Skip to content

New AI assistant answers: ChatGPT, Gemini, Copilot and AI Mode by API

SerpKite
Get API key
Docs menu / Search providers

Search providers

When Google is unavailable, engine=auto or a per-key setting lets SerpKite fall back to Brave, Bing, Yahoo, DuckDuckGo, Mojeek or Yandex, named in meta.engine.

View as Markdown

SerpKite is a Google SERP API, and search requests use Google by default unless the request or API key opts into another policy. Behind each Google request, SerpKite already retries internally when an attempt is blocked or slow. That failover is always on, costs you nothing extra, and never changes which search engine you get.

For workloads where some answer beats no answer (agents, chatbots, research tools), you can also let SerpKite fall back to other search engines when Google is blocked or times out. This is opt-in, per request or per key, and the response always says which engine produced it.

The engine parameter

engine is accepted by /v1/search, /v1/news, /v1/images, /v1/videos, /v1/maps and /v1/places (Yandex), and on /v1/place together with oid:

Value What happens
google (default) Google only. SerpKite retries Google internally before failing, never switching to another engine.
auto Google first; if Google is blocked or times out, the next enabled fallback engine that serves this vertical answers.
consensus (search only) Several independent engines in parallel, merged and ranked by agreement. See Consensus mode.
One provider, e.g. brave That provider only.
A list, e.g. ["google", "brave"] Only those providers, in SerpKite’s route order. In a query string, send engine=google,brave.
curl https://api.serpkite.com/v1/search \
  -H "Authorization: Bearer $SERPKITE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"q":"best espresso machine 2026","country":"us","engine":"auto"}'

Parameter errors (a bad country, an unknown parameter) never trigger a fallback: they return 400 straight away. auto and consensus can’t be combined with other values, and naming only providers that don’t serve the endpoint (or aren’t enabled) returns 400 invalid_request with a link to this page.

Which providers serve which verticals

Providerengine/v1/search/v1/news/v1/images/v1/videos/v1/maps/v1/places/v1/placeOther verticalsNotes
Google (default)googleYesYesYesYesYesYesYesYes, allThe default for every request. Every vertical and all Google-only blocks.
Brave SearchbraveYesYesYesYesNoNoNoNoIts own independent index.
BingbingYesYesYesYesNoNoNoNoMicrosoft's index.
YahooyahooYesYesNoNoNoNoNoNoResults largely come from Bing's index.
DuckDuckGoduckduckgoYesYesYesYesNoNoNoNoWeb results largely from Bing's index; news, images and videos from its own feeds.
MojeekmojeekYesNoNoNoNoNoNoNoWeb results only, from its own independent index.
WikipediawikipediaYesNoNoNoNoNoNoNoThe best matching article, from Wikipedia's official API. A source for engine=consensus (or by name), never an auto fallback.
YandexyandexYesNoNoNoYesYesYesNoWeb results from Yandex's own index (strongest for Russian and CIS queries), and organisations from Yandex Maps, the most complete map data for Russia and much of the CIS: engine=yandex on search, maps and places, oid on /v1/place. With engine=auto, the last fallback.

Each engine has a page with a live example response: Brave, Bing, DuckDuckGo, Yahoo, Yandex and Yandex Maps. Multi-engine search compares them side by side.

Google is the default configured provider; individual requests can still fail or be blocked. The fallback engines are enabled per deployment, so engine=auto only uses the ones that are currently switched on, and the route can change as SerpKite adds or pauses providers. Verticals not listed for a fallback engine (reviews, shopping, scholar, patents, autocomplete) are answered by Google only: engine=auto simply uses Google there, and naming a fallback engine returns 400.

Honest labelling: meta.engine and meta.route

A result is never relabelled. meta.engine names the provider that actually produced it, and meta.route lists each provider attempt behind a fresh result, in order, with its outcome and duration:

{
  "request": { "endpoint": "search", "engine": "auto", "q": "best espresso machine 2026", "country": "us" },
  "meta": {
    "request_id": "req_01J8ZK4M6Q2V7",
    "credits_used": 1,
    "cached": false,
    "engine": "brave",
    "route": [
      { "provider": "google", "outcome": "blocked", "ms": 2140 },
      { "provider": "brave", "outcome": "ok", "ms": 612 }
    ],
    "parse_quality": "ok",
    "latency_ms": 2790
  }
}
  • request.engine echoes what you asked for (google, auto, one provider or a list).
  • meta.route outcomes include ok, partial, empty, blocked, timeout and error.
  • Cache hits have no meta.route. The cache never mixes providers: a Google-only request is only ever served a cached Google result.
  • With format: "markdown" there is no meta; SerpKite’s response headers still carry the request ID and credits.

If you need Google specifically (rank tracking, SEO audits, anything that compares against google.com), keep the default and check meta.engine if you’re unsure how a key is configured.

What a fallback result lacks

Fallback engines map into the same response shape (results with title, link, snippet…), so your parser keeps working, but their pages simply don’t have Google’s extras:

  • No knowledge graph, answer box, People Also Ask, top stories or local pack unless that engine has an equivalent block. Fields a provider doesn’t have are absent or empty.
  • Ranking, snippets and result counts come from that engine’s own index, so positions are not comparable with Google’s.
  • parse_quality is judged per provider.
  • location (or a uule naming a place) is applied by Google, Brave and Yandex. Bing, DuckDuckGo, Yahoo, Mojeek and Wikipedia have no such setting, so with a location they are skipped (a 400 if you named one) rather than answering for somewhere else. An engine that can’t find the place is skipped the same way.

Billing follows the provider

You pay for the provider that answered, and only for that one. Every provider currently costs the same credits as Google for the same request, so a fallback never costs more. Failed, blocked and empty attempts along the route are free, just like a failed Google request: the request is settled once, at the price of the answer you get. X-Credits-Used and meta.credits_used show the actual charge. See Credits and billing.

Consensus mode

engine: "consensus" on /v1/search asks several independent search indexes at once and merges what they return, like a metasearch engine:

  • Up to three providers from its route answer, one per index. A provider that fails or comes back empty hands its turn to the next index in the route, so one flaky source doesn’t leave a single-engine answer. Yahoo and DuckDuckGo serve Bing’s results, so at most one of the three is Bing-backed. Wikipedia’s official API (the best matching article) can be one of them; it is only ever a consensus source, never an auto fallback.
  • Results are deduplicated by URL (ignoring www., a trailing / and tracking parameters such as utm_*) and ranked by how many providers returned them. Among equally agreed results, Wikipedia articles and results whose title and snippet contain your query’s words come first, then the best position any provider gave.
  • Each result has sources, the providers that returned it. meta.engine is "consensus" and meta.route lists every provider started.
  • include_domains and exclude_domains work too: every provider receives the domain operators and the filters apply to the merged rows. A provider whose rows were all filtered out is not billed.
  • It stops as soon as num unique results are in hand, or after a few seconds with whatever has arrived. Providers still running are cancelled (canceled in the route, timeout for the deadline) and ones never started cost nothing.
{
  "results": [
    { "position": 1, "title": "The Best Espresso Machines", "link": "https://www.seriouseats.com/best-espresso-machines-5185482", "domain": "seriouseats.com", "sources": ["brave", "bing"] },
    { "position": 2, "title": "Espresso machine - Wikipedia", "link": "https://en.wikipedia.org/wiki/Espresso_machine", "domain": "en.wikipedia.org", "sources": ["wikipedia"] }
  ],
  "meta": { "engine": "consensus", "credits_used": 3, "route": [ { "provider": "brave", "outcome": "ok", "ms": 640 }, { "provider": "wikipedia", "outcome": "ok", "ms": 410 }, { "provider": "bing", "outcome": "ok", "ms": 1180 } ] }
}

Billing: one page at each provider that returned results, added up (so up to three times a single search). Providers that failed, came back empty, were blocked or never started are free, and a cache hit (max_age) costs half. The whole reservation counts against your spend cap and key limit. Consensus answers are cached separately: they are never served to a single-engine request, or the reverse.

Per-key setting

In the dashboard, API keys have a switch: Allow fallback to other search engines (off by default). With it on, requests from that key that don’t send engine behave as engine=auto. An explicit engine always wins, so engine: "google" keeps a single request Google-only even on a key with fallback on.

This is useful when you can’t change the calling code, for example an agent framework or an MCP client that doesn’t expose the parameter.

Yandex

Yandex runs its own index, the strongest for Russian-language and CIS queries. SerpKite answers /v1/search from Yandex web search, and /v1/maps, /v1/places and /v1/place from Yandex Maps, in the same response shape as Google.

  • engine: "yandex" on /v1/search returns Yandex’s organic web results: title, link, domain and snippet, 10 per page. country picks the Yandex region (ru, ua, kz, tr, us, de and other large markets), location narrows it to a city or region (strongest for Russia and the CIS), and page pages through results.
  • There are no ads, answer boxes or related searches. num above one page walks Yandex’s pages like Google’s, one credit per page that returned results.
  • time and safe: "active" aren’t available on Yandex: a request that names Yandex returns 400 with them, and with auto Yandex is skipped.
  • meta.engine is yandex. It costs the same credits as Google, and failed, blocked and empty requests are free.
curl https://api.serpkite.com/v1/search \
  -H "Authorization: Bearer $SERPKITE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"q":"погода москва","country":"ru","engine":"yandex"}'

Yandex Maps

  • engine: "yandex" on maps and places searches Yandex Maps. ll and location place the search, and page pages through results. The first page has each place’s rating, rating_count and thumbnail; later pages have the rating only.
  • Without language, Yandex Maps answers in Russian: its own names, addresses and categories. language: "en" returns its English version (addresses and categories in English, place names as the business writes them), "tr" Turkish and "kk" Kazakh. English place names in q (“coffee in london”) are understood in English.
  • oid on /v1/place looks up one organisation by its Yandex Maps ID (the number in yandex.ru/maps/org/…/<oid>/) and implies engine=yandex. The result has oid and yandex_maps_url instead of the Google IDs, structured weekly_hours (closed days, 24-hour and past-midnight intervals explicit), photos, phones and website.
curl https://api.serpkite.com/v1/maps \
  -H "Authorization: Bearer $SERPKITE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"q":"кофейня","ll":"@55.757,37.615,14z","engine":"yandex"}'

With engine: "auto" (or a key with fallback on), search, maps and places try Google first and fall back to Yandex last. A Google place_id, cid or fid is never sent to Yandex, and an oid never to Google. Results from Yandex Maps have no place_id, so look them up again with their oid; /v1/reviews is Google only.

Always Google

Some surfaces stay on Google regardless of engine or the key setting:

  • The Custom Search compatible endpoint (GET /customsearch/v1), because it promises Google Custom Search results.
  • Every vertical without a fallback engine in the table above.

Related: Choose search providers, Common parameters, Errors, Caching.

Last updated: 2026-11-05