---
title: "Discussion on \"Best ChatGPT scraper APIs: answers and citations as JSON\" | Hashnode"
description: "Discussion on \"Best ChatGPT scraper APIs: answers and citations as JSON\". To get what ChatGPT tells users, with the pages it cites, as JSON, there are three routes: a scraper API that collects and parses the consumer answer, OpenAI's developer API with its web search tool,"
url: "https://hashnode.com/posts/best-chatgpt-scraper-apis-answers-and-citations-as-json/6abbd9924cd7a53d9b10b325"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Best ChatGPT scraper APIs: answers and citations as JSON](https://answerline.hashnode.dev/best-chatgpt-scraper-apis-answers-and-citations-as-json?utm_source=hashnode&utm_medium=feed)

To get what ChatGPT tells users, with the pages it cites, as JSON, there are three routes: a scraper API that collects and parses the consumer answer, OpenAI's developer API with its web search tool,

Sep 29 · 16 min read · answerline.hashnode.dev  
Updated Oct 13  
Read the article

-   [#chatgpt](https://hashnode.com/tag/chatgpt)
-   [#api](https://hashnode.com/tag/api)
-   [#seo](https://hashnode.com/tag/seo)
-   [#webscraping](https://hashnode.com/tag/webscraping)

## Written by

[AnswerLine](https://hashnode.com/@richardnemeth)

@richardnemeth

  
[View profile](https://hashnode.com/@richardnemeth)

Founder at AnswerLine — structured data from AI answer engines.

Founder of AnswerLine (answerline.dev): one API that returns the answers AI assistants and Google show users as structured JSON.

## Responses 1

Markdown supported · ⌘↵ to post

[OnlineProxy](https://hashnode.com/@onlineproxyio)

[Sep 30](https://hashnode.com/posts/best-chatgpt-scraper-apis-answers-and-citations-as-json/6abbd9924cd7a53d9b10b325/comment/6abd1150204cb1799c1059b5)

The bright data row is wrong, with the 5k free tier 6000 answers cost about $1.50 instead of $9.00, and the best fix is one column for effective usd per 1000 successful answers. Parse's 82 day study is the strongest case for reruns, just keep 52.6% to 28.9% survival separate from median 0.18 overlap and keep the small panel caveat

The 30 to 50 runs rule does not hold up since reruns correlate, close claims need more prompts not more runs. Tag empty arrays with not\_requested, not\_triggered, unsupported or parse\_failed so nothing served never reads as nothing captured