Skip to content

See what AI actually answers

The LLM Scraper API does not call the models' APIs. It captures the answer a real user sees in ChatGPT, Gemini, Google AI Mode, Perplexity or Copilot, from a specific country at a specific moment, with the sources the interface cites. That is the data the official APIs cannot give you.

  • Five engines: ChatGPT, Gemini, AI Mode, Perplexity, Copilot
  • Country-level capture through the residential network
  • Citations extracted as numbered sources with title and URL, not buried in prose
  • Web search on or off, follow-up prompts
  • Structured JSON: the answer text with [n] markers plus its citations
  • Batch hundreds of prompts in one asynchronous job
100 credits per request, any engine. Failed captures are free. No card to start.
MCP readyConnect your agent in one command:npx -y @datafuel/mcp initClaudeOpenAI / ChatGPTGeminiCursorMistral
LLM Scraper · live capture
Queued
Capturing · US
JSON ready
Request
engineopenai
prompt"best CRM for startups"
websearchtrue
proxy_countryUS
What the user sees
CChatGPT· New York, US · 09:41
For an early-stage startup, HubSpot [1] and Pipedrive [2] are the most recommended; Attio if you want a data-first CRM…
g2.comhubspot.comtechradar.com
{
"text":"HubSpot [1], Pipedrive [2]…",
"citations":[
{ "n": 1, "title": "G2", … },
{ "n": 2, "title": "HubSpot", … },
{ "n": 3, "title": "TechRadar", … }
]
}
Why it matters

Buyers ask AI first. Nobody tells you what it says.

2.5B+queries a day processed by ChatGPT
~48%of Google searches show an AI Overview
89%of B2B buyers consult generative AI while purchasing
37%of product-discovery queries now start in an AI interface
Google gives you Search Console

Impressions, clicks, position, queries. Twenty years of tooling to know where you stand.

ChatGPT gives you nothing

No impressions, no dashboard, no way of knowing what it says about you. The LLM Scraper API is that missing dashboard: the answers, the sources, the changes, by country.

Everything in one request

What the endpoint does, at a glance.
Five enginesopenai, gemini, google_ai_mode, perplexity, copilot — one parameter.
Geographic targetingproxy_country runs the query from a residential exit in that country.
Citation extractionThe sources the interface shows come back as numbered citations with title and URL.
Follow-up promptsfollow_up_prompt continues the same conversation, like a real user would.
Web search togglewebsearch: true lets the engine browse; false captures the model alone.
Point-in-time captureEach capture is a task with its own processed_at, so runs line up over time.
Structured JSONAnswer text with [n] markers and the citations they point to, in one predictable shape.
Batch jobsPOST /job with prompts[] runs hundreds of prompts asynchronously.
Capabilities per engineWhat each interface exposes to a real user, and therefore to you.
Engine
Country
Citations
Follow-up
Web search
ChatGPT
Perplexity
Gemini
Copilot
Google AI Mode

Follow-up support reflects the current release; engines add and remove interface features without notice, so the matrix is kept in sync with the docs.

01 · Same question, different answers

Where the user is changes the answer

The same prompt asked from Milan, New York and Singapore returns three visibly different answers, with different brands recommended. proxy_country runs the capture from a residential exit in that country, so you see exactly what a local user sees.

Geography changes the answer. No official API can show you that.

02 · Who gets cited

The sources are the signal that counts

As the answer streams in, every reference the interface shows becomes a numbered citation with its title and URL, matched to the [n] marker in the text. You get citations as data, not as footnotes to parse out of prose.

Citation frequency weighs roughly 35% in whether a source is included in AI answers. It is the number to track.

03 · How it changes over time

Model updates rewrite your visibility

Run the same prompts every week and your position in the answers becomes a time series. When a model update lands, the curve jumps, without notice and without a changelog. Continuous monitoring is the only way to see it happen.

Every capture is timestamped, so a change is a data point, not an anecdote.

04 · Multi-engine in parallel

One prompt, five engines, one table

One prompt goes to ChatGPT, Perplexity, Gemini, Copilot and Google AI Mode, one request per engine. The answers come back in the same JSON shape and line up in a comparable table: who is mentioned, who is cited, where.

Five calls, one per engine, the same JSON shape.

Use cases

What teams measure with it

The real jobs behind AI-answer monitoring, from brand teams to the platforms that serve them.

Brand visibility & share of voice

How often you appear, in what position versus competitors, on which engines.

Citation source tracking

Which domains the AI cites in your category. The direct input to a GEO strategy.

Competitor benchmarking

How the AI describes and recommends competitors compared with you.

Sentiment & accuracy

What the AI says about you, and whether it is correct. For finance, healthcare and regulated sectors this is brand governance with a verifiable record.

Content gap analysis

A competitor recommended on a category question where you are absent is usually a content gap.

Multi-market monitoring

How positioning shifts between the United States, Europe and APAC, engine by engine.

For platformsBuilding a GEO tool? Start here.

GEO platforms are customers, not competitors. Bring your own logic, pay per check, and skip running five browser farms yourself. Same API, same 100-credit price, batch and idempotency built in.

Per-check pricing100 credits
Batch promptsPOST /job · prompts[]
Safe retriesIdempotency-Key
Read the API reference
Why not the official API

The model API answers you. The interface answers your customers.

The first objection from anyone who knows the space, resolved in one table.

Official model API
LLM Scraper API
What you get
A datacenter response to your prompt
What a real user sees in the product
Geolocation
None
Country, state, city
Interface citations
Often absent
Extracted as a list of sources
Web search
Depends on the endpoint
Controllable per request
Coverage
One vendor at a time
Five engines, one call
Pricing

Pay only for successful requests

Pick a monthly plan, buy credits once, or run unlimited threads. Failed requests are never billed.

Every plan includes every endpoint and every interface (API, MCP). One shared credit balance: you only pay for volume, and only for successful requests.

How many requests a month?10,000 LLM captures / month
≈ 1,000,000 credits · Growth
103K10K35K120KCustom
Free

Test every endpoint with real credits. No card required.

$0one-time trial
1,000 credits, one-time10 LLM captures2 concurrent threads
  • Trial credits valid for 1 year
  • Every endpoint, API + MCP
  • Community support
GrowthBest fit

For production crawlers and agents that need headroom.

$79/ mo, billed monthly
1,000,000 credits / month10,000 LLM captures25 concurrent threads$7.90 per 1K captures
  • Everything in Starter
  • Auto top-up
  • Priority chat support
Business

For teams shipping data products on a schedule.

$199/ mo, billed monthly
3,500,000 credits / month35,000 LLM captures50 concurrent threads$5.69 per 1K captures
  • Everything in Growth
  • Usage alerts per API key
  • Account manager
Scale

High volume at the lowest per-page rate. Checks out instantly.

$499/ mo, billed monthly
12,000,000 credits / month120,000 LLM captures150 concurrent threads$4.16 per 1K captures
  • Everything in Business
  • Dedicated IP pool
  • 99.9% SLA
Enterprise

Committed volume, invoicing and governance.

Custom
Custom credits / month250+ concurrent threads
  • Custom volume pricing
  • Custom concurrency (250+)
  • SSO · audit log · DPA
Compare
FreeStarterGrowthBusinessScaleEnterprise
Credits / month1K one-time300K1M3.5M12MCustom
Concurrent threads2102550150250+
API keys131025UnlimitedUnlimited
JS rendering
Residential proxies
AI in-flight (your key)
Geo-targeting · 195+ countries
Batch multi-URL
Unlocker · Crawl · Map · Search · LLM Scraper
Auto top-up——
SupportCommunityEmailPriority chatAccount managerSlack + phoneDedicated + SLA
What does a request actually cost?1 credit = 1 basic page · costs don't stack: JS + Residential is 20 credits, not 5 + 10
1×Basic HTTPSimple HTML, datacenter proxy
5×JS RenderingHeadless browser for dynamic sites
10×Residential proxyReal residential IPs
20×JS + ResidentialFull power for demanding sites
FreeAI In-FlightLLM extraction on any Unlocker request with your own AI key, zero extra credits
FAQ

Frequently asked questions

Engines, targeting and billing, answered straight.

Engines & captureWhat is captured and from where.
Which models are supported?
Five engines today: ChatGPT (openai), Gemini (gemini), Google AI Mode (google_ai_mode), Perplexity (perplexity) and Copilot (copilot). You pick one per request with the engine parameter; a batch job runs all its prompts on one engine.
How does geographic targeting work?
Set proxy_country (ISO code) and the capture runs from a residential exit in that country, so the interface behaves as it does for a local user: language, sources and recommendations included.
Are citations always available on every model?
Citations are extracted whenever the interface shows sources. All five engines expose them; follow-up prompts are currently supported on ChatGPT and Perplexity. The capability matrix above is kept in sync with the docs.
What is the difference with the official model APIs?
The official API returns a datacenter response to your prompt, with no location, usually no interface citations and one vendor at a time. The LLM Scraper captures what a real user sees in the product, from a chosen country, with the sources shown, on any of five engines with one API.
Billing & volumeWhat a request costs and how often you can run it.
How often can I query?
As often as your concurrency limit allows. For monitoring, batch your prompts into a POST /job and run it on your schedule (hourly, daily, weekly); each task records when it was processed, so runs line up as a time series.
What exactly counts as a successful request?
A capture that returns the answer content. If the engine blocks, times out or returns nothing usable, the request fails and is not billed. Successful requests cost 100 credits regardless of engine or web-search setting.
Can I schedule monitoring?
Scheduling runs on your side today: call POST /job from a cron or workflow tool, then fetch results with GET /job/{id}/results. Use an Idempotency-Key so a retried run never bills twice.
ComplianceTerms of service and data handling.
Are there terms-of-service implications with the AI providers?
Automated querying of third-party AI assistants touches their terms of service. We capture only what is publicly shown to a signed-out user, rate-limit per engine and never store your prompts beyond the job's lifetime. Enterprise buyers can request our legal position paper and a DPA through their account team.
Get started

Ready to build?

Start with the free tier and scale as your project grows. No credit card, no sales call.

Talk to an engineer, not a chatbot.