Research

Original studies of how AI answer engines choose who to name and who to cite, published in full with their methods, data and limits.

  • What Decides Who AI Answer Engines Name and Cite

    Robert Griesmeyer · November 15, 2026 · 48,902 words · about 213 min

    AI answer engines decide which brands to name in their prose and which web pages to cite as sources. We study how those decisions are made, and how they depend on whether the engine ran a web search, using 68,636 stored answers to 7,711 prompts from seven commercial engines over five months of 2026, and 4,284 answers to fixed sets of category questions in which every brand named was extracted. A language model can take a name only from its parameters or from its context, and a web search works by appending retrieved text to the context: answers that returned sources were billed a median of 4,400 to 11,000 more input tokens than answers that did not. Three natural contrasts in the data support causal or near-causal readings. When one engine's search tool was withdrawn, its unprompted naming of tracked brands fell from 1.90% to 0.70% while four unchanged engines answering the same prompts rose from 1.18% to 3.53% (difference in differences -3.55 percentage points, 95% interval -7.11 to -1.52). Across 556 prompts that ChatGPT answered both with and without a search, it named the brand only when it had searched. And one engine reports the pages it retrieved as well as those it cited: of 45,035 retrieved results it cited 42%, with the first result cited 68% of the time and the ninth 32%. When engines do search, most of who is named is still decided by a prior the engines share: how often a model without search names a brand predicts which brands five searching engines name with a pseudo-R-squared of 0.52, and what was retrieved adds 0.03. Retrieval matters most for brands with no prior, which are named 80 times more often when their own site is among the sources, and still only 7% of the time. Citations are heavy-tailed over 61,222 sites; two engines answering the same prompt share under a tenth of their combined sources; and in 61% of answers that cite a brand's own site for a prompt that does not name the brand, the brand goes unnamed. Given retrieval, citation is predicted by position, by the match between a page's title and the question, and by short self-contained passages, and not by FAQ blocks, summary boxes or question headings. A single reading is unreliable: an unprompted mention recurs on the next reading of the same prompt less than half the time. We document five measurement artefacts, three of which produced strong false findings before correction, and specify two experiments that would close the main gaps.