AI Visibility Academy · Chapter 01

How an AI assistant builds an answer

2 min read · · CiteCue Team

An AI assistant answers in one of two ways. It writes from what the model learned in training, or it searches the web, reads some of the pages that come back and writes from those, linking the ones it used. Usually the assistant decides which route to take, and only the second one gives your website a direct say in the answer.

The steps, in order

Anthropic's documentation for Claude's web search tool lays the loop out plainly, and other assistants with live search follow a broadly similar pattern:

  1. Decide whether to search. Claude searches when a question depends on current or changing information, including facts about specific companies or products. Questions about stable knowledge it answers directly.
  2. Turn the question into searches. ChatGPT's help center says it rewrites your question into one or more targeted queries for its search providers, and may send narrower follow-ups after reading the first results.
  3. Read what comes back. The assistant takes in the results and can search again, several times in one answer, if they fall short.
  4. Write and cite. It writes one response and attaches links to the sources it drew on.

Google adds a step before all of this. It says AI Overviews only appear when its systems judge them additive to the classic results, so many searches never show one.

Which steps you can influence

Step Your influence
Whether it searches None. That depends on the question and the assistant.
What it searches for Indirect. You can't see the queries, but you can cover the questions around your topic.
Which pages it finds Real. A page has to be crawlable and indexed to be found at all.
What it takes from them Real. A clear answer in a sentence that stands on its own is easy to use.
What it knows from training Slow. It reflects what was written about you before the model's cutoff.

The model, its settings and the person's wording are out of your hands. The pages an assistant can reach, and how plainly they answer, are in yours. That is where your effort pays off.

Where to start

Pick three questions a buyer might put to an assistant before choosing a product like yours. Ask each one with web search on and open every cited link. For each answer, note three things: whether a page of yours was used, whether a rival's page was, and whether the answer came with no links at all, which suggests the model answered from what it already knew.

That one exercise shows which step you're losing at. Two related lessons go further: trained knowledge versus live search, and the hidden searches behind one question.

Sources

See where AI cites you, and where it doesn’t