Deep Research is ChatGPT’s background research agent: give it one prompt and it spends roughly 5–30 minutes browsing dozens of web pages and PDFs to produce a cited report. Its failures look different from a normal chat. Usually nothing visibly breaks — you just get a report that “looks complete but is full of generic statements,” a much thinner report than expected, or a hard Deep research request failed error.
Fastest fix: before blaming the prompt, confirm you’re actually getting the full version. As of June 2026, once you hit your monthly full-research cap, ChatGPT silently routes you to a lightweight Deep Research (an o4-mini-class model) that still browses but returns shorter, shallower reports with fewer sources. That single fact explains most “it suddenly got worse” complaints. Triage in this order: server status, quota tier, prompt quality, then blocked sites.
How to start it correctly (so it doesn’t quietly fail)
Two workflow facts trip people up before any “bug” does:
- Access path (as of June 2026): in the message composer, click the
+(Tools) menu and choose Deep research. There is no longer a standalone toolbar button — if you don’t pick it from that menu, you get a normal chat reply that only looks like research. - The clarifying-questions step is mandatory and load-bearing. Real Deep Research almost always replies first with 2–4 clarifying questions and a short, editable research plan before it goes away to browse. If your task started browsing instantly with no questions, you most likely didn’t actually trigger Deep Research (see the diagnosis table below). Answer the questions; a vague answer here is the #1 cause of a generic report.
Common causes
In rough order of frequency:
1. Monthly quota exhausted — silently downgraded to lightweight
Deep Research has a per-month cap by plan tier. As of June 2026 the published allowances are:
| Plan | Full Deep Research / month | Lightweight / month |
|---|---|---|
| Free ($0) | 0 | ~5 |
| Plus ($20) / Team | ~10 | +15 |
| Pro ($100) | ~50 | (shared/higher) |
| Pro ($200) | ~125 | +125 |
Past the full cap, the Deep research option still works, but ChatGPT automatically switches to the lightweight version. OpenAI’s own note: “Responses will typically be shorter while maintaining the depth and quality you’ve come to expect.” In practice that means fewer browsed sources and less synthesis — which reads as “it got dumber.”
How to verify: open avatar → Settings → look for your Deep Research / agent usage, or just compare the report against the diagnosis table in Step 1. A real full run is typically 1,500–4,000 words with 10–40 sources; a lightweight run is noticeably shorter with fewer.
2. Server-side incident — “Deep research request failed” / stuck on “Researching…”
This is a genuine, recurring failure that has nothing to do with your prompt. The task gets stuck on “Researching…”, or returns Deep research request failed, or shows a Starting… 0 button that never moves. During past incidents, failed runs also deducted quota even though no report was produced.
How to verify: check status.openai.com for an open “Increased error rates for Deep Research” incident. If one is posted, it’s not you — wait it out. If a failed run ate a credit during a confirmed incident, you can request a credit restore through ChatGPT support.
3. Prompt too broad — agent gives up early
Internally, Deep Research first does a “task decomposition,” splitting your prompt into ~5–10 sub-queries. If your prompt has no decomposable dimensions (e.g. “research AI for me”), the decomposer outputs heavily overlapping sub-tasks and the agent stops after a few rounds because marginal information gain is too low.
How to verify: look at the Sources count at the bottom of the report. Fewer than ~8, mostly from one domain (all Wikipedia / Medium), means it quit early.
4. Target sites block the crawler (robots.txt / Cloudflare)
OpenAI’s OAI-SearchBot and ChatGPT-User user agents are blocked by robots.txt or Cloudflare firewalls on some sites (X/Twitter, LinkedIn, Quora, Substack paywalls, corporate intranets, some news sites). Deep Research doesn’t stop when it can’t reach a page — those sources just go missing from the report, biasing the conclusion toward what it could reach.
How to verify: if a site you’d expect to appear (a leading industry blog) never shows up in Sources, its server is likely rejecting the crawler.
5. Output language mismatched with prompt language
If your prompt is in Chinese but you want an English report (or vice versa), the decomposed sub-queries can pull back the wrong-language sources, dropping citation quality. Cross-language research is its weak spot.
How to verify: the Chinese/English source ratio doesn’t match what you expected.
6. Too many input URLs / files in one task
Paste 10+ URLs or 3+ PDFs and the agent may sample rather than process all of them.
How to verify: the report’s citations don’t cover every URL you provided.
Shortest path to fix
Ordered by time-to-test — 30-second checks first.
Step 1: Confirm you got the full version (not lightweight, not a downgrade)
You can tell from the response itself:
| Signal | Full Deep Research | Lightweight / downgraded |
|---|---|---|
| Clarifying questions + editable plan before it runs | Yes | Often skipped |
| Progress UI (“Researching… / Browsing…”) | Yes, multi-stage, you can interrupt to refine | Brief or none |
| Total runtime | 5–30 minutes | Often <= 3 minutes |
| Sources block at bottom | 10–40 entries | Few, or sparse |
| Report length | ~1,500–4,000 words | Often <= 1,000 words |
If it’s the lightweight version, you’ve hit the full monthly cap: wait for the reset, upgrade a tier, or accept the shorter output for low-stakes questions.
Step 2: Rewrite the prompt to be decomposable + bounded
Bad → good prompt rewrite template:
Bad: "Research AI video generation tools for me"
Good: "Compare Sora, Veo 3, and Kling across these 4 dimensions:
1) max single-shot duration
2) character consistency across shots
3) commercial-use licensing
4) public pricing
Output as a table with source URL + publication date per row."The three required elements:
- Named entities — specific products / companies / papers / time windows, not abstractions like “the AI industry.”
- Explicit dimensions — list the columns you want compared.
- Output format — table / report with H2s / Markdown list.
Then answer its clarifying questions concretely. “Just go” is a wasted run.
Step 3: Restrict it to sources you trust (built-in, since Feb 2026)
Instead of hoping it finds good sources, tell it which ones to use. In the prompt window, open Sites → Manage sites and either:
- limit research to only the domains you enter, or
- choose “Prioritize these sites, but allow full-web search” to emphasize them while still searching the wider web.
Enter domains as a comma-separated list, e.g. nih.gov, who.int for medical topics or the official docs domains for a product comparison. This sidesteps “couldn’t reach site X” and sharply raises citation quality for medical, legal, financial, or government topics. You can still paste specific URLs in the prompt body (keep it to ~5 — more and it samples).
Step 4: Split the task into multiple runs
If one prompt spans 3 unrelated sub-questions (e.g. “market size” + “technical principles” + “business models”), run it 3 times, one sub-question each, and merge by hand. Deep Research goes deep far better than it goes wide.
Step 5: Try a different language
If a Chinese run pulls only low-quality Chinese sources, translate the same prompt to English and rerun — the English source pool is larger and usually higher quality. Then have GPT-5.5 translate the English report back into Chinese.
Step 6: Export + double-check citations
Deep Research occasionally “hallucinates citations” — the URL is real but the sentence it claims came from there isn’t actually there. For anything you’ll publish, click 3 random Sources entries and verify the quote.
How to confirm it’s fixed
A successful run, after the fixes above, should show: clarifying questions answered → an editable plan → a 5–30 minute browse with a live progress trail you could have interrupted → a final report of ~1,500+ words with 10+ distinct sources, including the domains you pinned in Step 3. If you still get a thin report with <= 8 sources, you’re either on the lightweight tier (Step 1) or your answers to the clarifying questions were too vague (Step 2).
Prevention
- Use Deep Research when you already know the direction and need evidence — not when you’re unsure what you want. For the latter, scope it in a normal chat first.
- Plan your monthly quota: list the 5–10 topics that truly need full Deep Research and rank them. Don’t burn full runs on trivia — use lightweight or normal chat for those.
- Pin trusted domains via Sites → Manage sites for any regulated topic; it’s the cheapest reliability win.
- Anything you’ll publish: hand-verify at least 3 citations. The model can’t guarantee this for you.
- Keep 3–5 reusable prompt templates (market comparison, technical due diligence, competitor research). A new topic in a known template beats starting from scratch.
FAQ
Why did Deep Research suddenly start giving short, shallow reports? You most likely hit your monthly cap for the full version and ChatGPT switched you to lightweight Deep Research without an obvious warning. It still browses, but returns shorter results with fewer sources. Check your usage in Settings, or wait for the monthly reset.
Does Deep Research use my full ChatGPT context window? No. The browsing agent manages its own context across sub-queries; it isn’t bounded by your chat’s in-app context. Plan limits are about the number of runs per month, not token length.
My task is stuck on “Researching…” or says “Deep research request failed.” Is it my prompt? Probably not. That pattern usually means a server-side incident — check status.openai.com. If there’s an open Deep Research incident, wait and retry later; don’t keep re-running, since failed runs have historically consumed quota.
A site I need never shows up in the sources. Why?
Many sites block OpenAI’s OAI-SearchBot / ChatGPT-User crawlers via robots.txt or Cloudflare (X, LinkedIn, paywalled Substack, intranets). Either paste the specific URLs into the prompt or pin the domain via Sites → Manage sites so it tries them directly.
Can I make it only cite trusted, authoritative sources? Yes. Open Sites → Manage sites in the prompt window, enter your domains as a comma-separated list, and choose either “limit to these sites” or “prioritize these sites, but allow full-web search.” This is the most reliable way to lift citation quality on medical, legal, or financial topics.
Related
- ChatGPT web search unavailable
- ChatGPT Canvas not working
- ChatGPT slow response
- Choose an AI search tool
- ChatGPT prompt improvement
Tags: #ChatGPT #Troubleshooting