When should an AI search the web instead of answering from model knowledge?

An AI should search the web when the answer depends on information that may have changed since the model was trained. The deciding factor is not whether the topic sounds difficult. It is whether freshness, exact attribution, or a specific external document matters.

Current information almost always needs search. Examples include market prices, weather, sports results, opening hours, laws, government guidance, software releases, product specifications, company leadership, travel restrictions, event schedules, and breaking news. Even a small change can make an otherwise polished answer wrong.

Search is also appropriate when the user names a particular page, paper, dataset, filing, product listing, or document that has not been provided in the conversation. The model should inspect that source rather than reconstruct it from memory. The same applies when the user asks for direct quotations or precise citations.

Stable knowledge often does not need live search. Explaining long division, rewriting a paragraph, brainstorming names, outlining a meeting agenda, or describing a well-established historical concept can usually be handled from model knowledge. Searching every request can add delay and irrelevant material without improving the result.

A good search-backed answer shows its work. It distinguishes facts found in sources from the model's inference, links directly to the pages used, and notes the date when freshness matters. It should prefer primary and authoritative sources. For news, it should distinguish the publication date from the date the event occurred.

Search does not remove the need for judgment. Results can be outdated, duplicated, promotional, or simply wrong. The model still needs to compare sources, resolve definitions, and say when evidence is incomplete. If no reliable source supports a claim, the correct response is uncertainty, not a fabricated citation.