surfaces: ["chatgpt"]Send a prompt, get back what a real person sees in ChatGPT — the answer, the sources it cites, the fan-out queries it ran — as one structured Envelope, with a screenshot and proof-of-page HTML kept as durable evidence. Captures run browser-first on our own stealth-browser fleet (guest sessions today), not through the sanitized OpenAI API.
curl -sS -X POST "https://api.aisearchapi.dev/v1/search" \
-H "Authorization: Bearer $AISEARCH_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"query": "best crm for startups",
"surfaces": ["chatgpt"],
"regions": [{ "country": "US" }, { "country": "GB" }],
"method": "auto"
}'
# → 202 { "jobId": "…", "children": ["….chatgpt.…"] }
# then poll GET /v1/jobs/:childId — or add ?mode=sync for an inline EnvelopeOne canonical Envelope — the same four sections every surface returns — with the fields ChatGPT actually populates. Field names below are the contract, not a summary.
answer.text · answer.markdown · answer.blocksThe answer as ChatGPT rendered it — plain text, normalized markdown (always populated), and typed blocks whose referenceIds point into the source list.
evidence.sources[]Every source with a role (cited / retrieved / related), a cited boolean, and charRanges — the exact character spans of the answer each source backs.
evidence.fanOutThe sub-queries ChatGPT actually searched, marked observed when read off the page — the raw material of GEO analysis.
evidence.mentionsThe derived brand-mentions layer: brand, ordinal position, sentiment, and the source ids supporting each mention. Separately versioned; null when not computed.
provenanceHow the capture ran: acquisition lane, consumer_ui fidelity, loginState, observed-vs-inferred model label with confidence, webSearch state, requested-vs-effective region.
job.artifactsDurable R2 keys — screenshot, proof-of-page HTML, raw payload. Not expired at 24h; your audit trail survives.
{
"job": {
"id": "job_8f3c2a1e.chatgpt.own-fleet.US",
"surface": "chatgpt",
"method": "own-fleet",
"status": "completed",
"warnings": [],
"artifacts": {
"screenshotKey": "shots/2026/06/30/job_8f3c2a1e.png",
"proofHtmlKey": "proof/2026/06/30/job_8f3c2a1e.html",
"rawKey": "raw/2026/06/30/job_8f3c2a1e.json"
}
},
"provenance": {
"acquisition": "own-fleet",
"fidelity": "consumer_ui",
"loginState": "logged_out",
"model": { "observedLabel": "GPT-5", "inferred": true, "confidence": 0.62 },
"webSearch": { "enabled": true, "known": true },
"surfacePresent": true,
"region": { "requested": "US", "effective": "US" }
},
"answer": { "text": "…", "markdown": "…", "blocks": [ … ] },
"evidence": {
"sources": [
{ "id": 0, "url": "https://…", "title": "…",
"role": "cited", "cited": true, "charRanges": [[12, 33]] }
],
"fanOut": { "provenance": "observed", "queries": ["…", "…"] },
"mentions": { "engine": "mentions@0.9.0", "confidence": 0.81, "items": [ … ] }
}
}Default browser-first capture of the real ChatGPT UI on our own fleet (guest sessions today). The official-api lane (OpenAI Responses) is an explicit pin for an api_surrogate baseline — auto never silently uses it. The managed-vendor lane is phase-2.
Phase-2 lanes are roadmap, not product — pinning one returns an honest 422 today. The full picture is on the capability matrix.
Browser-first on our own fleet: a real stealth-browser session (Bright Data Scraping Browser under our runtime) loads the actual ChatGPT UI as a guest, submits your prompt, waits for the completion signal, and normalizes the rendered answer into the Envelope. You get consumer_ui fidelity — what a person sees, including sources and fan-out — not a model API response.
No. For ChatGPT, `auto` routes to the own-fleet browser lane. The official-api lane (OpenAI Responses) exists as an explicit pin — method: "official-api" — when you want an api_surrogate baseline to diff against, and it is never selected silently.
Guest (logged-out) sessions today, and the Envelope says so honestly: provenance.loginState is recorded on every capture. The authenticated account pool that unlocks logged-in surfaces is phase-2 roadmap — we label it that way everywhere rather than selling it as live.
Five credits per successful capture, every Envelope field included — sources, fan-out, mentions, artifacts. Failed captures and empty surfaces cost nothing (charged-on-success), and new accounts start with 500 free credits.
Nine countries in v1: US, GB, DE, FR, CA, AU, IN, JP, BR — passed as structured regions objects ({ country, city?, language? }). Each region fans out into its own child job, and provenance records requested-vs-effective region plus the realized proxy egress so geo fidelity is auditable.
provenance.model separates the observedLabel (the UI string, e.g. "GPT-5") from a verified providerId, with an inferred flag and a confidence score. We never conflate a UI label with an official model id — if the identity is a guess, the Envelope says so.
500 free credits, charged only on success — and the same Envelope for every other surface when you add them.