ChatGPT Deep Research Not Working? Checks and Fixes

If ChatGPT deep research is not working, first identify where it stops: finding the feature, accessing sources, running the investigation, or producing the report. Test that stage with one small question before repeating your entire request. This helps you avoid spending another session on the same unresolved cause.

Keep the research question, verified source links, and remaining checks outside the failing conversation. fenbs can hold that work record while you troubleshoot, so a stopped report does not erase what you already established.

What should you check first?

1. Copy your question, required sources, and any usable findings into a separate record. Save the exact error and where the task stopped.
2. Confirm the signed-in account and workspace. Look for an access message, a usage notice, or a question waiting for your response.
3. Try one narrow public-source question without private files. Ask for a short answer and its source rather than a large report.
4. If the small task works, add your original sources one at a time. Check the first source or instruction that changes the result.
5. If the small task also fails, try a fresh conversation or another supported surface. Record the outcome before asking support or your administrator for help.

Treat this as a diagnosis, not a promise that refreshing will repair every failure. A source-access problem needs a different response from a missing feature. Changing your prompt cannot grant permission to a private document, and switching browsers cannot increase an account allowance.

Why can’t you find the deep research option?

Check the instructions for your surface. The official OpenAI documentation↗ (opens in a new tab) describes a Deep research plugin in the desktop app: switch to Work, open Plugins, and enable or install it if available. On the web, its instructions describe selecting Deep research from the + menu in Work when your account and workspace offer it.

Availability depends on account and workspace settings. Read an on-screen restriction as evidence, rather than assuming your account should match a screenshot someone posted. Ask a workspace administrator to check an organizational restriction; do not switch to a personal account to bypass a workplace data rule.

A missing research option is also different from an app that cannot answer anything. Use the separate general ChatGPT troubleshooting guide for broader failures; memory issues and project issues have their own checks.

Why are sources missing or unreadable?

Open the required source yourself and verify that the URL still points to the intended page. Check whether it needs a sign-in, has moved, or is only a preview. Your own ability to read a page does not prove the research task received the same access.

Ask ChatGPT which requested sources it actually accessed and which it could not use. Compare its answer with the report. A citation to a publisher’s abstract is not evidence that the full paper was read, and a company homepage cannot support a detailed claim buried in a separate filing.

If your organization permits it, attach an authorized copy of the relevant material and name the section needed. Keep the original URL and version beside it. When a connected source is the only failing part, the ChatGPT connector checks address that narrower problem.

Preserve the accessible sources and the missing ones separately in fenbs Links and listings, with a task for obtaining each missing document.

What should you do when a run seems stuck?

Inspect the latest visible activity and look for a clarification or approval request. An investigation with several sources can require more work than a short search answer, but elapsed time alone does not tell you whether progress has stopped. Avoid an arbitrary time limit that treats every long task as broken.

If you decide to stop, save visible findings first. Restart with a smaller question, fewer required sources, and a defined output. For example, replace “Research warehouse automation” with “Find primary evidence about picking errors in US warehouses; report the measurement method and limitations for each source.”

Change one thing at a time. If a public-source task works but adding a particular file repeatedly fails, record that contrast. A concise reproduction gives support something concrete to investigate, while repeated full reruns hide the difference between a temporary interruption and a consistent input problem.

How do you recover an incomplete or failed report?

Separate missing output from weak evidence. A report with headings but no support needs source verification. A missing download needs an output recovery attempt. Ask for the available findings in plain text, including source URLs and unresolved questions, before requesting another formatted file.

OpenAI’s guide to using ChatGPT↗ (opens in a new tab) recommends checking important claims, current source material, and missing information. Apply that review even when the retry succeeds. A polished replacement is not proof that earlier access gaps disappeared.

For a hypothetical purchasing review, keep “vendor documentation confirms compatibility” apart from “likely lower implementation effort.” The first needs the actual documentation; the second is a conclusion that may depend on assumptions. Do not turn either into an approved purchasing decision while recovering the file.

How to do this with fenbs

1. Create a free account, then add a project named for the research question. One person can start free, with no credit card.
2. Save the original question and checked source URLs. Label unverified findings clearly and keep sensitive documents in their approved storage.
3. Add separate tasks for the failing step and the unanswered research questions. Record what you tried and the observed result.
4. Save Where we left off with the last checked finding and the next diagnostic step. A supported connected assistant can read that context before continuing.

fenbs does not repair ChatGPT, restore an unavailable feature, or import the failed conversation. It fits when you need a dependable research record across sessions; a single disposable question may need only a short local note.

When should you escalate instead of retrying?

Escalate a repeatable failure after you have a small example. Include the surface, account type without private identifiers, error text, approximate time, and whether a simple public-source question worked. Share a sanitized screenshot if useful; exclude source documents or private report content unless the support channel is approved for them.

If you change tools while waiting, carry over only reviewed findings and open questions. Gemini research troubleshooting covers that product’s failures; the deep research workflow covers planning a successful investigation. Our AI at work hub has related work guides, while Claude Cowork troubleshooting addresses a different assistant environment.