When an engine answers without searching

Not every AI answer involves a search. Some engines decide, question by question, whether to look anything up. If they do not, the answer comes from what the model absorbed during training rather than from the web this week.

Keylight records how many searches happened behind every answer, and the evidence page says plainly when there were none. It says the engine answered from what it already knew, and that this reflects its training rather than the web this week.

Why we do not force a search

We could tell the engines to search every time. We deliberately do not.

A real person asking Gemini for the best tool in a category is also answered from memory, and also sees no sources. Reporting what a real person is told is the entire job. Forcing a search would make the engines neatly comparable and would stop the question being the one a real customer asks.

Which engines do what

Perplexity and Google AI Overviews search by definition. Claude, ChatGPT and Gemini each decide per question. DeepSeek cannot search at all, so every DeepSeek answer reflects training rather than the current web, and every DeepSeek answer says so.

What it means for your score

Two answers that look alike can be different kinds of evidence. An appearance won from training says the model absorbed your brand at some point. An appearance won from a search says the web said so this week. Both are real signals and they are not the same signal, which is why the count is on every answer rather than buried in a methodology note.