Web research

All

Some questions your knowledge base cannot answer: a market overview, the current state of a regulation, public facts on a topic. That’s what web research is for. It sends your question to OpenAI, which searches the web in the background several times and from different angles and writes a structured report with cited sources. The report arrives in your conversation automatically.

Web research is the counterpart to research mode, which goes through your internal documents completely. Which of the two fits is explained there under Research or Web?.

Not to be confused with web search in normal chat (Brave or Serper): it hands the chat a few search hits within seconds. Web research is a background run of its own that takes minutes. → Web search in chat

For users

Using the “Web” chip

In the chat composer, the Web chip (globe icon) sits next to the Research chip, provided your administrator has enabled web research.

The chip applies to exactly one message. It is read when you send and reset immediately. For a second web research run, switch it on again.
Web and Research are mutually exclusive. Switching one chip on switches the other off.
Switching conversations turns the chip off.

What happens after you send

The chat answers right away with a notice: ”🌐 The web research is running in the background at OpenAI. The report will be delivered to this conversation automatically as soon as it is finished — usually within 5 to 20 minutes. With the PII shield on, names were replaced by placeholders; the web search cannot look for them.” In the desktop app, “On Desktop the research only runs while the app is open.” follows.
The research usually takes 5 to 20 minutes. A spinner turns on the Web chip, and the label shows the time elapsed so far, for example “Web research 3 min”. There are no intermediate steps, because OpenAI reports none.
You can keep working — ask other questions, switch conversations, reload the page. The display on the chip stays as long as the run is going.
The report appears live in the same conversation. If you have switched in the meantime, a badge on the conversation in the list points to the new report.

As with research mode, each user has at most one research run at a time — Research or Web alike. → Limits & notes

How the report is structured

A structured text with headings in your interface language. The model is instructed to back every factual statement with a link, prefer primary sources and say openly what it could not verify.
Below it, “Cited sources”: a numbered list of the web pages the report relies on, each address once. The system appends this list, not the model.
At the very end the number of searches, for example “7 web search(es) made.”.
If the model hit its length limit, the text ends with “Note: The summary reached the model’s length limit and may be cut off at the end.”

Check important statements against the linked source — web sources can be outdated or wrong, and the model can misrepresent a source.

Privacy

Your question goes to OpenAI in the USA. OpenAI researches with its own web search. Web research does not read your knowledge base; only the question goes to OpenAI.
OpenAI keeps the request for about 10 minutes while the run is going, so that the result can be collected. It is not stored there permanently.
With the PII shield on, names in your question are replaced by placeholders (such as [PERSON_1]). The web search cannot look for a placeholder — a question about a specific person or company then gets little out of it. For such questions, switch the shield off for this one message (“PII Shield Off · 1 message”); the question then goes to OpenAI in plain text. Think about whether you want that first. → Switching the protection off on purpose
The finished report is stored in your conversation like any assistant message; placeholders in it are replaced by the real names on delivery.

When it doesn’t work

What you see	What’s behind it
No Web chip	Web research is not enabled, or no usable model is chosen. Only your administrator can change that.
”ℹ️ Web research is not available in this installation (switched off or no usable model). This question was answered normally.”	Your administrator switched web research off or disabled the model after you loaded the page. After a reload the chip is gone.
”ℹ️ A background research run is already in progress. This question was therefore answered normally.”	A research run of yours is still going. Wait for its report, then switch the chip on again.
The request is rejected with a note about the budget	Your cost budget has been reached. Contact your administrator.
”⚠️ The research could not start: its research agent is disabled or was deleted. …”	The Web Research agent is disabled or deleted on the Agents page. Only your administrator can switch it back on.
Research failed with “Reason:“	OpenAI reported an error — the reason names it, for example a model that isn’t enabled for your OpenAI access. If it says “OpenAI ended the web research without a report (status: …)”, no text came back from OpenAI. That does not mean “nothing found”; try again later.
Research (aborted) with “time limit reached or stopped manually”	The run exceeded its maximum duration or was stopped. It is then cancelled at OpenAI as well.
On desktop: “The application was shut down while the run was in progress.”	You closed the app during the run. Ask the question again and keep the app open.

The headings of these final messages are the same as in research mode. → If the run does not end normally

For administrators

Enabling it

You enable web research in system settings on the Web research card. → System settings

“Allow web research” switch — default: off.
“Web research model” — pick a model from the list.

A usable model is an enabled chat model on an OpenAI provider with the endpoint https://api.openai.com/v1 and stored credentials. gpt-5.6-sol is recommended. OpenAI has shut down the former deep research models o3-deep-research and o4-mini-deep-research; they no longer work, even though OpenAI still lists them. You add the model like any other. → AI models / providers

Users only see the Web chip when the switch is on and the chosen model is usable. Otherwise the card shows a notice:

“No usable model: …” — there is no suitable model.
“The chosen model is not usable (none chosen, disabled, deleted or moved to another endpoint). …”

If you disable the model later, sessions that are already open still see the chip until they reload. A question sent with it is answered normally, with the “not available” notice.

In the desktop app, the switch and model are settings of the desktop app itself, even when it is paired with a company server.

Only Managed-AI models

When Only Managed-AI models is on for chat, web research is not available: the question would go to OpenAI in the USA, and OpenAI is not a Managed AI provider. The chip then stays hidden. → Allow Managed AI models only

Run time

The maximum duration of a run is the Max run minutes of the seeded Web Research agent on the Agents page, 40 minutes by default. After that the run is aborted, at OpenAI as well. Provider and model stay empty on this agent: runs from the chat always use the model from system settings. → Agents

Max run minutes also lengthens the “one research run per user” lock. → Enabling research mode

If you disable or delete the Web Research agent, no web research starts any more; users get a message. To switch the feature off, the switch in system settings is the better choice — the chip then disappears.

Costs

Tokens of the model are booked like any model usage and count towards cost budgets. If a user’s budget is reached, the request is rejected right away, before a run starts.
The fee per web search is charged by OpenAI on top of the tokens. It is currently not booked and is therefore missing from booked costs and budgets. How many searches a run made is shown at the end of each report.

Research mode — research in the knowledge base
Tools: research — research_web and the agent behind it
The PII shield · Chat
For administrators: System settings · Enabling research mode