ChatGPT scraper API

What the ChatGPT API returns

Each answer comes back already read. You don't parse markdown or guess which product names are brands: the response lists them, in order, with the sources ChatGPT cited and the raw response behind it.

The ChatGPT app's answer vs the OpenAI API

Buyers ask ChatGPT in the app. The OpenAI API is a different product: a model you call as a developer, with its own defaults, that answers in whatever shape you ask for. Even with its web search tool turned on, it isn't the answer a buyer reads on chatgpt.com. To measure what buyers are told, you need the app's answer.

What each one measures

	OpenAI API	ChatGPT app, through aivisibilityskill
What you get	A model's reply to your prompt	The answer buyers see on chatgpt.com
Web search	Only if you call the web search tool	On, the way the app decides
Sources	Only what your call asks for	Every citation, in order, with its source type
Brands named	You parse them yourself	Listed with rank, recommendation and sentiment
Proof	None kept	Raw response with its SHA-256 hash
Billing	Per token	Per answer; failed answers are free

One question, three different answers

We asked “best CRM for early-stage startups” on each engine from the US on the same morning. Each card is what the API returned: how the answer opens, the brands it named in order, and the first sites it cited.

Where each answer ranks each brand (1 = named first; – = not named)

Brand	ChatGPT	Google AI Overview	Google AI Mode
Attio	1	3	1
HubSpot	2	1	2
Pipedrive	3	2	3
Zoho CRM	4	5	– not named
Close	5	– not named	4
Ringy	– not named	4	– not named
Salesforce	– not named	6 (to avoid)	– not named
Streak	– not named	– not named	5
Folk	– not named	– not named	6

All three name Attio, HubSpot and Pipedrive, but not in the same order, and after those three the lists split. They barely share sources either: ChatGPT and the AI Overview cite no site in common, and AI Mode shares only attio.com and hubspot.com with ChatGPT. Watching one surface tells you little about the others.

ChatGPT citations and sources

When ChatGPT searches the web, it cites the pages it read, and those pages decide which brands it names. In our study of 2,381 answers to 300 software buyer questions, ChatGPT cited 10 sources per answer on average, led by zapier.com, itechguides.com, rankite.com, dupple.com and techradar.com. It almost never cited YouTube or Reddit (YouTube once and Reddit 3 times in 5,986 citations), while Google's AI Overviews cited YouTube in 85% of their answers.

The API returns every citation in the order ChatGPT gave it, with the page title and a source_type, so you can see at a glance whether a brand is named because of its own site, a review site, Reddit or the press.

• Position is the order ChatGPT cited the page in, starting at 1.
• Domain and title tell you exactly which page it was.
• Source type groups pages the same way for every engine, so you can compare ChatGPT with Google AI Overviews and AI Mode.

Read the full study: what AI engines cite when buyers ask what to buy.

Track brand mentions in ChatGPT over time

ChatGPT searches and samples every time it answers. In our study it never cited exactly the same set of sites twice for the same question (0 of 298), so one check is a sample of one, and a ChatGPT rank tracker has to ask repeatedly and count. Create a project with a brand, its competitors and its buyers' questions, put it on a daily or weekly schedule, and read:

• GET /v1/projects/{id}/visibility: how often ChatGPT names the brand, with error bars.
• GET /v1/projects/{id}/competitors: each competitor's share of the answers.
• GET /v1/projects/{id}/sources: the sites ChatGPT cites for those questions.
• GET /v1/projects/{id}/history: all of it over time, by day or week.

Add a webhook and your server hears when each run is done. To check one product by hand instead, see how to check if ChatGPT recommends your product.

Call the ChatGPT API from TypeScript or Python

The SDKs wrap the same request, retry safely without charging twice, and wait for the answers for you. Free sandbox keys (aiv_test_…) return realistic sample answers, so you can build the whole feature before spending a credit.

TypeScript (npm install @aivisibilityskill/sdk)

import { Aiv } from "@aivisibilityskill/sdk";

const aiv = new Aiv({ apiKey: process.env.AIV_KEY! });

const job = await aiv.checks.create({
  questions: ["best CRM for early-stage startups"],
  engines: ["chatgpt"],
  locale: "en-US",
});
const done = await aiv.jobs.wait(job.id);

for (const answer of done.answers ?? []) {
  console.log(answer.brands?.map((b) => `${b.rank}. ${b.name}`));
  console.log(answer.citations?.map((c) => c.domain));
}

Python (pip install aivisibilityskill)

import os
from aivisibilityskill import Aiv

aiv = Aiv(api_key=os.environ["AIV_KEY"])

job = aiv.checks.create(
    ["best CRM for early-stage startups"],
    engines=["chatgpt"],
    locale="en-US",
)
done = aiv.jobs.wait(job["id"])

for answer in done["answers"]:
    print([(b["rank"], b["name"]) for b in answer["brands"]])
    print([c["domain"] for c in answer["citations"]])

For thousands of questions, send them as one batch of up to 20,000 answers and get one webhook when it's done. Full guides: SDKs, API reference, OpenAPI 3.1 spec.

ChatGPT API pricing

You pay per answer, in credits. Pick the speed per request. A dollar buys 1,000 credits on a pack; monthly and volume plans include more. An answer reused from a recent identical one (max_age_hours) costs half, and an answer that fails costs nothing.

Every account gets the same API and every engine, starting with 500 free credits and no card. Volume plans start at $500 a month for 750,000 credits, with invoicing and a 99.5% uptime commitment: see volume plans.

Countries and languages

Ask ChatGPT from 192 countries and territories, in the buyer's language, with a locale like en-US, de-DE, pt-BR or ja-JP. Google AI Overviews, AI Mode and Gemini cover 240. GET /v1/locales lists every supported pair for free, and a pair that isn't supported is refused before anything is charged.