RAG pipelines are only as good as what they retrieve, and model snapshots go stale the day they ship. Search-engine result APIs give you links without answers; model APIs give you answers without the live retrieval a consumer surface runs. Grounding on captured AI-surface answers gives you both — current, source-backed answer text your pipeline can quote, verify against, or decompose — in one contract per surface.
Real endpoints, real field names — nothing below is pseudocode.
step 1 / 4
Add ?mode=sync (or Prefer: wait) to block for inline Envelopes instead of polling — built for a pipeline step where a user is waiting. Sync and async are the same endpoint and body, with no sync surcharge.
curl -sS -X POST "https://api.aisearchapi.dev/v1/search?mode=sync" \
-H "Authorization: Bearer $AISEARCH_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"query": "current EU AI Act enforcement status",
"surfaces": ["perplexity"],
"regions": [{ "country": "DE" }]
}'
# → 200 with the terminal Envelope inlinestep 2 / 4
Ask for ?view=flat and skip the nesting: surface, status, text, markdown, sources — the five fields a grounding step consumes. The full four-section Envelope is one query-param away when you need provenance.
{
"surface": "perplexity",
"status": "completed",
"text": "the answer as plain text",
"markdown": "the answer as markdown",
"sources": [{ "url": "https://…", "title": "…" }]
}step 3 / 4
Use answer.markdown as grounded context and evidence.sources as the citation candidates your product shows. Because the shape is identical across surfaces, swapping or ensembling grounding surfaces is a config change, not a new parser.
step 4 / 4
A surface with no answer completes as surfacePresent:false (never a fabricated answer your RAG step would ingest as truth), and an unbacked lane fails loud with DRIVER_UNAVAILABLE. Honest empties are a feature in a grounding pipeline — branch on them.
Every surface returns the same canonical Envelope, so these fields mean the same thing whether the capture came from ChatGPT, Perplexity or an AI Overview.
answer.markdownAlways-populated, normalized markdown — drop-in grounded context for a generation prompt.
evidence.sources[]Citation candidates with urls and titles — what your product displays as "where this came from".
provenance.capturedAtFreshness stamp per capture — your pipeline knows exactly how current its grounding is.
provenance.surfacePresentA truthful empty instead of a hallucination-shaped blank — branchable in code.
job.status + warningscompleted / partial / failed with machine-readable reasons — grounding steps need failure semantics, not mystery nulls.
Sync mode is bounded server-side and designed for single-surface, single-region calls where a human is waiting; if the wait runs long it degrades gracefully to a 202 you poll. Real browser captures take longer than a cached-index lookup — if you need sub-second SERP snippets and no answer layer, a classic SERP API is the better fit and we say so. Sync concurrency is plan-limited; high-QPS grounding should mix sync for interactive paths with async + webhooks for the rest.
Terms used here, defined precisely: Envelope · Absence-is-data · API surrogate
A search API returns links and snippets; your pipeline still has to read, rank and synthesize them. A captured Perplexity or ChatGPT answer has already done live retrieval and synthesis on the real surface, and arrives with its citations attached — one hop to grounded, source-backed context. Many pipelines use both: google_search for breadth, an answer surface for synthesis.
It runs the same real-browser acquisition inline, so it is bounded by genuine surface latency — captures take longer than an index lookup, and we will not quote a fake number. The honest engineering answer: use sync for the interactive path where the grounding is the product, cache aggressively, and push everything else through async + webhooks.
That failure mode is designed out: an absent answer is an explicit surfacePresent:false with a surface_absent warning (and costs nothing), and a failed lane is an explicit failed status with an acquisition_failed reason. Your pipeline branches on contract fields instead of guessing whether an empty string means "no answer" or "broken scraper".
Yes — the same captures work offline: batch-submit query sets, store the Envelopes, and you have dated, source-attributed, provenance-stamped answer data for evals and fine-tuning corpora. The durable raw payloads (job.artifacts.rawKey) preserve the verbatim upstream for reproducibility.
500 free credits, charged only on successful captures — the whole workflow above runs on two endpoints.