The short answer
Web search is on by default. The model you selected answers from current sources instead of only from its training data, it costs no credits, and the toggle is per request rather than per model, so it applies to whichever model you pick.
The toggle in the composer is permission rather than a command. With it on, Whizi asks the model whether this particular question needs the web before it spends a search, so a request for a poem does not trigger one and a question about this week does. Switching it off tells the model it has no way to look anything up, which is worth doing when you want an answer drawn only from the conversation.
The per-request part is the useful bit. A model with no retrieval of its own normally answers a question about this month from a training set that ended months ago. The toggle puts a search layer in front of whichever model you picked, which means you can ask Claude, or a Llama model, or a small fast model, a question about today and get a sourced answer.
If the toggle is on but an answer came back unsourced, or the message errored, the failure surface is in web search not working.
When it changes the answer
Three categories where the difference between search on and search off is not subtle.
Anything with a date in it. Prices, model releases, product launches, company news, version numbers. Without search, a model answers confidently from whatever was true when it was trained, and gives you no signal that the information is stale. This is the single most common way AI answers go quietly wrong.
Anything obscure or specific. A small company, a niche tool, a local regulation, a particular error message. Training data covers these thinly or not at all, and thin coverage is exactly the condition under which models invent plausible detail.
Anything you intend to act on. Search produces sources you can open. A claim with a link you can check is a different object from a claim without one, even when both turn out to be correct.
When to leave it off
Search is not free of cost to answer quality, which is why the toggle exists at all.
For reasoning, writing and analysis of something you supplied, search adds nothing and can actively distract. If you have uploaded a document and want the model to reason about that document, pulling in web results introduces material that is not yours and was not asked about. Same for creative writing, code review of your own code, and working through a problem step by step.
The rough rule: switch it on when the answer depends on facts about the world, switch it off when the answer depends on thinking about something already in the conversation.
| What you are asking about | Leave search on | Why |
|---|---|---|
| Prices, releases, launches, company news, version numbers | On | Without it the model answers from training data and gives no signal that the information is stale |
| A small company, a niche tool, a local regulation, a particular error message | On | Training data covers these thinly, and thin coverage is when models invent plausible detail |
| Anything you intend to act on | On | It produces sources you can open, which is a different object from a claim without one |
| Reasoning about a document you uploaded | Off | Retrieved pages introduce material that is not yours and was not asked about |
| Creative writing, or code review of your own code | Off | The answer depends on thinking about something already in the conversation |
Reading the sources properly
A sourced answer is more checkable, not automatically more correct. Two failure modes survive retrieval and are worth watching for.
The first is a source that exists but does not support the claim. The model retrieved a page, and the page is real, but the specific number or statement attributed to it is not in it. Opening one or two cited links on any answer you plan to rely on catches this quickly.
The second is a source that is itself wrong or out of date. Search does not evaluate credibility, so a stale blog post ranks like an authoritative one. For prices and specifications especially, a vendor page beats a roundup article, and both beat a forum comment.
If a claim matters, the fastest verification is to ask a second model the same question with search on, in the same conversation. Where two independent retrievals agree on a specific number and a source, confidence is reasonable. Where they diverge, that is your finding.
- The toggle is in the composer, next to attach, and it is on by default
- Web search costs no credits
- The toggle is per request, not per model, so it works with whichever model you pick
- Whizi searches only on the questions that need it, not on every message
- Switch it off for reasoning about a document you uploaded
- Open at least one cited link on any answer that matters
- Cross-check important claims by asking a second model with search on
Frequently asked questions
Can Claude search the web in Whizi?
Yes. Web search applies to whichever model you selected, so Claude Sonnet 5 or Claude Opus 5 answers from current sources. This is one of the practical advantages of a multi-model workspace: you are not limited to whichever models happen to ship with their own retrieval.
Does web search use extra credits?
No. A message costs the same whether search is on or off, because the credit rate is set by the model alone. The only thing that changes the cost of a message is which model answers it.
Why did the model give me outdated information?
Either search is switched off in that conversation, or the model judged the question answerable without it and answered from training data. Check the toggle first. If it is on and the answer carries no sources, ask again and say you want it looked up; if the answer changes, the first one was stale.
Should I leave web search on all the time?
For most work, yes, and it is on by default: Whizi only searches when a question needs the web, so leaving it on costs nothing on the messages that do not. Switch it off for a conversation about material already in the chat, such as a document you uploaded, where retrieved pages pull the model away from your own file.