ChatGPT Search Scraper: answers, cited sources and product cards

ChatGPT Search Scraper collects answers from anonymous ChatGPT sessions, together with the sources cited in each answer and product cards when they appear. Use it to gather the raw evidence behind AI search visibility research, compare answers to a consistent question set, or build a dataset for your SEO reporting.

The actor returns one dataset row per question. Each row includes plain answer text, a Markdown version, ordered sources and compact product information. Questions are independent: the next question starts a new conversation. No ChatGPT account, OpenAI API credential, brand configuration or competitor list is required as input.

What the ChatGPT Search Scraper returns

• The answer as plain text (answer_text) and as Markdown with headings, lists and tables (answer_markdown).
• The cited sources in the order ChatGPT shows them: URL, page title, domain and position, with tracking parameters removed and duplicates merged.
• Product cards when ChatGPT shows them: product name, position, numeric price, currency, merchant and product URL.
• Context for every row: the question, its position in your list, the country, a UTC capture time and the conversation identifier.

ChatGPT Search Scraper use cases

SEO teams can collect the answers their clients care about and examine which websites appear as cited sources. Agencies can keep the same question list across several collection runs and compare the resulting evidence in their own reporting tools. Brands can inspect recommendations and product mentions without having to manually copy every answer into a spreadsheet.

ChatGPT rank tracking and AI visibility

Run the same question list on a schedule and count how often a domain appears in citations, at which position, and whether a product shows up in products. That is the raw material of a ChatGPT rank tracker: you keep the method in your own spreadsheet, BI tool or database instead of trusting a black-box visibility score. Because every row carries the question, country and capture time, a change in a domain's citations can be traced back to the exact answer.

Content and competitor research

Read which pages ChatGPT cites for the questions in your niche. The domain field makes it simple to group sources by publisher, and answer_markdown keeps comparison tables intact, so you can see which vendors are named, in what order, and which reviews or pricing pages back up the answer.

Product and shopping questions

For buying questions ChatGPT often shows product cards. The actor returns them as structured rows with a numeric price and currency, so merchants and brands can see which products and shops are recommended for a question in a given country.

This is a raw collection tool. It does not calculate brand visibility scores, infer sentiment, identify competitors, or schedule repeat audits. Those calculations depend on your reporting method and the questions you select. Save the question, country and capture time with the answer so that later comparisons keep their context.

Input

queries is required. On Apify, supply one string with one question per line, both in the input form and through the Apify API. Leading and trailing whitespace is removed and blank lines are ignored. Question order is preserved, including duplicate questions. The input accepts at most 10,000 questions per run; completing a large list requires a sufficient run timeout (see the FAQ). Every question must contain at most 4,000 characters. An empty question list returns an input error without producing an answer. The direct API and MCP also accept an array of question strings.

{
  "queries": "what is 17 times 23",
  "country": "US"
}

country is optional and defaults to US. Use an ISO alpha-2 code such as NL, DE or GB; lowercase codes work too. countryCode is accepted as an API alias, and an empty alias means US. When both names are present, country takes precedence. Unsupported countries are rejected before collecting answers. The accepted country list reflects ChatGPT's supported countries and territories.

Country describes where the question is sent from. It does not guarantee local sources, local shops or a particular answer language. Write your question in the language and geographic context you want to research. For example, the recorded Dutch robot vacuum question returned Dutch answer text and EUR product cards, while an English phone plan question sent from Great Britain asked for clarification and showed no cited sources.

How to use the ChatGPT Search Scraper

1. Enter your questions in the queries field, one per line.
2. Optionally pick the country the questions are sent from.
3. Start the run and open the Answers and sources view, or export the dataset as JSON, CSV or Excel.

ChatGPT scraper API: call it from your own code

Every Apify actor is also an API. This call runs the actor for one recorded question and returns the dataset rows in the response. It suits a handful of questions; for long lists, start an asynchronous run and read the dataset when it finishes.

curl -X POST "https://api.apify.com/v2/acts/s-r~chatgpt-search/run-sync-get-dataset-items?token=<YOUR_APIFY_TOKEN>" \
  -H "Content-Type: application/json" \
  -d '{"queries": "best crm software for small business 2026", "country": "US"}'

The same input works from the Apify Python and JavaScript clients, from Make, Zapier and n8n, and from scheduled tasks in Apify Console.

Output

The dataset contains an ai_answer row for each processed question. Core fields include source, fetched_at, engine, query, country, answer_shown and answer_text. source is chatgpt-search and engine is chatgpt. Capture timestamps are UTC. query_index keeps the question's position in the normalized input list.

answer_text removes citation chip labels while retaining the answer's prose, entity names and inline product names. Inline product references also keep their names in answer_markdown, without prices or merchant button labels. answer_markdown preserves headings, lists and tables. Each nested citations entry includes a URL, title, domain and position. Tracking parameters beginning with utm_ are removed from clean URLs. Duplicate sources within one answer are combined by clean URL, in order of first appearance.

Product cards are optional. When ChatGPT shows them, products contains product names, positions, numeric prices, currencies, first merchants and product URLs when available. Missing price or merchant information remains null. A product mention is evidence that a card appeared in the response; it is not a verified price quotation, inventory check or recommendation from the actor.

This excerpt is taken from an actual recorded response to the arithmetic question above:

{
  "query": "what is 17 times 23",
  "answer_shown": true,
  "answer_text": "17 × 23 = 391.",
  "citations": [],
  "products": []
}

For direct API integrations, the same sources and product cards also appear in separate citations and products lists with their question and capture context. Source rows additionally preserve the original cited URL, attribution and publication timestamp when the answer supplies them. Product rows include ratings, review counts, image URLs and offers when those fields exist on the card.

A shopping question returns sources and product cards together. Excerpt from the recorded US response to a headphones question, with the answer text left out and the lists shortened to two entries (the full row has 5 sources and 5 products):

{
  "query": "best noise cancelling headphones under 200 dollars",
  "country": "US",
  "fetched_at": "2026-10-04T22:26:35.905972Z",
  "answer_shown": true,
  "citation_count": 5,
  "product_count": 5,
  "citations": [
    {
      "url": "https://www.rtings.com/headphones/reviews/best/by-price/under-200",
      "title": "The 5 Best Headphones Under $200 of 2026 - RTINGS.com",
      "domain": "rtings.com",
      "position": 1
    },
    {
      "url": "https://www.rtings.com/headphones/tools/compare/sony-wh-ch720n-wireless-vs-anker-soundcore-space-q45-wireless/38908/34852?usage=19",
      "title": "Sony WH-CH720N vs Anker Soundcore Space Q45 Wireless: Which Headphones Are Better? - RTINGS.com",
      "domain": "rtings.com",
      "position": 2
    }
  ],
  "products": [
    {
      "product_name": "Anker Soundcore Space Q45",
      "position": 1,
      "price": 129.99,
      "currency": "USD",
      "merchant": "Micro Center",
      "url": "https://www.microcenter.com/product/653777/anker-soundcore-space-q45-adaptive-noise-cancelling-wireless-bluetooth-headphones-black"
    },
    {
      "product_name": "Sony WH-CH720N",
      "position": 2,
      "price": 89.99,
      "currency": "USD",
      "merchant": "Target",
      "url": "https://www.target.com/p/sony-whch720n-bluetooth-wireless-noise-canceling-headphones-black/-/A-88533038?TCID=OAG&afid=chatgpt"
    }
  ]
}

How it compares

	ChatGPT Search Scraper	Typical alternative
Answer	Plain text without citation labels, plus Markdown with tables	One text field with Markdown and inline image links
Sources	Ordered list with URL, title, domain and position; tracking parameters removed, duplicates merged	URL and title with utm_source tracking left in, no domain field
Product cards	Separate rows: name, position, numeric price, currency, merchant, URL	Names and prices only inside the answer text
Questions without an answer	Returned with an English reason and not charged	Varies
Search queries ChatGPT used	Not available in the anonymous client (query_fan_out is empty)	Offered as a field by some tools, often empty
Account needed	No ChatGPT account or OpenAI key	Varies

Pricing and free plan

Billing counts completed answers with non-empty answer text: each one is a single chatgpt-result-scraped event, the same event name other ChatGPT search actors use, so switching does not change how your usage is counted. Sources and product cards are included with the answer and are not charged separately. A completed answer without citations still counts as an answer. Unanswered questions and invalid input do not generate answer charges. Check the actor's Pricing tab for current event prices and any platform fees.

Users without a paid Apify plan receive at most ten dataset rows per run. The question list is capped before collection begins. Paid users can submit larger lists; processing time grows with the number of questions. Repeated questions remain separate requests and separate completed answers.

FAQ

Does every answer include web sources?

No. ChatGPT decides whether to search and what to cite. Arithmetic, clarification requests and questions without relevant results can produce complete answers with an empty citations list. The actor preserves that distinction instead of inventing sources.

When ChatGPT shows only lightweight source chips (domain names without citation URLs in the page), citations stays empty. The actor never invents URLs, so empty citations alongside visible source names is expected behaviour rather than a defect.

Can I request ChatGPT's internal search queries?

The anonymous client does not expose query fan-out. query_fan_out is therefore always an empty array. It should not be interpreted as proof that ChatGPT did not perform a search.

How are long question lists processed?

The actor follows continuation cursors automatically. Allow roughly 3–4 questions per minute (about 15–20 seconds per question), with additional time for retries and waiting during busy periods. The default Apify run timeout is 1,200 seconds (20 minutes), typically enough for roughly 60–70 questions, not the full 10,000-question input limit. Concurrent runs share collection capacity fairly, so each run may need more time during overlap. Increase the run timeout in Apify's run options before submitting a large list; for 500 questions, allow at least three hours and extra time for retries. If the timeout guard is reached, the run returns the completed answers with partial: true and a warning. Only completed answers are charged.

Direct API clients keep the same questions and country, pass next_cursor as cursor, and continue until the cursor is null. Every collection request has a bounded time budget, so larger lists are delivered in several pieces. A retryable busy response should be retried with pauses; the Apify actor handles this automatically within its run timeout.

What happens when a question fails?

A failed question can carry answer_shown: false, null answer text and an English error reason alongside successful questions. If no answer completes, the request returns a structured source error. Temporary timeouts can be retried. Previously delivered answers remain available when a later question cannot finish.

Are results identical across runs?

No. ChatGPT answers and search results can change between runs. Keep the full response, question wording and timestamp when comparing measurements, and use repeated observations before drawing conclusions about visibility.

Is this the same as the OpenAI API with web search?

No. The OpenAI API answers through developer models with their own settings and requires an API key that you pay for. The ChatGPT Search Scraper records what an anonymous visitor sees in ChatGPT itself: the answer, the sources it cites and the product cards it shows. Use it when you need to know what ChatGPT users are shown, not what a model returns through the API.

Do I need a ChatGPT account or an OpenAI key?

No. Questions are asked in anonymous sessions. You only need an Apify account to run the actor.

Which countries are supported?

Any country or territory where ChatGPT is available, set with an ISO alpha-2 code such as US, GB, NL or DE. Countries where ChatGPT is not offered are rejected before any answer is collected, so they never cost anything.