Yes. Claude can search the web, in the Claude app, through the API and inside Claude Code. It does not search on every request though. Claude holds a large amount of knowledge from training and only reaches for live retrieval when the question calls for current information.
We measured what that retrieval looks like in practice. On 13 July 2026 we ran 20 buyer-intent prompts three times each through four engines via API (gpt-5-mini, gemini-2.5-flash, claude-haiku-4-5 and Perplexity sonar), 240 answers in total. Claude returned sources on 100% of its 60 answers, averaging 8.8 sources per answer drawn from 194 distinct domains. It was also the most stable engine of the four: it changed its top recommendation between identical consecutive runs only 28% of the time.
One caveat up front: those citation rates come from the API. Consumer apps may show fewer sources, or none, even when retrieval happened behind the scenes.
When Claude searches and when it answers from memory
Claude has two ways to answer. Training memory covers everything it learned before its knowledge cutoff: stable facts, concepts, how-to knowledge. Retrieval covers everything else: current prices, recent releases, "best X in 2026" comparisons, anything where a stale answer would be wrong.
The model decides per query. Ask it to explain a marketing concept and it answers from memory with no sources. Ask it to recommend software for your team and it searches, because recommendation queries are exactly where training data goes stale fastest.
This matters for interpreting sources. When Claude cites a page, that page shaped the answer you just read. Claude and Perplexity expose their citations openly, while ChatGPT and Gemini often hide theirs. In our test, Gemini wrapped 429 of its 580 source URLs (74%) in vertexaisearch.cloud.google.com grounding redirects that mask the real source. Claude's citations pointed at real domains.
Can Claude Code search the web?
Yes. Claude Code, Anthropic's command-line coding agent, can search the web and fetch pages when its web tools are enabled. It uses them the same way the chat product does: it decides when a task needs current information, such as looking up a library's latest API, checking documentation or verifying an error message against recent forum threads.
The tools are permission-gated. If your organisation has disabled them, Claude Code answers purely from training memory and will usually say so when a question needs fresher data. If you are comparing agents or assistants for a team, this is worth checking in your own environment rather than assuming.
The same retrieval-versus-memory logic applies: boilerplate code comes from memory, while anything version-sensitive should trigger a search.
What 240 answers show about Claude's retrieval
Our 20 prompts were buyer-intent questions, the kind that trigger search: which tool, which provider, which product. Here is how the four engines compared.
| Engine (model) | Answers with sources (API) | Avg sources per answer | Distinct domains cited | Top-pick change rate between identical runs | Brand-set overlap between runs |
|---|---|---|---|---|---|
| Claude (claude-haiku-4-5) | 100% | 8.8 | 194 | 28% | 67% |
| ChatGPT (gpt-5-mini) | 100% | 15.0 | 445 | 35% | 42% |
| Gemini (gemini-2.5-flash) | 100% | 10.7 | 109* | 44% | 54% |
| Perplexity (sonar) | 100% | 8.3 | 142 | 43% | 61% |
*Gemini's domain count is deflated by its redirect wrapping: 74% of its source URLs hide the real domain behind a Google grounding redirect.
Want to see this in action?
See how every major AI model talks about your brand. Free to start.
Claude's citation pattern stands out for how flat it is. Its 529 citations spread across 194 domains, and the top three domains account for just 6% of all citations. Compare Perplexity, where Reddit alone took 71 of 498 citations (14%) and YouTube another 40 (8%). ChatGPT cast the widest net, 900 citations across 445 domains, but paired that with the least consistent brand sets: only 42% overlap between identical runs.
One pattern held across all four engines: Forbes was the only domain to appear in every engine's top-cited list. Beyond that, the engines barely agree. Pairs of engines named the same top brand between 20% of the time (ChatGPT and Perplexity) and 53% (Gemini and Claude). If you have read our Claude vs ChatGPT comparison, this will not surprise you: they are running different retrieval stacks over different source pools. ChatGPT's retrieval behaviour has its own quirks, which we covered in can ChatGPT search the internet.
The most stable engine of the four
Top-pick change rate
How often the top recommendation changes between identical runs
SparkToro found the same AI query changes its answer roughly 70% of the time. Our run-to-run numbers were lower but confirm the direction: Gemini changed its top pick 44% of the time between identical consecutive runs, Perplexity 43%, ChatGPT 35% and Claude 28%.
Claude also kept 67% of its recommended brand set between runs, the highest of the four. That combination, most stable top pick plus most stable brand set plus fully exposed citations, makes Claude the cleanest engine to establish a visibility baseline on. If Claude stops recommending you, that is more likely a real signal than noise. On ChatGPT, a single disappearance could just be the 58% of the brand set that churns anyway.
Stability cuts both ways though. A single manual check on any engine is still a coin flip, which is why spot-checking fails as a measurement method. Even the steadiest engine changed its top pick more than a quarter of the time.
What this means for your brand
Three practical takeaways from the data.
First, Claude's flat citation distribution (top three domains at 6% of citations) means there is no single gatekeeper site to win. Broad third-party coverage matters more than any one placement, which matches Ahrefs' finding that AI visibility correlates most with third-party mentions and video rather than on-page work. Semrush adds a sharper warning: 62% of AI citations never name the brand, so being cited and being recommended are different problems. Our guide on how to get cited by AI covers the mechanics.
Second, measure on repeated runs, not single checks. Our whole dataset is 3 runs per prompt, and even that only samples the variance.
Third, if you track one engine as your baseline, Claude is the defensible choice: sources on every API answer, real domains behind the citations and the lowest churn of the four. We built Honeyb, our AI visibility platform, around exactly this kind of repeated multi-engine measurement.
Want to know whether Claude, ChatGPT, Gemini and Perplexity recommend you today? Run a free AI visibility check.
