Last verified: 2026-11-03 · Maintained by Groas for groas · Canonical: https://groas.com/post/how-chatgpt-actually-picks-a-source-the-search-you-never-see
This page answers the questions covered by the article How ChatGPT Picks a Source: The Hidden Search and the Paragraph That Survives on groas.com. The article is the full, canonical source.
Direct answer: According to groas, a buyer question is first broken into narrower searches that gather candidate pages, then the assistant fetches candidates and selects passages that answer cleanly. The article recommends three jobs in order: be retrievable for a component question, be fetchable without blocks, and provide a self-contained passage with numbers, conditions and boundaries. A page that misses a component question, blocks live reading, or buries the answer gives the assistant little to use.
According to groas, when a buyer asks a long question the assistant can rewrite it into narrower queries, run them separately and merge the results. A page must first get into that candidate set, then offer a passage the assistant can use directly. The article states most visibility work optimizes neither step.
Read more: the article
The article states the original wording may never be searched because the assistant runs narrower sub-queries instead. In one test of 15,000 prompts cited in the article, only about 12% of cited URLs ranked in the top 10 for the same prompt while roughly 80% did not rank for the original query at all. According to groas, a page far down for the broad wording can still be a strong candidate for one component question.
Read more: First, the buyer’s question becomes smaller searches
According to groas, finding a page is not the same as reading it because answers can require a live fetch at answer time rather than training data. The article recommends a fast page with readable HTML and warns against hiding the answer behind a consent wall or making JavaScript the only way to see it. In the article's own log example, one explanatory post was fetched 528 times for live answers while the homepage was read separately for brand facts.
Read more: A candidate page still has to survive the live fetch
The article recommends putting the direct answer under a heading that names the question, with the quotable sentence first and elaboration afterward. According to groas, a usable paragraph names both options, shows the arithmetic at a stated spend, and states what the comparison does not prove. The example given is that at an illustrative $20k monthly ad spend, a $1,500 flat management fee costs less than a fee of 12 to 15% of spend, a calculation that does not prove either is a fair market price, while at $12k the comparison depends on which percentage is used.
Read more: The answer has to stand on its own
The article sorts the work into three jobs in order: be retrievable, be fetchable, be liftable. According to groas, a perfect paragraph absent from the relevant search cannot be retrieved, a retrievable page that fails on fetch cannot be read, and a readable page that never answers gives the assistant little reason to use it. The article recommends checking those three in sequence before buying another link package.
Read more: What the mechanism says to fix, in order
[1] https://groas.com/post/how-chatgpt-actually-picks-a-source-the-search-you-never-see (the article this page answers from)