Agentic Search

Agentic search is what happens when an assistant answers a question by running its own sequence of searches. It reads what comes back, decides that is not enough, searches again, and only then answers, instead of returning one ranked list and stopping. The user asks once; the system may search a dozen times.

What the system is actually doing

A classic search engine takes a query and returns documents ranked against it. The decision about what is relevant is made once, and the person does the rest of the work. An agent inverts that: it treats the question as a goal, and searching as one of several moves available to it. It will reformulate the question, search on the reformulation, notice that two sources disagree, search a third time to break the tie, and only then write an answer.

The practical consequence is that a page is no longer competing for one position against one query. It is competing to be retrieved by any of several queries the system invented, and then to survive being read alongside the other sources it was retrieved with. Being ranked first for the phrase the person typed is neither necessary nor sufficient.

What it changes for a page

Two things get more valuable and one gets less. More valuable: being unambiguous about what the page is and who it is from, because an agent comparing sources resolves conflicts partly on which source it can identify; and being right about specifics, because a claim that contradicts the rest of the retrieved set is the claim that gets dropped.

Less valuable: the parts of a page that exist to hold a visitor’s attention. An agent does not scroll, does not bounce, and is unmoved by the thing above the fold. It reads what was served and moves on. A page built entirely around holding a human reader will be read by an agent as the two paragraphs of substance it contains.

One question people ask

Is this different from an AI assistant just using a search engine?

In degree, and the degree matters. An assistant that runs one search and summarises the top results is still doing ranked retrieval with a summary on top. An agent that searches, evaluates, searches again on what it learned, and stops when it judges it has enough is running a loop whose length depends on the question. The second is much harder to be absent from by accident and much harder to game, because there is no single query to target.

Sources

This entry describes the mechanism in general terms. Nothing on this page reports a measured figure; where this site reports numbers about crawler behaviour, the method is on the methodology page.