ChatGPT API — answers, sources and search queries

CHATGPT asks ChatGPT with web search and returns the answer, every source it found, the searches it ran, and the product and place cards it drew.


taskType	`CHATGPT`
Reads	`prompt` (falls back to `query`)
Credits	1 per answer
Country	Optional, default `US`. `CN` and `HK` are asked from a Singapore exit with a local device location.
Location	Accepts `payload.location`. See Regions and location.

Request

curl -X POST https://api.querying.ai/v1/async/task \
  -H "Authorization: Bearer $QUERYING_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "taskType": "CHATGPT", "payload": { "prompt": "best CRM for a small team", "country": "US" } }'

Response

Field	Type	Description
`text`	string	The answer as plain text.
`markdown`	string	The answer as Markdown, photos in their card’s place. On by default.
`model`	string	The model ChatGPT answered with.
`sources`	array	Every source: `position`, `url`, `label`, `footnote`, `datePublished`, `searchQueries`, `cited`, `citedText`.
`searchQueries`	string[]	The web searches ChatGPT ran. On by default.
`shoppingCards`, `inlineProducts`	array	Product cards and product names linked in the answer. On by default (`include.shopping`).
`ads`	array	Sponsored items in the answer. On by default (`include.ads`).
`entities`	array	Places and products named in cards, each with its `type`, such as `local_business` or `product`, and its `name`.
`entityDetails`	array	The detail panel of each card. Only with `include.entityDetails`. See Card details.
`map`	array	Places shown in cards: address, coordinates, phone, rating, and the `icon` emoji ChatGPT pins them with. `provider` is the feed the record comes from (`serp` and `bright-feed` for Google Maps, `yelp-feed` for Yelp, `tripadvisor-feed` for Tripadvisor); `placeId` and `providerUrl` are the place’s ID and page in that feed. `resolvedBy: "address"` marks a place ChatGPT found by the address the model wrote. `localName` is the record’s other name, such as a Korean store’s Korean name, and `closed` is `permanently` or `temporarily`.
`citationPills`	array	The sources behind each inline citation chip.
`layout`	object	The card layout, when ChatGPT draws the answer as cards. See below.
`forms`	array	Follow-up forms in the answer, when there are any. See Conversations.
`conversation`	object	`id`, `turn` and `expiresAt` of the conversation the answer belongs to.
`html`	object	The screen ChatGPT showed. Only with `include.html`.
`rawResponse`	array	Parsed stream events. Only with `include.rawResponse`.

For example, a trimmed response:

{
  "text": "For a small team, HubSpot CRM and Pipedrive are the usual picks ...",
  "markdown": "For a small team, **HubSpot CRM** and **Pipedrive** are the usual picks ...",
  "model": "gpt-5-3",
  "sources": [
    {
      "position": 1,
      "url": "https://www.example.com/best-crm-small-business",
      "label": "The best CRM software for small businesses",
      "searchQueries": ["best CRM for small teams"],
      "cited": true,
      "citedText": ["HubSpot CRM and Pipedrive are the usual picks"]
    }
  ],
  "searchQueries": ["best CRM for small teams"],
  "conversation": { "id": "8f6c1d2e-...", "turn": 1, "expiresAt": "2026-10-19T12:00:00.000Z" }
}

Each source carries searchQueries (the searches whose results included it), cited (whether the answer cites it inline) and citedText (the answer text each citation backs). Sources that no search returned, such as products, have no searchQueries.

Card layout

layout appears when ChatGPT draws the answer as cards. Its nodes follow the order ChatGPT shows them:

Node	Holds
`row`, `box`	Groups of cards.
`title`, `text`, `badge`	The card copy.
`image`	The photo: `images[].url`, `thumbnailUrl` and `pageUrl`.
`markdown`	The rest of the answer.
`map`	In `places`, the `position` of each place it shows.
`form`	A form that asks you to pick a follow-up question. Its `choice` nodes hold the options and the one selected at first; its `button` node holds the question ChatGPT would send next (`prompt`). Forms are left out of `text` and `markdown`.

Include flags

CHATGPT honors every response flag. All are on by default except rawResponse and html; send false to turn one off.

Flag	Default	Effect
`markdown`	`true`	Adds `markdown`.
`searchQueries`	`true`	Adds `searchQueries`.
`shopping`	`true`	Adds product cards.
`ads`	`true`	Adds `ads`.
`rawResponse`	`false`	Adds `rawResponse`. See Markdown and raw responses.
`html`	`false`	Adds `html`, 0.25 credits when the page is stored. Rejected with `422 EVIDENCE_UNSUPPORTED` together with `location`, `CN`, `HK` or `continueFrom`. See HTML pages.
`analysis`	`false`	Adds `analysis`, 1 credit. See Answer analysis.
`citationAnalysis`	`false`	Runs a Citation Analysis on the answer.
`entityDetails`	`false`	Adds `entityDetails` at no extra cost. See Card details.

Card details

With include.entityDetails, every card in the answer (a product, place, app or brand) comes with the detail panel ChatGPT opens when a reader clicks it: a description, specifications, reviews and the pages behind them. Panels cite their own pages, so they show sources the answer text leaves out.

Field	Description
`name`, `category`, `query`	The card as the answer shows it, its kind (`product`, `local_business`, `software`, `organization`) and the name ChatGPT looked it up by.
`status`	`ok`, or `failed` with an `error`.
`markdown`	The panel as Markdown, with citation chips written like the answer’s.
`sources`	Each cited page once, in the order first cited: `position`, `url`, `label`, `site`, `description`.
`products`	The shopping card a product’s panel opens on: `title`, `price`, `merchant`, `rating`, `num_reviews`, `url` and `offers`, shaped like `shoppingCards[].products`. Only when the panel shows one.
`images`	The panel’s photos.
`complete`	`false` when ChatGPT stopped writing the panel early. The text and sources up to that point are included.

error is ENTITY_DETAIL_FAILED when the panel request fails, ENTITY_DETAIL_EMPTY when the panel has no text, and ENTITY_DETAIL_UNAVAILABLE when the answer lacks the conversation or turn credentials needed to open its panels. Card details also work with payload.location, CN, HK, fallback answers and include.html. A panel ChatGPT has written before returns in about a second; one it writes on request takes 10 to 30 seconds, with up to 4 panels at a time. Up to 20 cards per answer are opened. Panels add no credits, and each follow-up answer gets panels of its own.

"entityDetails": [
  {
    "name": "Brooks Adrenaline GTS 25",
    "category": "product",
    "query": "Brooks Adrenaline GTS 25",
    "status": "ok",
    "markdown": "## Brooks Adrenaline GTS 25\n\nThe **Brooks Adrenaline GTS 25** is a **stability road running shoe** ... ([Brooks Running](https://www.brooksrunning.com/...))\n\n### Key specifications\n\n| Feature | Details |\n| --- | --- |\n| Heel-to-toe drop | 10 mm |\n...",
    "sources": [
      { "position": 1, "url": "https://www.brooksrunning.com/...", "label": "Men's Adrenaline GTS 25 Running Shoes | Brooks Running", "site": "Brooks Running" },
      { "position": 2, "url": "https://www.rei.com/product/249691/...", "label": "Brooks Adrenaline GTS 25 Road-Running Shoes - Men's | REI Co-op", "site": "REI" }
    ],
    "images": ["https://images.openai.com/..."],
    "complete": true
  }
]

Follow-up questions

payload.followUps asks up to 4 more questions in the same conversation after the first answer, in order. Each item is one of:

{ "prompt": "..." }: a question of your own.
{ "select": { "button": 0, "choices": { "budget": "Under $100" } } }: presses a follow-up form’s button in the previous answer.

button counts the answer’s buttons from 0 and defaults to the first. choices picks options by each choice node’s name, or as an array in the form’s order; a pick is an option’s value or label, or its position from 0. Choices you do not pick keep the option ChatGPT selected, so { "select": {} } sends the button’s own prompt.

{
  "taskType": "CHATGPT",
  "payload": {
    "prompt": "best CRM for a small team",
    "country": "US",
    "followUps": [{ "prompt": "Which of these has the best free plan?" }]
  }
}

The top-level fields stay the first answer. followUps in the response lists each further answer with the question that was sent (prompt), what a select pressed (selection: the button label and each choice’s value), and the same fields as the first answer. When a select cannot be pressed, the remaining follow-ups are not asked, and followUpsStopped gives the follow-up’s index, a code and a message:

Code	Meaning
`NO_FOLLOW_UP_FORM`	The previous answer showed no form.
`BUTTON_NOT_FOUND`	No button at that index.
`CHOICE_NOT_FOUND`	No choice with that name.
`OPTION_NOT_FOUND`	No option matches the pick.
`PROMPT_UNAVAILABLE`	The button had no question to send.

The task still completes with the answers it got. Each answer costs 1 credit: the task holds credits for every follow-up and charges only for the answers it returns. include.analysis, include.citationAnalysis and include.html apply to the first answer. Other engines reject followUps with 400 VALIDATION_ERROR.

Continue a conversation later

To read an answer before you choose the next question, send a new CHATGPT task with payload.continueFrom set to the id of a completed task, plus either a prompt or a select that presses a form in that task’s last answer (the same rules as in followUps).

{
  "taskType": "CHATGPT",
  "payload": {
    "continueFrom": "8f6c1d2e-...",
    "select": { "button": 0 }
  }
}

An answer that shows a form carries forms: each form’s buttons, whose order is select.button, and its choices with their name, the option selected at first and the options.
conversation.id is the task that opened the conversation, turn counts its answers so far, and expiresAt is the last moment it can be continued, 3 hours after its first answer.
A continuation asks under the last answer of the task you name, so naming the same task twice branches the conversation.
It keeps the original country and location. Leave them out or send the same values.
A task of another account or engine, a task that has not completed, or an invalid select is rejected with 400 VALIDATION_ERROR. An expired conversation is rejected with 400 CONVERSATION_EXPIRED. Neither is charged.
A continuation costs the same as any ChatGPT task.

Get started

Core concepts

Answer engines

Search and web data

Request options

Research

GEO monitoring

Account and data

ChatGPT API — answers, sources and search queries

Request

Response

Card layout

Include flags

Card details

Follow-up questions

Continue a conversation later

​Request

​Response

​Card layout

​Include flags

​Card details

​Follow-up questions

​Continue a conversation later

​Related

Request

Response

Card layout

Include flags

Card details

Follow-up questions

Continue a conversation later

Related