CLAUDE LABJP
FORK — Claude Code 2.1.212 changes what /fork does: it copies your conversation into a new background session with its own row in claude agents, so you can keep working. The old in-session subagent is now /subtaskLIMITS — WebSearch calls are now capped at 200 per session by default, and subagent spawns get the same 200 ceiling, so a runaway search or delegation loop stops on its ownMCPBG — MCP tool calls running past two minutes now move to the background automatically, keeping the session usable. Tune the threshold with CLAUDE_CODE_MCP_AUTO_BACKGROUND_MSPLANFIX — Fixed plan mode auto-running file-modifying Bash commands such as touch and rm without a permission prompt or an SDK canUseTool callbackSONNET5 — Claude Sonnet 5 is running on introductory pricing of $2 per million input tokens and $10 per million output. After August 31 it moves to $3 and $15IPO — Bankers are reportedly lining up investor meetings for Anthropic ahead of a possible public listing as soon as OctoberFORK — Claude Code 2.1.212 changes what /fork does: it copies your conversation into a new background session with its own row in claude agents, so you can keep working. The old in-session subagent is now /subtaskLIMITS — WebSearch calls are now capped at 200 per session by default, and subagent spawns get the same 200 ceiling, so a runaway search or delegation loop stops on its ownMCPBG — MCP tool calls running past two minutes now move to the background automatically, keeping the session usable. Tune the threshold with CLAUDE_CODE_MCP_AUTO_BACKGROUND_MSPLANFIX — Fixed plan mode auto-running file-modifying Bash commands such as touch and rm without a permission prompt or an SDK canUseTool callbackSONNET5 — Claude Sonnet 5 is running on introductory pricing of $2 per million input tokens and $10 per million output. After August 31 it moves to $3 and $15IPO — Bankers are reportedly lining up investor meetings for Anthropic ahead of a possible public listing as soon as October
Articles/Claude.ai
Claude.ai/2026-04-27Beginner

When Claude's Web Search Isn't Working: 7 Things Worth Checking

Walk through the seven most common reasons Claude's Web Search fails to return fresh results — from model selection and toggle state to robots.txt blocks and corporate proxies.

web-search4troubleshooting87claude-ai15beginner11

You enabled Web Search in Claude.ai, asked a perfectly reasonable question about something that happened last week, and Claude responded with cached training data and a polite note that it cannot access real-time information. I've hit this enough times across the four sites I run that I've stopped guessing — there's a checklist that catches the cause within a few minutes almost every time.

The seven items below are ordered by how often they're the actual culprit. If you start at the top and work down, most issues resolve before you reach item four.

1. Does the model you're using actually support Web Search?

The first thing to verify is your currently selected model. Web Search is not available across every Claude variant.

  • Claude Sonnet 4.6, Claude Opus 4.6, Claude Haiku 4.5 — all support Web Search
  • Claude Haiku 3.5 and earlier — Web Search is unavailable, and the tool icon stays greyed out
  • Some models running Extended Thinking — Web Search may be temporarily disabled while reasoning is in progress

Switch to Sonnet 4.6 or Opus 4.6 from the model selector and resend your prompt. In my experience, this single fix accounts for more than half of the incidents I see. The most common pattern is starting a chat on a lighter Haiku model for speed, getting deep into a conversation, and forgetting to upgrade once the question shifts to something that needs current information. The icon staying greyed out is the visual signal that the model itself is the limitation, not your settings.

2. Is the Web Search toggle actually on?

Look at the row of tool icons below the input box. The globe icon should be highlighted when Web Search is active. If it's faded, click it once to enable. The state is visually subtle, so it's worth zooming in or trusting the tooltip text rather than your peripheral vision.

The non-obvious trap here is project-level overrides. You can have Web Search enabled at the global level, but a specific Project may have the tool turned off in its own settings. Open Project settings → Tools and confirm Web Search is on for that project. Settings inside a project always win over your default account-wide preferences, which is great for keeping focused project contexts but easy to forget when you can't figure out why search isn't firing.

If you're using a workspace provided by your organization, an admin may have restricted tools at the workspace level. If your account refuses to enable Web Search even after toggling, that's the most likely cause and worth a short conversation with whoever manages your team's Claude deployment.

3. Is your prompt actually inviting Claude to search?

Claude only invokes Web Search when it judges that the question requires fresh information. Questions that look answerable from training data will be answered from training data — even when you'd prefer otherwise. The model is making a quiet cost-versus-benefit decision about whether searching is worth the latency.

Likely to skip search:
"Tell me about Anthropic's latest models."
→ Claude may answer from training data alone

Reliably triggers search:
"As of April 2026, list Anthropic's latest models with their release dates. Please search the web."
→ "search the web", "as of [date]", "release dates" force the lookup

Phrases like "search the web", "find the latest", "as of [month/year]", and "verify against the official site" produce noticeably more consistent behavior. I've added a line to my custom instructions that says to always use Web Search when factual currency matters, so I don't have to repeat it on every prompt. If you're a heavy researcher, that single sentence will save you a remarkable amount of friction over weeks of use.

The corollary is that you can also explicitly tell Claude not to search when you want a faster response from training memory alone. The toggle plus prompt language together give you fine-grained control.

4. Have you hit a rate limit?

Claude.ai's Web Search has per-hour and per-day quotas that vary by plan. The Free tier has fairly tight limits, and a burst of research-style prompts can lock you out temporarily without much warning.

The symptoms split into two flavors. Sometimes you get an explicit notice that the rate limit has been reached. Other times — and this is the more dangerous case — Web Search silently skips, and Claude responds from training data only. The reply still looks confident and well-formed, so you can easily ship the result downstream without realizing it's stale. Always sanity-check whether the answer actually contains current information when freshness matters.

Two ways out:

  • Wait an hour or two for the limit to reset automatically
  • Upgrade to Pro or Max for substantially higher quotas

If you research-search dozens of times a day like I do, Pro is the practical baseline. The math is straightforward: a single afternoon of competitive research can easily burn the Free tier's daily allowance, and the time you spend waiting out the lockout is worth more than the subscription cost.

5. Is the target site blocking AI crawlers via robots.txt?

This one is easy to miss. Claude's Web Search respects each site's robots.txt. If a site blocks AI crawlers — Anthropic-ai, ClaudeBot, and similar — its content won't appear in your search results, no matter how cleanly you phrase the prompt.

When a single site refuses to load, open https://example.com/robots.txt in a browser and look for entries like:

User-agent: Anthropic-ai
Disallow: /

User-agent: ClaudeBot
Disallow: /

If those rules are present, that site has opted out of AI search. Your workaround is to ask Claude to consult several alternative sources instead — news outlets, official posts on X, GitHub Discussions, related blog coverage — so the model has multiple paths to the same information. Casting a wider net usually surfaces the underlying facts even when one specific source is off-limits.

6. Is your network or proxy interfering?

When you use Claude from a corporate or campus network, enterprise proxies and SSL inspection can quietly block the backend calls that Web Search relies on. The chat itself loads fine — it's only the search-time API calls that fail — which makes this category of issue particularly confusing to diagnose.

A quick triage:

  • Try the same query from a home Wi-Fi network
  • Try again over your phone's hotspot
  • If those work and your office network doesn't, the problem is on the corporate side

The cleanest fix is to ask your IT team to allowlist api.anthropic.com, claude.ai, and the search backend domains. A short justification — Claude.ai is a generative AI service from Anthropic, used to accelerate research workflows — usually moves the request along. If your organization is already using other AI tools, point to those existing approvals as precedent.

7. Final checks before giving up

If none of the above changes anything, work through these one by one. They cover the long tail of less common issues:

  • Clear your browser's cache and cookies — and site data, not just cookies
  • Open Claude in a different browser (Chrome → Safari, Edge → Firefox)
  • Try a private or incognito window to rule out extension interference
  • Check status.anthropic.com for ongoing incidents

If the status page shows an active issue with the search component, there's nothing to do on your end except wait. Anthropic's uptime has been strong in recent months, but every few months there's a multi-minute window where things genuinely are broken upstream. Bookmarking the status page alongside this article saves a lot of self-doubt during those rare outages.

A quick way to confirm Web Search actually ran

After tweaking any of the items above, the fastest way to verify the fix worked is to check whether Claude's response includes inline citations or a "Sources" footer. When Web Search runs successfully, the response either cites URLs directly in the text or appends a list of source links. If those are absent and the answer still feels current, double-check by asking Claude to paste the URLs it consulted. Claude will either confirm it searched and cite the sources, or it will tell you it answered from memory — which is the cleanest possible diagnostic when you're not sure whether your fix landed.

I keep a tiny test prompt for this: "What was today's most-read article on Hacker News? Please search the web and cite your sources." If Web Search is healthy, it returns within a few seconds with a clear citation. If something's still wrong, the failure mode is usually obvious from the response alone.

One concrete next step

In my experience, Web Search failures cluster in three places: model selection, the toggle, and how the prompt is phrased. Check those three first, every time, and you'll resolve the majority of incidents before needing to dig further. Bookmark this article — the next time Web Search misbehaves, work the list from the top.

Share

Thank You for Reading

Claude Lab is ad-free, supported entirely by members like you. We publish practical guides daily with implementation code, benchmarks, and production-ready patterns. If you've found it useful, we'd love to have you on board.

  • Copy-paste ready implementation code
  • New advanced guides published daily
  • $5/mo or $10 for lifetime access
View Membership →

If you found this article helpful, a small tip ($1.50) would mean a lot to us. Your support helps keep this site ad-free and covers server and hosting costs.

Related Articles

Claude.ai2026-05-05
Claude Keeps Making the Same Mistake — Why Corrections Don't Stick and How to Fix It
Why Claude repeats mistakes even after correction, and how to make your instructions stick permanently using Projects, custom prompts, and smarter conversation design.
Claude.ai2026-05-01
When Claude won't send your message: a diagnostic order that always finds the cause
When the send button does nothing, the spinner runs forever, or your message vanishes — here's the order I work through to find the real cause within five minutes.
Claude.ai2026-04-23
When Claude Says 'Failed to Send Message' — A Practical Triage and Recovery Guide
Why Claude shows 'Failed to send message' — broken down into five real causes, with a quick triage order and practical fixes for each.
📚RECOMMENDED BOOKS
Build a Large Language Model (From Scratch)
Sebastian Raschka
LLM Dev
Prompt Engineering for LLMs
Berryman & Ziegler
Prompting
AI Engineering
Chip Huyen
AI Eng
* Contains affiliate links
See all →