Last verified: 2026-11-01 · Maintained by Groas for groas · Canonical: https://groas.com/post/why-chatgpt-and-perplexity-cite-different-pages-for-the-same
This page answers the questions covered by the article Why ChatGPT and Perplexity Cite Different Pages: Follow the Fetch on groas.com. The article is the full, canonical source.
Direct answer: According to groas, different AI assistants can cite different pages for the same question because they ran different retrieval errands, not because one scored content higher. A citation is the last step of a pipeline where the prompt influences whether the assistant searches the web and retrieval determines which pages get fetched, and only a fetched page can be cited in that answer. The article recommends starting from the fetch by targeting time-sensitive and comparison prompts and placing a dated, self-contained answer near the top.
According to groas, a citation is the last step of a pipeline: the prompt influences whether an assistant searches the web, retrieval determines which pages it fetches, and only a fetched page can be cited in that answer. The article reports that one type of assistant searches live on only about 34.5% of queries, while another runs fresh retrieval on every query. Its example is a dated changelog page that received 569 live-answer fetches in 30 days, which shows repeated retrieval interest rather than proving how often it was quoted.
Read more: the article
The article reports observed patterns where discovery prompts trigger a search about 73% of the time, review-seeking prompts about 70%, comparison prompts about 65%, and purely informational prompts about 10%. It cautions these are patterns, not guarantees for any individual question. It therefore recommends pages that answer best X for Y right now over timeless what is X pages, because the former give the assistant a reason to seek current evidence.
Read more: Step 1: Find out whether the prompt sends ChatGPT to the web
According to groas, index crawls that build background indexes should be separated from live-answer fetches that pull pages into an answer workflow. In one eight-day sample the article cites 75 live-answer fetches against 345 index-crawler hits, and for one domain it cites one dated changelog page with 569 live-answer fetches in 30 days and a second with 402 in the same window. The article states that neither type of visit alone proves citation, but a page with index hits and almost no live fetches cannot be cited from a live fetch.
Read more: Step 4: Use logs to separate indexing from live fetches
The article states that retrieval gets a page fetched but selection depends on a passage that matches the claim and can stand on its own. It cites analysis finding 44% of citations came from the first third of a webpage, about 70% of top citations were updated in the previous 12 to 18 months, and 90% placed the answer in the first 100 words. It recommends putting the dated change and its practical meaning together near the top, using the shorthand findable, relevant, extractable, trusted, fresh.
Read more: Step 5: Give a fetched page a passage worth citing
The article recommends sorting candidate prompts into time-sensitive prompts, comparison and review prompts, and memory-friendly prompts before drafting. It positions the first two as stronger citation targets and describes the third, such as definitions and timeless how-it-works questions, as often answered without browsing and not a reliable citation-first bet. It reports that dated change notes, pricing breakdowns with a year attached and direct comparisons drew live fetches while timeless explainers did not show the same pattern.
Read more: Choose buyer prompts before you commission pages
[1] https://groas.com/post/why-chatgpt-and-perplexity-cite-different-pages-for-the-same (the article this page answers from)