Follow one question through an assistant: "Which project management tool is best for a small agency?" The searches, pages and answer in this story are our own example; the study figures at the end are real.
Memory or search
First, the assistant decides how to answer.
It can answer from memory: what it learned in training, up to its cutoff date. Or it can search the web first. Questions about products, prices and anything recent usually push it to search.
Fan-out
One question becomes several searches.
Google calls this query fan-out: "breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf". Each search looks at the question from another side.
The index
The searches run against an index.
An index is a catalogue of pages that crawlers read ahead of time. Each search lights up the pages that match it; pages missing from the index cannot be found at all.
Reading
The model reads a handful of pages.
It takes in the passages that best match the question and leaves most of the results unread.
The answer
It writes the answer and cites some sources.
The names in the answer come from what it read. ChatGPT search, launched on 31 October 2024, answers with links to its sources, as do the other assistants that search.
What gets cited
Pages that match the question closely are cited more.
In Ahrefs' study of 1.4 million ChatGPT prompts (15 April 2026), the titles of cited pages were closer in meaning to the prompt than titles of pages it did not cite: 0.602 against 0.484, and 0.656 against its closest fan-out search.
Why answers move
Ask again, and different pages can light up.
Search results change, the model draws its words with some chance, and each assistant has its own index and ranking. Press the button to rerun our example with a different set of pages; it is a simulation, not real search data.
An assistant answers in one of two ways: from what it learned in training, or by searching first and writing from what it finds. Which sources it reads, and which names it mentions, follow from that choice.
The short version
From memory, an assistant can only repeat what its training data said, up to the date that data ends. With search, it turns a question into several searches, reads a handful of pages and writes an answer from them, often with links. Studies of what gets cited find that pages matching the question closely, and stating facts plainly, are used most.
Follow a question through
Figure 1From a question to an answer, when the assistant searches
1. The question arrives"Which project management tool is best for a small agency?"
2. Memory or search?The assistant decides whether it can answer from what it learned in training or needs current information.
3. The question fans outSeveral searches at once, for example: "project management software for small agencies", "agency project management tools compared", "project management tool pricing 2026".
4. Pages come backEach search returns pages from a search index, as titles, links and passages of text.
5. The model readsIt reads the passages that best match the question, and leaves most of the results unread.
6. The answer is writtenIt writes its answer from what it read, names the tools it judged relevant, and links to some of its sources.
3 / 6
A simplified picture. The searches in step 3 are our own example; each assistant decides differently when to search, and how they rank pages is not published.
From memory
Without a search, an assistant writes from what it learned in training. That knowledge stops at a cutoff date and reflects what was written before it. Pre-training explains how that memory forms.
With a search
When a question needs current or specific information, the assistant searches first. Google describes the method in its AI Mode as query fan-out: "breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf".[1] ChatGPT search, launched on 31 October 2024, answers with links to the sources it used.[2]
The search results come back as text. The model reads the most relevant passages and writes its answer from them, the same pattern as in Function calling, with search as the tool.
Which pages get used
The ranking rules inside each assistant are not published, so what is known comes from outside studies. Two recent ones point the same way.
Ahrefs studied 1.4 million ChatGPT prompts in April 2026. The titles of cited pages were closer in meaning to the prompt than the titles of pages ChatGPT did not cite, at 0.602 against 0.484 on a similarity scale, and closer still to at least one of its fan-out searches, at 0.656.[3]
A study posted to arXiv in April 2026 traced 21,143 citations from 602 prompts across ChatGPT, Google's AI Overviews and Gemini, and Perplexity. The pages the assistants drew on most heavily were "longer, more structured, semantically aligned, and richer in extractable evidence": definitions, numbers, comparisons and step-by-step instructions. Perplexity and Google cited more sources per answer; ChatGPT cited fewer but leaned on each one more.[4]
Both studies describe patterns, not rules. They show what cited pages tend to have in common, not that adding those features will get a page cited.
Why the same question gets different names
Three things make answers move. The model draws each word by weighted chance, as Temperature shows. The search results it reads change from day to day. And each assistant has its own index, its own ranking and its own feedback training. A brand named in one answer can be missing from the next.
What Visibian measures
Visibian asks seven assistants the questions a brand's buyers ask, on a schedule, and records whether the brand is named, how it is described and which sources are cited. What we measure, what we send to the AI providers and what we keep is on our transparency page.
From Visibian, which measures what seven AI assistants say about a brand. See what Visibian does.
