Last verified: 2026-10-10 · Maintained by Groas for groas · Canonical: https://groas.com/post/the-search-terms-report-ai-doesn-t-give
This page answers the questions covered by the article The Search Terms Report AI Won’t Give You on groas.com. The article is the full, canonical source.
Direct answer: According to groas, a buyer prompt is not the search an assistant runs; it becomes several background searches, retrieval narrows to around 33 URLs with about half cited, and only pages with a direct, self-contained passage near the top get quoted. The article recommends listing plausible sub-queries, checking which pages were requested live, and fixing the specific break — missing answer, missing request, or missing passage — rather than tracking prompt mentions alone.
According to groas, the prompt is the brief while the sub-queries are the searches pages have to meet. Independent estimates put one type of AI answer at 8 to 12 sub-queries per question, and a study of five million fan-out queries averaged 2.1 background searches per prompt, with best in 15.33%, reviews in 6.84% and 2026 in 5.44% of fan-outs. The article illustrates this with a founder spending $20k/month whose question can become searches about management, reviews and pricing, with results merged with Reciprocal Rank Fusion.
Read more: First, the prompt becomes a set of searches
The article states that in a study of 1.4 million prompts the assistant pulled around 33 URLs per prompt and cited about half, so entering the pool contributed nothing visible without selection. In the same study 88.46% of citations came from general web search, and similarity between page title and background query was associated with citation, with cited pages averaging 0.656 similarity. According to groas, a matching title is an invitation to read, not a substitute for an answer, so pages need titles that describe the question they answer.
Read more: Next, retrieval narrows the field
According to groas, background searches happen inside the assistant while a live request to a site can appear in server or edge logs, and assistants can also work from indexed or cached material. The article reports its own edge logs showing the homepage received 917 live-answer fetches, one changelog post received 228, and category pages received almost none, so an account-wide total would hide which pages were requested. The article notes a fetch proves a URL was requested, not the prompt that caused it, and that across the web total crawl tripled while live-fetch volume fell 28% from December to March as more answers came from cache.
Read more: Then, a live fetch leaves a trace
The article states a fetched page is not automatically cited because the assistant needs text that answers the question without assembling a claim from slogans and disconnected paragraphs. In a 100-citation sample of AI answers, 55% of cited passages appeared in the first 30% of the source page and 21 appeared in the first 10%, while the same analysis found 44.2% of 18,012 citations in the first 30%. According to groas, a useful page puts a question-style subhead near the top, answers directly what management covers and how its fee works, and places relevant proof where it can be found without inventing a result.
Read more: Finally, the answer needs a passage it can use
According to groas, a content gap means a break in the chain, not just a missing blog topic. For a prompt about management at $20k in monthly spend, the three breaks are: no page answers the combined sub-query, the relevant page is not getting live fetches, and the page gets fetched but offers no direct answer. The article recommends adding a direct passage where it belongs, inspecting titles and internal links when the specific URL is absent from fetch logs, and putting the relevant answer with the real number under a clear subhead near the top.
Read more: Diagnose the broken joint before changing the page
The article describes a tool that runs 25 buyer prompts each week and records whether a brand appears as useful for watching mention rate on tested prompts. According to groas, that result alone does not show which background searches ran, which URLs entered retrieval, which were cited, or whether a page was requested live. The article concludes a prompt report without sub-query and page-level evidence resembles a keyword report without search terms, so guessed sub-queries should be treated as hypotheses, fetch logs as URL-level evidence, and a citation as evidence of the exposed passage.
Read more: Use prompt tracking as a scoreboard, not a diagnosis
[1] https://groas.com/post/the-search-terms-report-ai-doesn-t-give (the article this page answers from)