associate with my brand? Which claims are supported by credible evidence? Which claims are influencing recommendations? Which important claims are missing entirely? Those questions move beyond visibility. They begin measuring how AI constructs an understanding of a brand — a network of reinforcing evidence rather than a single marketing message.

<h2>AEO Updates Takeaway</h2>

The most valuable unit of knowledge in AI search may no longer be the document. It may be the claim. The brands that understand, strengthen, and support their most influential claims could have a significant advantage as AI systems become increasingly claim-centric rather than document-centric.

<h3>References</h3>

[1] AskChem Research Team, <em>AskChem: Claim-Centric Retrieval for Scientific Literature</em>, 2026.