App Logo
chatgpt.com

Ask ChatGPT

TL;DR

Automatically ask ChatGPT on chatgpt.com. See what ChatGPT tells a stranger about your market and which pages it cites to back that up.

About

A growth lead at a French payroll startup sends "What payroll software should a 20-person company in France use?" from a signed-out browser every Monday and logs which vendors the reply names and which review sites it cites, since those are the pages worth getting listed on. Signed out is the closest thing to a stranger asking, but you get little beyond the reply text and its cited references. A signed-in account adds the model id, every query ChatGPT typed and every page those searches pulled up, though saved memory can shape the reply and the chat is kept in that account's history. Those queries are worth a close read, since one question can fan out into many (a September run typed 9 across 2 search calls). Each run is one question in a fresh chat, and setting mode to work is saved on the account, so your own next chat starts in Work too.

Input

question required

The question to send as the first message of a new chat.

Output

calls required

One entry per search call. A call often runs a whole round of queries at once and answers one list for all of them. What the page does not receive: which query returned which result, the call's arguments, and which search backend answered it. Empty when it did not search. Null signed out, where ChatGPT does not show them.

model required

Which model answered, by ChatGPT's own id (e.g. "gpt-6"). Null when the app does not say, which is always the case signed out.

answer required

The last entry of `answers`.

answers required

One entry per reply, usually one; ChatGPT sometimes answers twice and asks which reply you prefer, and neither entry is more official than the other: read the length before reading answers[0]. Each is the reply as markdown, each citation a markdown link to its source, widgets (images, forms) left out.

sources required

Every page the searches handed the model (`given`), once each, in the order it got them: the pool the answer could draw on, as opposed to `references`, what it cited, and to the results it never got. Empty when it did not search. Null signed out, where ChatGPT does not show them.

activity required

The progress lines ChatGPT showed while it worked, in order and in the display language, e.g. "Searching the web", "Searched 8 websites". Signed out, the only trace of the searches it ran. Display text: its wording and its counts are ChatGPT's, not a count of `sources`. Empty when it showed none.

searches required

One entry per round of web searches, in order: what the model typed and the results its searches returned. Empty when it did not search. Null signed out, where ChatGPT does not show them.

references required

Every source the answer cited or linked, signed in or out, deduplicated by url and in the order the answer surfaced them: the pages behind the citation pills. Not everything the model got: that is `sources`. Empty when the answer cited nothing, which is a fact about that reply: the same question searches on one run and not the next.

stopReason required

Why generation ended, as ChatGPT records it: "stop" for an answer the model finished. Null when ChatGPT records no reason, and signed out. A finished answer can still be only a sentence announcing work it never did; that is what the model returned, not a truncation.

searchContext required

What the app asked for and what the server says about this turn, as sent to the page. Signed out, every field the server sends is null, and a field is null when the app did not send it.

conversationId required

The conversation's id. Null if the app did not create one.

webSearchQueries required

The searches ChatGPT issued, in order. Empty when the reply did not search. Null signed out, where ChatGPT does not show them: unknown, not none.

Example output

Shape only: placeholder values generated from the output schema, not a real run.

{ "calls": [ { "call": 3, "tool": "…", "results": [ { "ref": "…", "url": "https://example.com/item/123", "rank": 3, "type": "…", "given": true, "title": "Example", "pubDate": "2026-01-15T09:30:00Z", "snippet": "…", "attribution": "…", "thumbnailUrl": "https://example.com/item/123" } ] } ], "model": "…", "answer": "…", "answers": [ "…" ], "sources": [ { "url": "https://example.com/item/123", "title": "Example", "attribution": "…" } ], "activity": [ "…" ], "searches": [ { "queries": [ "…" ], "results": [ { "ref": "…", "url": "https://example.com/item/123", "type": "…", "given": true, "title": "Example", "pubDate": "2026-01-15T09:30:00Z", "attribution": "…" } ] } ], "references": [ { "ref": "…", "url": "https://example.com/item/123", "title": "Example", "attribution": "…" } ], "stopReason": "…", "searchContext": { "tool": "…", "useCase": "…", "planType": "…", "toolInvoked": true, "locationUsed": "…", "clusterRegion": "…", "requestedModel": "…", "thinkingEffort": "…", "modelExperience": "…", "locationIsPrecise": true }, "conversationId": "abc123", "webSearchQueries": [ "…" ] }

FAQ

Open a new ChatGPT chat, send one question, wait for the answer to finish (up to 3 minutes), and return the answer, the sources it cited, and the web searches it ran.
Ask an AI agent connected to Reduck to run reduck/chatgpt.com/ask, or run it from a terminal with the Reduck CLI: npx @reduck-ai/cli@latest run --script reduck/chatgpt.com/ask
You do not need one. "Ask ChatGPT" drives the real chatgpt.com pages in a browser, so it works whether or not chatgpt.com offers an API for this.
Required: question.
It returns calls, model, answer, answers, sources, activity, searches, references, stopReason, searchContext, conversationId, webSearchQueries.
No. It only uses pages of chatgpt.com that are reachable without signing in.
It makes changes on chatgpt.com, like sending, posting, booking or buying something.
Ask an AI agent connected to Reduck to run reduck/chatgpt.com/ask, or run it from a terminal with the Reduck CLI: npx @reduck-ai/cli@latest run --script reduck/chatgpt.com/ask
It is part of Reduck's official curated catalogue.
ChatGPT decides on each reply whether to search the web, so one run can cite several pages and the next none at all. When it does search, the queries it types can change between runs, and the pages it reads and cites shift with them. An empty references list means that reply skipped the search, not that the run failed.
No. OpenAI describes each API request as independent and stateless, so the reply comes from the model and the messages you send, without the history or saved memory a ChatGPT account brings to the app. API usage is billed apart from any ChatGPT plan, and the standard web search tool costs $10 per 1,000 calls plus the retrieved content billed as tokens at the model's rate.
You can, but spread them out. When several runs share one signed-in account at the same moment, ChatGPT has been seen sending tabs back to the home page mid-answer, and Reduck's Ask ChatGPT script reopens the chat up to two times, then stops with an error. If ChatGPT refuses a message (a rate limit or a moderation block), the run ends at once with ChatGPT's own error text instead of waiting out its three-minute budget.
It can. OpenAI says ChatGPT search estimates a general location (country, state or city) from your IP address and that a VPN can change that guess, so the same question asked from Paris and from Brussels may come back different. Signed-out replies sometimes report the place in searchContext.locationUsed, and to track one market the sensible habit is to run from a browser in that market and keep it there.

Related scripts

chatgpt.com

ChatGPT whoami (signed-in account)

Report which ChatGPT account this browser is signed in as: the user id, email and name. When the browser is signed out of ChatGPT, the run fails with a message asking you to sign in to chatgpt.com, so it can be used to check a session before running other ChatGPT scripts.

chatgpt.com

ChatGPT billing: list your ChatGPT subscription invoices

ChatGPT subscription invoices, not the OpenAI API platform. Lists the billing history of the ChatGPT account currently open in this browser, as its Billing page shows it: only that account's invoices, never those of the user's other accounts. To get another account's invoices, switch ChatGPT to that account, then run this again. A Business, Team or Enterprise workspace open in the browser has no personal Billing page, and the run fails saying so. Each invoice comes with its date, amount, currency, payment status, plan and a link to the invoice; the result also names the active account and lists every account the user belongs to. Amounts are in the currency's smallest unit, so 1917 means 19.17 EUR. The Billing page shows the most recent invoices first and this script cannot page further: when ChatGPT says older invoices exist, complete is false and next_cursor holds the cursor ChatGPT gave, but those older invoices are not returned. When the browser is not signed in to ChatGPT, the run fails with a not-signed-in error. Pass hosted_invoice_url to the Stripe invoice PDF download script to get the PDF.

chatgpt.com

Get ChatGPT memory setting

Report whether memory is turned on for the signed-in ChatGPT account: the "Enable memory" switch in Settings > Personalization, which lets ChatGPT use saved memories and past chats to personalize its answers. Requires a signed-in session; a signed-out browser fails with an error that says so. Use set_memory to change it.

Start building not clicking

Get started for free