Does a spinning indicator mean the research is still progressing?

It shows activity in the interface, but it does not prove that useful findings are being produced. Record the last visible change, save the output you can access, and use a smaller request to investigate a repeatable failure.

Will changing the chat model fix Research access?

Not necessarily. Research uses its own designated models, and organization settings can restrict access to them. Ask your administrator to check the current Research requirements rather than assuming the model selected for ordinary chat controls Research.

Can I trust the citations in an unfinished report?

Open each source and check that it supports the nearby claim. Keep verified findings separately from unsupported statements and unresolved questions. An unfinished report should not become an approved conclusion just because it contains links.

Can another assistant continue after Claude fails?

Yes, if you provide a usable checkpoint: the question, scope, checked source links, unresolved issues, and next step. It cannot recover unseen output from Claude. A shared work record can provide the starting point, but the new assistant still needs access to any sources it must inspect.

Claude Research Doesn’t Work? Checks and Recovery Steps

When Claude Research will not start or finish, separate access problems from a stalled session. Save the useful findings before testing a smaller request.

September 21, 2026Updated October 21, 20266 min read

If Claude Research doesn’t work, check the error message, service status, Research access, and web search before retrying. If it starts but freezes or never produces a usable answer, save the visible work and test one smaller question in a fresh conversation. A slow response alone does not tell you which part failed.

Tools like fenbs keep the research question, checked findings, and next steps outside the stalled chat, so a retry can begin from a clear record. That helps preserve work; it does not repair Claude’s service or recover an answer you never received.

What should you check first?

Copy the exact error and note when it appeared, including your time zone. Record whether Research never started, stopped during searching, or failed while producing the answer.
Open the Claude status page (opens in a new tab) and look for an incident matching the affected app or connectors. An unrelated incident does not establish the cause of your failure.
Confirm you are signed into the intended account and workspace. Compare your access with Claude’s current Research documentation rather than assuming another account has the same capabilities.
Check that Research and web search are available for the conversation. For work accounts, ask the administrator about relevant organization restrictions.
Save any useful output. Then try a narrow public-web question with no private sources. Change one thing at a time so you can tell what helped.

Write down the result of each check. “A short public-web request completed, but the request using our internal documents failed” gives you a useful distinction. Repeatedly submitting the original question gives you less information and makes it harder to track which output belongs to which attempt.

Why won’t Research start?

Claude’s Research guide (opens in a new tab) describes eligible access and says Research requires web search. It also explains that Research uses designated models regardless of your chat selection; organization restrictions that disable those models can prevent it from working. Ask your administrator to check the current requirements on that page.

For web search itself, Claude’s web search instructions (opens in a new tab) describe both organization controls and chat settings. Some interfaces have a web search toggle, while the newer experience handles searching without that toggle. A missing switch therefore does not, by itself, prove that search is unavailable.

Once access is confirmed, explicitly ask Claude to use Research for a bounded question. For example: “Use Research to compare the published accessibility documentation for these two products. Use official sources and list anything you cannot verify.” This tests the research path without adding a large document collection or several unrelated goals.

How do you isolate a frozen run?

Before refreshing or closing anything, copy the question and any findings you can see. Note the last visible activity and whether the interface still responds. An activity indicator is a clue, not proof that useful research is still progressing. Avoid inventing a fixed deadline after which every run must be broken.

Use a small test to narrow the failure
What you observe	Next test	What to record
Research never begins	Try one short public-web question after checking access.	Whether the request starts and whether any error appears.
The interface appears stuck	After saving visible work, reopen the conversation or try another browser session.	Whether the same saved output appears and the controls respond.
Only a large request fails	Test one subquestion with fewer documents.	The smallest request that succeeds and the scope still unfinished.
Public sources work; internal sources do not	Test access to one known internal source separately.	The source requested, account used, and exact refusal or omission.

These are diagnostic comparisons, not guaranteed fixes. If a new browser session works, investigate the original session before changing unrelated settings. Save your work before clearing site data, since that can sign you out. On a managed device, involve your IT team when a test requires changing network or browser controls.

Why are connected sources missing?

Research that returns a web-only answer may be encountering a source restriction rather than freezing. The Research connector guidance (opens in a new tab) says Research started on mobile searches the web without using connectors, and local MCP servers configured in Claude Desktop are unavailable to Research. MCP is a standard way for AI apps to use outside tools.

For supported connected sources, check that the intended connector is enabled for the chat and name the source you want searched. Test a document you can open yourself. Ask Claude which source it actually consulted, then follow its citation. A confident summary without the requested source does not establish that the source was read.

Keep separate problems separate: use our Claude Desktop MCP checks for a failing desktop connection, Claude Code troubleshooting for coding sessions, and Claude Cowork troubleshooting for that workspace.

How can you retry without losing verified work?

Treat a partial answer as material to review, not a finished report. Open the original sources and keep only claims you can support. If a claim cannot be checked, label it unresolved. Do not ask the next session to accept a polished summary as evidence merely because the earlier session produced it.

Question and boundaries: the decision the research supports, the audience, and what is outside scope.
Verified findings: each useful claim, its source link, the relevant passage or section, and when you checked it.
Open questions: missing evidence, conflicting sources, and assumptions that still need review.
Recovery step: the failed attempt, the smallest successful test, and the next question to investigate.

For example, a design studio comparing booking systems might retain checked accessibility documents while leaving payment handling unresolved. Its next request should investigate that missing evidence, rather than repeat the whole comparison. Save that boundary in fenbs’s Where we left off; the AI handoff workflow helps the next connected assistant continue from it.

How fenbs helps

Keep the research checkpoint ready for a retry

Add the research task to fenbs and keep checked findings, source links, and unresolved questions with the project. Save Where we left off before changing conversations, including the exact next question. A connected assistant can read that record before continuing, so you can review the remaining work without rebuilding the explanation.

Create a free account

Free for one person · No credit card · Works with Claude, ChatGPT, Cursor and more

When should you stop retrying and ask for help?

Follow explicit messages before changing prompts. Claude’s error guide (opens in a new tab) distinguishes usage limits, excessive input length, capacity constraints, and service incidents. A usage refusal calls for the account’s stated next step; a length error calls for less input. Capacity constraints may occur without a listed outage, so a clear status page does not settle every failure.

If a small request still fails after access checks, collect a concise report: exact error, time, app or browser, workspace, whether connected sources were involved, and the tests already tried. Use Get help through Claude’s account menu (opens in a new tab). Exclude confidential document contents from screenshots unless your organization authorizes sharing them. Avoid endless retries that reproduce the same result without new evidence.

How to do this with fenbs

Create a free account, free for one person with no credit card, and add the research project to your kanban board.
Add a task for the research question. Include its boundaries and the recovery plan; keep checked findings and source links in project context notes.
Record the decisions a person has made in Decisions and rules. Keep unanswered research questions open, and save Where we left off with what failed and where to resume.
Connect a supported assistant using the integration guides. Ask it to read the project context before continuing and save progress before stopping.

fenbs is useful when findings become ongoing work or move between assistants. For one disposable question, a saved note may be enough. fenbs does not import Claude chats automatically or complete a failed Research run for you.

For different symptoms, see Gemini Deep Research recovery or ChatGPT memory troubleshooting. The AI at work hub covers other ways to keep useful work moving between sessions.