Switch gone: Claude fetches fresh pages only when the task needs them

The switch is gone, the model decides

The old routine asked a lot from the user: open settings, find the search switch , turn it on, and remember it exists. Most people never did, so the assistant answered from older training knowledge while nobody noticed the gap. It felt like sitting in a dim room and forgetting the lamp has a switch .

There is no switch in the new routine. I type a question, and Claude works out whether fresh data is needed. When it is needed, it fetches it; when it is not, it relies on stored knowledge. Think of a skilled librarian who decides which question sends them to the shelves and which one they can answer from memory. No settings, no forgotten control, no stale answer.

Anthropic started rolling this out to Pro and Max users on September 16th, 2026. A missing control in settings is the sign you are on the new version. The Claude help center confirms the same point: in the new experience there is no toggle, and Claude browses when it helps. I find that small detail telling, because quietly shipped changes tend to last.

Research mode writes the file instead of glancing at pages

The second layer is research mode. Regular search means the assistant glances at a few pages and answers. Research mode follows the question through chained searches, corrects its course as it finds material, and ends with a cited, structured file. ArsTechnica reported last year that on tangled questions this work can run for up to 45 minutes and sweep through hundreds of sources. One is a quick glance, the other lands on the desk as a file.

Starting it takes one step: type /deep-research in the chat box or pick research from the plus button. It ships on Pro, Max, Team, and Enterprise plans, not on the free plan. The Claude help center notes that research mode can also draw on connected mail and documents. A broad question is therefore treated less like a message and more like a small project.

So which one fits which job? A single fresh fact is enough for automatic search. Mapping every tool that launched in the past month and picking the useful ones calls for research mode. The Taku tech journal makes this split clear: the assistant decides alone whether a search is needed, and the user cannot force one. My lesson is simple: narrowing the question is my job, browsing is its job.

The agent setup described in the video runs on the same logic. Keyword work goes like this: Claude pulls fresh search data, spots gaps in existing content, and says which piece to write next. The example claims this routine reached 1,300 clicks a day. I care less about the number and more about the discipline: an aide that asks the same question every day and files the answer in a table.

In the content layer, the assistant places found terms next to our own background and drafts from that mix. Every claim arrives with citations , so checking before publishing gets easier. The video points to updated Google guidelines dated October 1st, 2026, saying every machine-written line needs a human review before release. I keep that step: the machine drafts, I sign off.

Then comes the community layer: the assistant watches what members ask most and predicts which tutorial should open next. That used to take a full-time person. The SwapBiswas blog lands on the same point while explaining the two modes: broad jobs need five or more search steps, and the report arrives at the end. As the crowd grows, the assistant becomes more valuable.

The picture in front of me is plain: most owners still search by hand, hunt for switches, and never open research mode. That window will not stay open forever. I would start small: ask research mode one question about my field each morning and read the file over breakfast. Six months from now, the gap will separate the hardest workers from the earliest builders.