How can I tell when AI searched the web?
The difference between remembering and searching.
You ask about a company's latest results.
One answer comes from what the model learned in training. The other comes from a web page it found thirty seconds ago.
They look identical: same tone, same confidence, same neat paragraphs. But one could be a year out of date and partly invented, and the other is based on something real.
Recall: answering from what the model learned in training. It's fast, and it's all the model has on its own.
Search: the app runs a web search, puts the pages it finds on the desk, and the model answers from them.
Search isn't part of the model. It's a tool the app gives it, like handing someone a phone so they can check.
Signs it searched:
No sign of any of these? Assume it's recall.
Simon Willison, a programmer who writes a lot about AI tools, noticed that search finally became genuinely useful in 2025.
The change was that newer models can search several times while they work things out: search, read, notice a gap, search again. Earlier models searched once and hoped for the best.
Willison's complaint about one popular app was that it "doesn't show me what it's searching for".
That matters. If you can see the searches, you can spot a bad one. Maybe it searched for the wrong company or the wrong year. If you can't see them, you're trusting it blind.
Even when it does search, it can:
That's why the three-click check in "AI's sources need checking too" still applies.
Ask for a search when facts matter. "Search the web for this and show me your sources."
Look for the signs. No sources shown means it came from memory.
Prefer tools that show their searches. Glance at what it searched for. A bad search means a bad answer.