Source Aggregation

Source aggregation is the stage of an AI search pipeline that collects the results of every search the system ran for a question, removes duplicates and weak or stale material, ranks what remains, and selects the small set of passages the model will write from and cite. It sits between retrieval and writing: query fan-out produces many candidate results, source aggregation narrows them down, and multi-source synthesis turns the survivors into an answer.

The gap between what an engine looks at and what it cites is large, and vendors document it. OpenAI's web search guide distinguishes inline citations, which 'show only the most relevant references', from a sources field that 'returns the complete list of URLs the model consulted', and notes that 'the number of sources is often greater than the number of citations'. The same list includes real-time feeds labeled oai-sports, oai-weather and oai-finance (OpenAI). Promptwatch records this split for tracked responses as three tiers: inline citations, footnote sources, and search results the model browsed but did not cite. A page in the last tier was retrieved and then lost during aggregation.

Perplexity has described the stages in some detail for the search stack behind its answer engine. Results from lexical and semantic retrieval are merged into one candidate set with 'comprehensiveness over precision' as the priority. Prefiltering then removes 'clearly non-responsive or stale content'. Several ranking stages follow, each more expensive than the last, ending with cross-encoder reranking models that perform 'the final sculpting of the result set'. Other engines do not publish their pipelines, but the outputs show the same funnel.

Aggregation applies several filters, and each suggests what survives:

• Each passage is judged on its relevance to the specific sub-query that retrieved it.
• When many pages say the same thing, only one or two are kept, so content that repeats widely available information is easy to drop.
• Stale pages are filtered early for time-sensitive questions.
• Citations tend to be spread across several domains rather than given to one site for everything.

This explains citation patterns that look odd from a classic SEO view. A specialist page can beat a major publisher if it is the only source with a particular figure. A high-ranking page can be retrieved and still lose, because a reranker found a more direct passage elsewhere. And a brand can be well known yet absent from citations, because every fact it publishes also appears on stronger or more recent pages. Content with real information gain, meaning facts, data or analysis not available elsewhere, is the hardest to deduplicate away.

For GEO, the practical diagnostic is to compare the pages an engine retrieved with the pages it cited for your prompts. If your page is retrieved but not cited, the problem is aggregation: sharpen the passages, add specifics, and update dates. If it is never retrieved, the problem sits earlier, in crawling, indexing or retrieval coverage. The AI search optimization guide covers the content side in more depth.