The API for GEO

All the analysis data GEO needs in one API — real-browser answers from seven engines, cause-and-effect between answers and their sources, GEO monitoring, and more to come. Any country, structured JSON.

ChatGPT

Gemini

Perplexity

Google AI Overviews

Google AI Mode

Naver AI Brief

Naver AI Mode

2,000 free credits · no card required

querying.ai/dashboard/playground

How do AI search engines choose which sources to cite?

Response

COMPLETED · 1ad65a0f-…

COMPLETED

CHATGPT10.5s2 credits6 sources

Searched for

how AI search engines choose sources to cite retrieval ranking authority relevance

AI search engines generally choose citations through a retrieval-and-selection pipeline, not by simply picking the highest-ranking web page like a traditional search engine. The exact algorithms are proprietary and vary by product, but the broad process looks like this: answerbench.org

Find candidate sources

The system first gathers potentially relevant material from search indexes, web crawls, licensed databases, or other retrieval systems.
A user question may be rewritten into multiple related searches so the system can find different angles of the answer. bluefrog.ai

Match sources to the question AI systems look for sources whose content actually answers the user’s intent, often using semantic matching rather than only keyword matching. A page about “best running shoes for flat feet” may be preferred over a general shoe article because it contains the specific information needed. thestacc.com
Score reliability and usefulness Common signals include:

Relevance: Does the source directly support the claim being made?
Authority: Is the source credible for this topic?
Freshness: Is it current enough, especially for news, prices, laws, or software?
Factual clarity: Does it provide specific, verifiable information?
Extractability: Can the system easily identify a passage that supports the answer? lightspacelabs.com

Select evidence for specific claims The system is usually not asking “which website is the best overall?” It is asking something closer to: “Which source best supports this particular sentence?” A government page might be chosen for a legal fact, a research paper for a scientific claim, and a company’s documentation for a product feature. answerbench.org
Generate the answer and attach citations After selecting useful passages, the AI produces a response and links citations to the claims those sources support. Some systems cite many sources; others prefer fewer sources that they rely on more heavily. arxiv.org

A few things that often surprise people:

The top Google result is not guaranteed to be cited. AI systems may choose a lower-ranked page if it provides a clearer or better-supported answer. lightspacelabs.com
Being authoritative is not enough. A highly reputable site can be skipped if the relevant information is buried, vague, outdated, or difficult for the system to extract. answerlab.au
Citations are not a perfect audit trail. A citation means the system used or found the source relevant, but it does not automatically mean every statement in the answer is fully verified by that source. answerbench.org

A simple mental model: AI search first finds possible evidence, then chooses the evidence that best supports the answer it wants to give. The winning source is often the one that is both trustworthy and easy to use as evidence. bluefrog.ai

Sources

1How AI Answer Engines Choose Which Sources to Cite — AnswerBenchhttps://answerbench.org/articles/how-ai-answer-engines-choose-sources
2How do AI assistants decide which sources to cite? | Bluefroghttps://bluefrog.ai/answers/assistants-choose-which-sources-to-cite/
3How AI Search Engines Choose Sources to Citehttps://thestacc.com/blog/how-ai-search-engines-cite-sources/
4How AI Search Engines Decide Which Sources to Cite | Lightspace Labs Blog | Lightspace Labshttps://www.lightspacelabs.com/blog/how-ai-search-engines-decide-which-sources-to-cite
5From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platformshttps://arxiv.org/abs/2604.25707
6How AI Search Engines Decide Which Sources to Cite | AnswerLabhttps://answerlab.au/blog/how-ai-decides-citations

See the answer. Know why. Track it.

One API, three layers of GEO: collect the answers your customers actually see, trace which sources shaped them, and keep score over time.

Answers

Real consumer answers from ChatGPT, Gemini, Perplexity, Google AI and Naver AI — per country, as structured JSON.

GEO Research

Feed an answer and its citations back in and see exactly which source shaped which sentence — with verified evidence behind every claim.

Monitors

Scheduled prompt sets scored for brand presence and citations — your share of AI answers, measured over time.

Seven engines, one API — and you can see who's up

Every engine, across regions, at scale. Status is measured, not asserted: an engine under repair says so instead of showing a green light it hasn't earned.

ChatGPT

What each surface is, and what comes back

Seven consumer surfaces, not seven wrappers around one model. Each returns its own structured payload — the field names below are exactly what lands in your webhook.

ChatGPT

ChatGPT as anyone gets it for free, answering with web search on. So you get more than prose: the queries it fanned out to before answering, the citation pills behind them, and its shopping cards. Whichever model actually served the answer comes back in `model`, so you always know what you measured.

textsourcesmarkdownsearchQueriesshoppingCardsmodel

Gemini

The consumer Gemini app, not the Gemini API. Every source carries Google's own 0–100 grounding confidence, so you can weight citations instead of treating them all alike.

textsourcesmarkdownmodel

Perplexity

Perplexity's answer, with the widest structured payload here: the queries its search model actually issued, the follow-ups it suggests, and its media, shopping, place and hotel blocks.

textsourcesmarkdownsearchQueriesrelated_queriesvideosimagesshopping_cardsplaceshotels

Google AI Overviews

The AI Overview box on a Google results page. It is conditional — a query Google writes no overview for says so plainly — and the organic results and People Also Ask from the same fetch come with it.

aiovervieworganicResultspeopleAlsoAsk

Google AI Mode

Google's AI Mode tab, a conversation rather than a summary box. The full answer arrives in `text` and `markdown` alike — page furniture already stripped — with every source Google cited.

textsourcesmarkdown

Naver AI Brief

The AI 브리핑 box on Naver's integrated search. Naver only briefs queries it judges briefable, so an empty briefing is a real answer rather than a failure — and the organic web documents beneath the box come along either way.

textsourcesmarkdownorganicResultsrelatedQuestions

Naver AI Mode

Naver's AI 탭 — always on, conversational, and the one surface that returns Korean shopping cards with price, discount rate, review score and delivery.

textsourcesmarkdownproducts

Why not just call the LLM API?

When you call a provider's API, you get model output — not what your customers see on their screen.

Consumer surface ≠ LLM API

ChatGPT's consumer app applies search grounding, citations, regional routing, and real-time retrieval that the OpenAI API does not. GEO and AEO analysis needs 'what does the customer actually see' — not 'what can the model generate'. querying.ai captures the real consumer surface.

Structured data beyond the answer

Ask Perplexity one question and ten fields come back — including the queries its search model actually issued, which is the closest thing there is to a list of what it thought your brand was about. ChatGPT returns its search fan-out and shopping cards; Naver AI returns Korean product cards with price, discount rate, review score and delivery.

The same prompt is a different answer per country

Ask about the best running shoes from Seoul and from Chicago and you get different brands, different shops, different citations. Every task takes a `country`, and the engines are queried from that market — so a rank you report for KR is a rank a Korean customer would actually see.

An empty answer is a finding, not an error

Google AI Overviews and Naver's AI 브리핑 are conditional: for many queries the engine writes no summary at all. Learning that your category's top query gets no AI answer is a GEO result worth having, so we report it as one — never dressed up as a failure, and never invented to fill the field.

Up to 85% lower cost

OpenAI Responses API with web search: ~$10 per 1,000 calls. querying.ai ChatGPT: $4/1k on Basic, $1.90/1k on Max. Not the same product — but for teams that need consumer-surface data, the savings are concrete. Failed tasks are never charged.

Status you can check before you trust it

Every engine's health is computed from real task outcomes in a 30-minute window, not from a synthetic ping that can pass while customer traffic fails. It is published per engine at status.querying.ai, so you can see which surface was healthy on the day your data was collected.

Built for these teams

querying.ai gives every team that needs to understand AI search a single, structured data source.

GEO & AEO Monitoring

Track how AI search surfaces mention your brand, products, and competitors — across engines, countries, and prompts.

Competitive Intelligence

See what AI engines recommend in your market in real time: which brands get cited, which products surface.

Agency Dashboards

Monitor dozens of clients across engines and regions simultaneously with batch APIs and webhooks.

Research & Analysis

Study AI search behavior across markets with structured data for academic research, content strategy, and SEO.

Submit a task. The answer comes to you.

Every task is async — real engines take seconds to answer. Post it, get an id back, and the result hits your webhook as JSON: the answer, citations, and timing.

01You submit a task

curl -X POST https://api.querying.ai/v1/async/task \
  -H "Authorization: Bearer vt_live_…" \
  -H "Content-Type: application/json" \
  -d '{
    "taskType": "GEMINI",
    "payload": { "prompt": "best wireless earbuds 2026" },
    "webhook": { "url": "https://yours.app/hooks/querying" }
  }'

02Your webhook receives

{
  "task": {
    "id": "8f2c1e40-…",
    "taskType": "GEMINI",
    "status": "COMPLETED",
    "createdAt": "2026-07-17T04:12:00.000Z"
  },
  "credits": { "creditsToCharge": 1, "creditsCharged": 1 },
  "response": {
    "text": "The best wireless earbuds in 2026 are …",
    "sources": [
      { "position": 1,
        "url": "https://example.com/review",
        "label": "Best earbuds, reviewed" }
    ],
    "timing": { "totalMs": 7400 }
  }
}

Retried five times with backoff if your endpoint is down. Read the docs ↗

Pay for scale, not for surprises

Bigger plans buy cheaper credits — down to $0.95 per 1,000 on Max. Failed tasks are refunded, so an engine outage is not your bill.

Basic

$20/month

10,000 credits included

A single product's tracking, or a build you're still shaping.

≈ 5,000–10,000 tasks / month

10 concurrent tasks
Every engine we serve
Every region we serve
Webhooks + polling
Email support

Recommended

Pro

$200/month

190,000 credits included

A brand tracked daily across engines, countries and prompts.

≈ 95,000–190,000 tasks / month

25 concurrent tasks
Everything in Basic
Batch submission
Priority support

Scale

$500/month

500,000 credits included

Several brands, or one brand tracked wide and deep.

≈ 250,000–500,000 tasks / month

50 concurrent tasks
Everything in Pro
Usage review on request
Priority support

Max

$1000/month

1,050,000 credits included

An agency or platform running many brands at once.

≈ 525,000–1,050,000 tasks / month

100 concurrent tasks
Everything in Scale
Usage review on request
Priority support

Enterprise

Custom

Volume pricing

Committed volume, invoicing, and terms that fit procurement.

Everything in Max
Custom engine mix
Invoice billing
Shared support channel

Talk to us

Negotiated

Task estimates vary by engine because each engine uses 1–2 credits.

See what a credit buys, per engine →