Get Hotel Sources

Per-provider normalized-host recurrence over the period-selected runs.

For each publisher-host-capable provider (OpenAI, Perplexity, Claude), returns the normalized hosts that recurred across this experiment's source-captured results, each with n = distinct source-captured results the host appeared in, alongside M = the provider's source-captured result count for the period (captured-zero included, null-capture excluded). Hosts are normalized (lowercased, leading www. stripped) — NOT registrable domains; subdomains are preserved.

Gemini is surfaced as a distinct reconstructed state: its captured URLs are vertexaisearch redirects with no derivable publisher host, so hosts are reconstructed best-effort from Gemini's own source labels (GeoResultSource.provider_title, i.e. its web.title) — never the redirect URL — with a real M and a per-occurrence unresolved_sources count for labels that are not domain-shaped.

Hosts are ranked and trimmed to a single global top-N set (ranked by the summed n across all providers, keyed by host string), so every returned host appears for every provider that cited it — no per-provider false-absence. providers stays provider-major (per-provider M/state/hosts); the response also carries the global total_hosts (pre-trim distinct-host count) and the contributing runs.

limit/offset select the [offset, offset + limit) window of that global ranking. They bound the response only: total_hosts is always the full pre-trim distinct-host count, so it stays the pagination denominator, and an offset past the end returns empty hosts for every provider rather than an error. Sending neither reproduces the historic top-_SOURCES_TOP_N response exactly.