Back to blog
Web Scraping·August 14, 2026·9 min read

ChatGPT Scraper API: get the answers ChatGPT actually shows users

Why the official OpenAI API can't measure ChatGPT visibility, and how Anakin's ChatGPT Scraper API returns the answers users actually see - text, citations, and search queries - with per-country geo targeting for AI visibility and GEO tracking.

T

Tosh Kothari

Anakin Team

Radar diagram of the chatgpt.com retrieval layer: one real ChatGPT answer with 47 sources, extracted as structured JSON fields answer_text, citations, web_search_query, and country.

Your brand shows up in some ChatGPT answers and not others, and nothing in the OpenAI developer platform tells you why. That's because ChatGPT and the OpenAI API are two different systems. The API sells you the model. chatgpt.com wraps that model in a retrieval layer that decides what to search, which sources to cite, and which brands to name (OpenAI, 2024). AI visibility lives in that layer, so if you want to measure it, you have to query the product itself.

Key takeaways

  • ChatGPT and the OpenAI API are different systems. The API serves the model. chatgpt.com adds a retrieval layer that runs web searches, picks sources, and cites them. GEO tracking measures that layer, and no official endpoint exposes it.
  • A ChatGPT Scraper API sends your prompt to the real ChatGPT and returns the answer a user would see: the text, the citations, and whether a web search ran.
  • Anakin's ChatGPT Scraper API is a dedicated action in the Wire catalog. One call, 2 credits, and you get back the full answer plus every source ChatGPT cited, as structured JSON.
  • Anakin patches the action when ChatGPT ships frontend changes, and Wire shows 99.95% uptime over the last 90 days. Failed calls aren't billed.
  • Pass a country code (and, for the US, a state code) and the query runs from that market. ChatGPT answers differently by location, and you can measure each market separately.

Why the OpenAI API can't tell you what ChatGPT says

The OpenAI API returns what the model knows. ChatGPT returns what the model finds. Ask ChatGPT "best project management tool for small teams" and the product decides whether to search, runs the searches, picks the sources, and writes an answer with citations. The answer changes with the user's location and with whatever is in the index that week. None of that happens in a plain chat/completions call. OpenAI's API can search the web (OpenAI API docs), but that's a search you configure yourself, not the retrieval layer chatgpt.com runs for its users.

This matters because the answers people act on come from the product. You can send the same prompt to the API a thousand times and never see what a ChatGPT user sees, because their answer went through the product's retrieval and yours didn't. If you're tracking whether ChatGPT recommends your brand or cites your docs, the official API is measuring the wrong system.

AI visibility is how often, and how favorably, an AI assistant names your brand or cites your content when it answers relevant questions. Generative Engine Optimization (GEO) is the work of improving that (Princeton GEO study, 2024). Both start from the same raw input: the real answers ChatGPT gives to real questions, collected often enough to see movement. There's no official endpoint for "what would ChatGPT say", and that's the gap a scraper API closes.

What a ChatGPT Scraper API does

A ChatGPT Scraper API takes your prompt, runs it through the actual ChatGPT product, and hands back the response as structured JSON. Doing this yourself means a browser automation stack: Playwright, residential proxies, session handling, and an engineer on call, because ChatGPT streams its answers, sits behind anti-bot systems, and ships frontend changes without notice. A scraper that worked in January quietly breaks in March. For a handful of queries a day, DIY is viable. Past that, the maintenance is the product, and a managed API is how you stop paying for it in engineering time.

When you compare managed options, three numbers decide it: success rate, cost per query, and how fast the vendor recovers when ChatGPT changes something.

Anakin's ChatGPT Scraper API

ChatGPT has its own dedicated scraper in Anakin's Wire catalog, with its own action, its own parameters and its own docs page. It isn't a generic scrape endpoint pointed at chatgpt.com, and it isn't bundled behind some multi-engine wrapper. You call the chatgpt action, it runs your prompt on the real product, and it returns the answer with everything around it parsed into fields. Gemini and Perplexity are separate scrapers in the same catalog, each with its own action and docs page. (If you want one prompt run across several engines side by side, that's Anakin's AI Visibility product, built on top of these scrapers.)

The parameters you'll use most (full reference):

  • prompt - what you'd type into ChatGPT. Required, up to 4,096 characters.
  • web_search - ask ChatGPT to search the web. The response tells you whether a search actually ran, via web_search_triggered.
  • country - a 2-letter code. The session runs through a proxy in that country, so you see the answers that market sees. Anakin's proxy network covers 207 countries.
  • state - a 2-letter US state code, used with country set to us, so local answers ("near me") follow that state.
  • messages - optional earlier turns (up to 5), sent before your prompt in the same conversation so ChatGPT answers with that context. Each turn costs 2 credits.
  • include_html - adds the rendered HTML to the response. It's off by default, and the structured fields cover most uses.

A single-prompt call costs 2 credits, and failed calls cost nothing. When ChatGPT ships frontend changes, Anakin patches the action upstream, so your integration keeps working. That rests on reverse-engineering experience going back to 2017, and the Wire product page shows 99.95% uptime over the last 90 days (checked 25 Oct 2026).

Calling it

curl https://api.anakin.io/v1/wire/task \
  -X POST \
  -H "X-API-Key: $ANAKIN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "action_id": "chatgpt",
    "params": {
      "prompt": "best web scraping API for AI agents",
      "web_search": true,
      "country": "us"
    }
  }'
# returns {"status": "processing", "job_id": "...", "poll_url": "/v1/wire/jobs/..."}

# poll until "status" is "completed" (or "failed")
curl https://api.anakin.io/v1/wire/jobs/JOB_ID \
  -H "X-API-Key: $ANAKIN_API_KEY"

The first call returns a job_id, and the second polls it until the job completes. The Python SDK (pip install anakin-sdk) does both for you:

import os
from anakin import Anakin  # pip install anakin-sdk

client = Anakin(api_key=os.environ["ANAKIN_API_KEY"])

result = client.wire.run(
    "chatgpt",
    {
        "prompt": "best web scraping API for AI agents",
        "web_search": True,
        "country": "us",
    },
)
answer = result.data["data"]  # the job's data envelope: {"status", "data", "meta"}
print(answer["answer_text"])

You can also pass a webhook_url and get the result pushed to you instead of polling (webhooks docs).

What comes back

Here's the data field from a real completed job, run 2026-08-14 and truncated for length:

{
  "prompt": "best web scraping API for AI agents",
  "title": "Web Scraping API Search",
  "model": "gpt-5-6",
  "country": "us",
  "web_search_triggered": true,
  "web_search_query": "best web scraping API AI agents Firecrawl Browserbase Apify Bright Data ScrapingBee 2026",
  "answer_text": "If you're building AI agents that need live web access, I'd shortlist these in 2026: ...",
  "citations": [
    {
      "url": "https://www.stork.ai/blog/best-web-scraping-tools-for-ai-agents-2026",
      "title": "Best Web Scraping Tools for AI Agents in 2026 (Compared) | Stork.AI",
      "domain": "Stork",
      "date_published": "2026-07-13T00:00:00Z",
      "cited": true,
      "position": 1
    }
  ]
}

Three fields do the heavy lifting for GEO work:

  • answer_text is the complete answer a user would have read, with markdown and raw variants alongside it. This is where you check whether your brand got named.
  • citations is every source ChatGPT pulled, as structured records: URL, domain, publish date, position, and whether it was actually cited in the answer. This run returned 47 of them. If you want to know whose content ChatGPT trusts for your category, it's this list.
  • web_search_query shows you the search ChatGPT actually ran under the hood, and web_search_queries lists every search when it ran more than one. Your prompt was "best web scraping API for AI agents". ChatGPT searched for specific brand names. That rewrite is exactly the kind of thing GEO teams need to see, and it's a field in the response.

There are more: references, search_sources, entities, and shopping fields for commerce-style prompts. The three above are where most tracking pipelines start.

Getting started

Create an API key from the anakin.io dashboard and the calls above work as-is. New accounts start with 300 free credits, which covers 150 single-prompt ChatGPT calls. Usage is billed in credits per completed call. If you're running volume (thousands of prompts across markets on a schedule), contact Anakin for enterprise and custom plans, which come with direct support from the engineering team.

What this gives AI SEO and GEO agencies

If you run an AI SEO or GEO agency, this API turns "does ChatGPT recommend my client" into a weekly number. Define the prompts that matter for the client, often hundreds or thousands of them ("best <category> for <audience>" and its variants), run them on a schedule in each target market, and store the responses. Mention share is a pass over answer_text. Source share is a pass over citations. The trend line is the diff between weeks. If you'd rather not run the scheduler yourself, an Anakin Wire monitor can run the chatgpt action on an interval and alert you when the answer changes.

That's the deliverable: share of ChatGPT answers naming the client, per market, per week, with the answer text as evidence. And because citations shows whose content ChatGPT is pulling instead, the same data hands you the gap analysis for the next content cycle: which domains own the category today, and what they published that got cited.

Frequently asked questions

What is a ChatGPT Scraper API?

A ChatGPT Scraper API is a service that submits prompts to the real ChatGPT product and returns the answers programmatically as structured JSON. It captures the retrieval-layer behavior (web search, citations, brand mentions) that the official OpenAI API does not expose.

Why not just use the official OpenAI API?

The OpenAI API serves the model without ChatGPT's retrieval layer. It doesn't run ChatGPT's web searches, pick its citations, or vary by user location, so it can't tell you what ChatGPT actually says to users. For AI visibility tracking, the product is the thing being measured.

Is there a ChatGPT Search API?

OpenAI does not offer a public API that returns ChatGPT's search-backed product answers. Its API can search the web, through a web_search tool in the Responses API or dedicated search models in Chat Completions (OpenAI API docs), but you configure that search yourself, and its retrieval and citations are not the same system as chatgpt.com. Scraper APIs exist to close that gap.

Does the API return ChatGPT's citations?

Yes. Every search-backed answer comes with structured citation records: URL, domain, title, publish date, position, and whether the source was actually cited in the answer text. A single response can carry dozens of them, and the web_search_query field shows the search ChatGPT ran to find them.

Can you track ChatGPT answers in specific countries?

Yes. Pass a 2-letter country code and the query runs through a proxy in that country, across 207 countries and territories. For the US you can also pass a 2-letter state code. ChatGPT answers location-sensitively, so per-country tracking shows differences a single-region setup misses.

What does it cost?

A single-prompt ChatGPT call costs 2 credits, each extra conversation turn adds 2 more, and failed calls cost nothing. New Anakin accounts start with 300 free credits, and the Pro plan lists 5,000 credits for $19/month (Anakin pricing, checked 25 Oct 2026). For comparison, cloro's paid plans start at $30/month for 37,500 credits (its free tier includes 500 credits), and SerpApi, which covers Google's AI surfaces rather than ChatGPT, starts at $25/month. For high-volume or multi-market tracking, Anakin also offers enterprise and custom plans.

Does Anakin cover engines besides ChatGPT?

Yes. Gemini and Perplexity are separate scrapers in the Wire catalog, along with Google AI Mode, Google AI Overview, Naver AI Briefing and Baidu's Wenxin (ERNIE). Each has its own action and parameters, listed on the AI engines docs page. Anakin's AI Visibility product runs one prompt across ChatGPT, Gemini and Google AI Overview side by side when you want to compare how each platform responds.

In a 2026 Semrush survey of US B2B professionals who use AI at work, 71% said they use ChatGPT for product research (Semrush, 2026), so its answers are now a channel you can measure. Try Anakin's ChatGPT Scraper API and see how ChatGPT answers the questions your customers are asking, and who it cites when it does.