How AI assistants search the web before they answer

Many AI assistants answer in four steps. They decide whether they need fresh information, turn your question into one or more web searches, read what comes back, and then write an answer, often with links to the sources they used. The answer you see is the last step of a longer process, and the earlier steps decide which brands can appear in it at all.

What happens between your question and the answer?

Here is the sequence in plain terms. The details differ between assistants, but the shape is shared.

1. You ask. For example, "What is the best project tool for a ten-person agency?"
2. The assistant decides whether to search. For a question about the current market, it usually will. For a question it can answer from what it already learned, it may not. See when an AI answers without searching.
3. It writes search queries. Your question becomes short searches, often several. These are not your words. See search queries vs. prompts.
4. A search service returns results. Pages, snippets, titles and addresses.
5. The assistant reads some of them. It does not use everything it finds. See pages found vs. pages cited.
6. It writes the answer. It combines what it read with what it already knows, names brands, and attaches citations.

Each step filters. A brand that is not in the search results cannot be recommended on the strength of them. A page that is found but never read cannot shape the answer.

Why does the search step matter for brands?

Because it changes what an assistant can say. An assistant that relies only on what it learned in training knows the world up to some past point. One that searches can reflect pages published recently and can pick up new products, price changes and fresh reviews.

It also moves influence outward. The assistant's judgment is partly a judgment about which pages it found and trusted. That is why the sources a category relies on, such as review sites, comparison pages and forums, tend to matter as much as a brand's own site. For a longer treatment, see the kinds of pages AI assistants cite most.

Is it the same for ChatGPT, Claude, Gemini and Perplexity?

The outline is the same. The behaviour is not.

• Whether to search is decided differently by each assistant. Some search for nearly every product question. Others sometimes skip it.
• How many searches and how much reading vary by assistant and by question.
• How sources are shown differs: inline links, a numbered list, a side panel, or nothing.
• How much is said differs too. Some assistants name a few brands, others name many.

We have separate guides for ChatGPT, Claude, Gemini and Perplexity. The broader point is that you cannot learn from one assistant how the others behave. Edition 0 of the Citewake index puts measurements on these differences; the summary is in what four AIs did when we asked them 180 buying questions.

What does a research lab do with this process?

It records it. A lab that asks an assistant through its official interface can keep more than the final text:

• the answer, word for word;
• the citations as the provider structured them;
• the searches the assistant ran, and the pages it surfaced, where the provider reports them;
• the model, its version and the settings.

That is the difference between "ChatGPT recommended X" and "ChatGPT ran three searches, found these pages, cited these two, and recommended X". The second can be examined. The first is a screenshot. The way the lab keeps these records is described in the methodology.

One design choice is worth knowing. A lab could run a single search itself and hand the same results to every assistant, which would make the comparison tidy. But then it would measure the lab's search, not the assistants'. Which pages each assistant finds and trusts is a large part of why they recommend different things, so a fair study lets each assistant do its own searching.

What does this mean if you want to be found?

Three practical points follow from the sequence.

• Be findable by the searches an assistant would write. Those are often plainer and more specific than your marketing language.
• Exist on the pages assistants read. Third-party roundups, reviews and community answers carry weight. See how to get onto the roundup and review pages AI relies on.
• Do not block the readers. If crawlers cannot fetch your pages, they cannot be read. See AI crawlers and robots.txt.

Why do the answers vary if the process is the same?

The process is the same in outline, but nearly every step contains a choice, and choices can go differently each time. The assistant may write different queries on a different day. A search service may return results in a different order. The assistant may read the second result in one run and the fourth in another. A small change early on can send the answer in a different direction, and that is a large part of why one question can give a different set of brands on a second asking. It is also why a careful study asks every question more than once and treats the spread as part of the finding, rather than hunting for the one answer that looks most typical.

How can you see this for yourself?

Ask an assistant a buying question with search switched on, and look at what it shows. Open the sources. Ask again in a fresh chat and compare. You will often see different sources and sometimes different brands, which is the reason a single answer is a poor guide. The Citewake checker gives a quick first picture if you would like one for your own site.