ChatGPT Results Scraper: Answers, Citations & Screenshots
Pricing
from $2.70 / 1,000 query answereds
Scrape ChatGPT answers and search results for any question, with no login and no API key. Get cited sources with URLs, images, tables, product cards and a screenshot of each answer, from 20 countries. Export JSON, CSV or Markdown, run via API, schedule runs or integrate with AI workflows.
images field lists the pictures of the answer: image cards, pictures of image mosaics and the thumbnails of citation pop-ups, with imageUrl, alt, pageTitle, publisher, context, sourcePosition and collageCount. It is an empty list on an answer without pictures and on error items.answerHtml now keeps the src of images (https only, no icons), and answerMarkdown shows each image of the text as . answerText is unchanged. The image links point to ChatGPT's image cache and may stop working later, so download the images you want to keep.country is now a menu of 20 countries (AT, AU, BE, BR, CA, CH, DK, ES, FR, GB, GR, IN, IT, JP, MX, NL, PL, SE, TR, US). Only countries where the answer comes back localized for that country are listed. Lowercase codes are still accepted in API calls.
The country menu now lists United States first. The accepted values and the default (US) do not change.
Germany is removed from the country menu: the hosted browsers rarely reach chatgpt.com from Germany. A run with country set to DE now stops before any request, with the list of accepted codes.
Every query now waits a fixed 120 seconds for its answer. The timeout status refers to that duration.
provider now takes the values primary and fallback (and fixture for local test runs only). Some countries are served directly by the fallback browser network.
The run log now holds one line per failed attempt.
The Actor is now delivered as a prebuilt bundle. This has no consequence for your inputs or outputs.
The files of each query now start with their kind: answer-<slug>.md, item-<slug>.json, screenshot-<slug>.png and table-<slug>-<n>.csv (they were <slug>.md, <slug>.json, <slug>.png and <slug>-table-<n>.csv). The run files results.md, results.json and citations.csv keep their names. The links in the dataset item (files, screenshotUrl) always point to the right key.
The key-value store has a schema: in the Storage tab, the files are grouped in seven tabs (answers, items, screenshots, tables, and the three run files). Each file is stored with a plain content type (text/markdown, text/csv, application/json, image/png).
Citation bullets now list their pages in sources, and links outside citation bullets are included. Each bullet in answerMarkdown shows the numbers of its pages, for example [2][3], and [?] marks a bullet whose pop-up could not be read. The Actor opens at most 20 bullets per answer within 20 seconds, and this time is part of responseTimeMs.
Each page behind a citation bullet is now a source, with its domain, its title, the site name in the new publisher field and the excerpt shown in the pop-up in the new snippet field. Its url is filled when the page could be opened without leaving the answer, and is null otherwise. sources[].url can therefore be null. The links of the text have publisher and snippet set to null.
citations.csv has two new columns, publisher and snippet. A null value gives an empty cell.
engine and maxWaitSeconds input fields. If an input still contains them, they are ignored and the run log shows a warning.language input and the language output field are removed: ChatGPT answers in the language of the question. Older inputs that still send language are accepted and the field is ignored.model, searchQueries, followUps and streamCaptured are removed. They are gone from the dataset item, from item-<slug>.json and from results.json.places field is removed. The map that ChatGPT shows to logged-out visitors says that map data is unavailable, so it never held a place. It is gone from the dataset item, from item-<slug>.json and from results.json.sources, places, products, tables, status, response time and a screenshot link..md, .json, table CSVs, .png) and per run (results.md, results.json, citations.csv).query-answered event per answered query, and no charge for error items.ok, invalid_input, blocked, no_answer, timeout and engine_error, with the run always ending as succeeded.searchQueries, followUps and model are empty in this version and are present so the output shape stays stable.places is often empty because map cards frequently report that map data is unavailable.