Skip to content

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

SerpKite
Get API key

Product

Changelog

New endpoints, parameters and dashboard features, by API version. For Google's own changes, see the Google SERP changelog.

v1.12.3

Lower prices for AI answers

ChatGPT answers now cost 4 credits (was 5), Gemini 3 (was 5) and Google AI Mode 2 (was 3). Copilot Search and Brave AI are unchanged.

AI assistants

  • Lower prices for AI answers: ChatGPT 4 credits (was 5), Gemini 3 (was 5), Google AI Mode 2 (was 3); Copilot and Brave unchanged.
  • A cached answer (max_age) still costs half, the batch lane bills the same price, monitors pay the new price per run, and failed, refused and empty answers stay free.
v1.12.2

Free credits once per person

A new account signed up from the same device or network as another one gets no free credits, and its requests get 403 account_restricted until it buys a pack.

Search API (api.serpkite.com/v1)

  • New error 403 account_restricted on every charged endpoint (search, batches, crawls, extract, map, answer, monitors and MCP) for an account whose free use is paused. The message links the billing page; buying any pack lifts it.

AI assistants

  • /v1/gemini answers faster: about 12 seconds at the median, most within 5 to 30 (was about 38, most within 10 to 65).
  • /v1/chatgpt and /v1/gemini: an answer still being written at the 90-second budget is no longer lost. The request fails with 503 upstream_timeout (free) and a message that the answer is being finished in the background; the same request sent again within 10 minutes returns it, waiting for it if needed, billed once. Batch jobs collect it on their own retry.
  • A Gemini reply that is only its "having a hard time" message is retried on another connection instead of failing the request outright.

Dashboard and privacy

  • A repeat sign-up gets no sign-up, account-linking or monthly free credits, and the dashboard explains why with a link to buy credits. Invited team members are exempt.
  • Sign-ins record keyed hashes of the network and of a random sk_did device cookie (never the IP address), kept 30 and 90 days. The Privacy Policy and Cookie Policy are updated.

Agent frameworks

  • AI assistant tools: @serpkite/ai-sdk 0.6.0 adds chatgpt, gemini, copilot, ai_mode and brave_ai (factories serpkiteChatGPT, serpkiteGemini, serpkiteCopilot, serpkiteAIMode, serpkiteBraveAI; serpkiteTools adds them only when only names them), langchain-serpkite 0.4.0 adds SerpKiteAssistant, and serpkite[crewai] 0.8.1 adds SerpKiteAssistantTool.
v1.12.1

A clearer free daily limit

The daily_limit_reached error names the limit, the reset and the billing page; the dashboard shows today's usage and the owner gets one email a day when the limit is hit.

Search API (api.serpkite.com/v1)

  • 429 daily_limit_reached says how many credits the free daily limit allows (500), that it resets at 00:00 UTC and that any pack at app.serpkite.com/billing removes it. Batches, crawls, extract, map, answer and MCP return the same message.
  • Every endpoint sends Retry-After with daily_limit_reached, counting down to 00:00 UTC.
  • /v1/chatgpt answers faster: about 17 seconds at the median, most within 10 to 30 (was about 22, most within 15 to 40). Its text and markdown no longer carry the questionnaire ChatGPT sometimes appends to a recommendation or the "View retailer listing" links of product cards, and a product shown with two offer cards is listed once in products.

Dashboard and SDKs

  • The overview and Limits page show today's usage against the free daily limit, with a notice from 80% used; the account owner gets one email the first time the limit is hit each UTC day.
  • SDKs 0.8.1 (TypeScript, Go) raise daily_limit_reached at once instead of retrying it; Go adds serpkite.CodeDailyLimit.
  • The account owner gets one email the first time a request gets 402 insufficient_credits, re-sent only after credits are added and the balance runs out again.
v1.12.0

AI assistant answers: ChatGPT, Gemini, Copilot, AI Mode and Brave

Ask ChatGPT, Gemini, Copilot Search, Google AI Mode and Brave's AI answer by API as a logged-out user, and track how often they name or cite your brand.

Search API (api.serpkite.com/v1)

  • /v1/chatgpt, /v1/gemini, /v1/copilot, /v1/ai_mode and /v1/brave_ai send a prompt to the assistant as a logged-out visitor and return its answer (text, markdown), cited sources, search_queries, products and places: ChatGPT 4, Gemini 3, Copilot Search 3, Google AI Mode 2 and Brave AI 2 credits per answer. Failed, refused and empty answers are free.
  • track (up to 20 brands or domains) returns mentions: whether the answer names each term, cites it, and where.
  • Batches accept the assistant endpoints at their full price; monitors ask a prompt every run and report each answer with new_sources and mentions.
  • /v1/search returns Google's AI Overview as ai_overview when the page shows one, at no extra cost.

MCP server and SDKs

  • 23 MCP tools: new chatgpt, gemini, copilot, ai_mode and brave_ai.
  • SDKs 0.8.0: TypeScript chatgpt, gemini, copilot, aiMode, braveAi; Python chatgpt, gemini, copilot, ai_mode, brave_ai; Go ChatGPT, Gemini, Copilot, AIMode, BraveAI.
v1.11.0

Images paging: 100 per page, num honoured, no slow empty pages

Image pages continue their positions, num trims a page of up to 100 images, and a page past the last one answers at once instead of failing after a slow retry.

Search API (api.serpkite.com/v1)

  • /v1/images: one page is a block of up to 100 images (1 credit), and page: 2 now starts at position 101 (it was numbered from 11, overlapping page 1). num (1–100, default all) returns only the first num images of the page; before, it was ignored.
  • A Google page past the last one on images, news, videos and shopping is recognised as the end of the results: it returns empty results in about a second, free. Before, images and news retried it in a browser for up to 15 seconds and sometimes answered 503 upstream_blocked.
  • /customsearch/v1 with searchType=image: start=11 returns images 11–20 of the first block, as Google's CSE does; before, it returned the start of a different block.

MCP server and SDKs

  • The MCP images tool takes num; @serpkite/ai-sdk 0.5.0 adds num to its images tool.
v1.10.0

Cited answers, structured extraction and deep research

Claude writes cited answers from live Google results, extract fills your JSON schema or summarizes each page, and research tasks end in a cited report. Failed AI steps are never billed.

Search API (api.serpkite.com/v1)

  • POST /v1/answer: an answer with [n] citations, written by Claude from live Google results. depth: "fast" costs 3 credits; depth: "standard" also reads the top pages (8 credits with 3). Streams over Server-Sent Events, or returns JSON for output_schema. A failed answer refunds the model part.
  • include_answer: true on /v1/search and /v1/news adds a cited answer (+2 credits, +4 with include_content), refunded with a meta.warnings entry when the model fails.
  • /v1/extract: output_schema (+ prompt) returns each page as JSON (+4 credits per page) and summary: true adds a summary (+4). Boolean and enum schemas are answered by OpenAI's Decisions API, always schema-valid.
  • POST /v1/research: async deep research ending in a cited Markdown report: lite 30, standard 75, deep 200 credits, charged only when the report is written. Poll, cancel, or get a signed research.completed webhook.
  • Short pages that are bot checks, login walls, soft 404s or empty app shells are no longer billed or cached by webpage, extract, crawl and include_content; page monitors skip cosmetic changes.

MCP server

  • 18 tools: new answer, research and research_result; search takes include_answer, extract takes summary, output_schema and prompt.
v1.9.0

Depth searches in order, honest blocked answers, batch depth pricing

Deep searches fetch their pages in order and return more of Google's results, a search Google blocked is reported as blocked instead of empty, and from 23 November 2026 batch depth requests are billed per page.

Search API (api.serpkite.com/v1)

  • num above 10 fetches its pages in order (each starts once the page before it answered, or 1.5 s after it started, at most two at a time): Google refuses the same query arriving twice at once. Two pages in a row without results, empty or failed, end the walk. num: 100 takes roughly 10 seconds; prices are unchanged.
  • A stray "did not match any documents" page past the first is fetched once more on another session before it counts as empty.
  • A search that was blocked and then only got an empty page returns 503 upstream_blocked (not charged) instead of 200 with empty results.
  • engine: "yandex" on /v1/search returns Yandex web results (up to 15 a page; country picks the region), and engine: "auto" falls back to Yandex last. It can't apply time or safe: "active": a request naming it gets 400, not charged.
  • From 23 November 2026, batch depth requests cost 0.5 credits per page that returned results, without the 7-credit cap: a batch num: 100 goes from 3.5 to 5 credits. Batches queued before that date keep today's price; realtime prices are unchanged.
v1.8.0

Consensus with domain filters and more fallback engines

Consensus accepts domain filters and replaces a failing source, Bing and DuckDuckGo answer images, videos and news, map searches read one level deeper, and Google images and videos return the right rows.

Search API (api.serpkite.com/v1)

  • engine: "consensus" accepts include_domains and exclude_domains: every provider receives the operators, the filters apply to the merged rows, and a provider whose rows were all filtered out is not billed.
  • Consensus replaces a provider that fails or comes back empty with the next index in the route, with at most three answering providers.
  • Bing and DuckDuckGo answer /v1/images and /v1/videos; Bing, DuckDuckGo and Yahoo answer /v1/news; DuckDuckGo and Yahoo join web search after Bing. Named engines or engine: "auto" fallbacks.
  • /v1/map with search reads up to 8 of the start page's links one level deeper when the sitemap is thin. Still 1 credit.
  • Fixed: Google images paired image_url with the previous row's page; about a third of Google video searches returned empty results; bot-check pages served with HTTP 200 were returned and billed as the page by extract, webpage and crawl (now upstream_blocked, free).
v1.6.0

Place details and structured opening hours

/v1/place returns one place's full Google Maps listing by ID, every place gains structured hours, country and time zone, and Maps results come from Google Maps search data in every language.

Search API (api.serpkite.com/v1)

  • POST /v1/place looks up one place by place_id, cid or fid: address, country, coordinates, time zone, categories, hours, phones, website, rating, description and up to 10 photos, in language. 1 credit; an identifier with no place returns empty results and isn't billed. Batches accept endpoint: "place", and the MCP server has a place tool.
  • New place fields on /v1/maps, /v1/places and /v1/place: weekly_hours (local {open, close} intervals per weekday, next_day past midnight), international_phone_number, country, timezone, description, google_maps_url, and photos on /v1/place.
  • /v1/maps and /v1/places read Google Maps search data: addresses, categories, coordinates, websites, phones and hours are filled in every language, Russian included, and both return the same results. ll searches the map at that point and zoom (14z by default) wherever the request comes from; location searches the named place.

SDKs and dashboard

  • sk.place({ place_id }) (TypeScript), sk.place(place_id=...) (Python) and c.Place(ctx, serpkite.PlaceParams{PlaceID: ...}) (Go).
  • The playgrounds can call /v1/place with a place ID.
v1.5.1

Stricter filters, safer crawls and monitors

Domain filters are checked harder and enforced on cached results too; monitors keep only the fields they rerun; crawls survive deploys.

Search API (api.serpkite.com/v1)

  • Domain filters reject entries that could never match (ports, site: operators, odd characters; internationalised domains are converted to punycode), drop duplicates, apply to top_stories and the answer box too, and are enforced on cache hits. GET /v1/search takes repeated keys (exclude_domains=a.com&exclude_domains=b.com). The request echo shows start_date / end_date instead of the derived tbs.
  • /v1/rank rejects include_domains, exclude_domains, boost_domains and highlights (they would change the reported position).
  • /v1/webpage with format: compact or markdown now includes links and image_links when asked for.
  • /v1/extract always has a deadline: timeout defaults to 50 seconds (at most 90), pages still loading stop and are free, and html is capped at 2 MiB per page. The SDKs no longer retry an extract on a 5xx or a timeout.
  • Monitors reject fields they don't keep (400 instead of silently dropping them) and are billed to the key that last created, changed or ran them. POST /v1/crawl answers 503 where async workers are off; DELETE /v1/crawl/{id} answers 404 for an unknown crawl. Crawls cap their results at 16 MB and go back to the queue when a deploy interrupts them.
v1.5.0

Async crawls and monitors

POST /v1/crawl reads a whole site in the background, and monitors rerun a search or check a page on a schedule and report only what is new.

Search API (api.serpkite.com/v1)

  • POST /v1/crawl starts an async crawl: up to 1,000 pages and 10 link hops from a start URL and the site's sitemaps, filtered by path regexes, best first with query, honouring robots.txt and Crawl-delay. Poll GET /v1/crawl/{id} (kept 24 hours), cancel with DELETE, or receive the signed crawl.completed webhook. 1 credit per page read (0.5 from cache); limit credits are reserved and the rest refunded.
  • Monitors (/v1/monitors): saved /v1/search or /v1/news requests that run hourly to weekly and report only results they haven't seen, or /v1/webpage URLs reported when their content changes. New results go to the run history (GET /v1/monitors/{id}/runs) and, when set, a signed monitor.results webhook. Each run costs what its request costs; empty runs are free.
  • The remote MCP server has crawl and crawl_result tools. The SDKs cover crawls and monitors.

Dashboard (app.serpkite.com)

  • New Crawls and Monitors pages: crawl status and credits, monitor runs, pause, resume and delete.
v1.4.0

Extract up to 20 URLs per call

POST /v1/extract reads up to 20 URLs (HTML or PDF) in one call, with query-ranked highlights, and bills only the pages that come back.

Search API (api.serpkite.com/v1)

  • POST /v1/extract: 1–20 URLs as markdown, text or html, fetched in parallel. query + highlights (1–10) add the most relevant passages per page (BM25); include_links, include_images, max_tokens, max_age, country and timeout (up to 90 s, default 50) work as on /v1/webpage.
  • 1 credit per URL that comes back, 0.5 from cache. URLs that fail are listed in failed with a code (invalid_request, upstream_error, upstream_timeout, empty_page) and cost nothing; the response is 200 either way.
  • The remote MCP server has an extract tool. The SDKs have extract (TypeScript, Python) and Extract (Go).
v1.3.0

Map a site

POST /v1/map lists a site's URLs from its sitemaps and start page, optionally ranked by a search phrase.

Search API (api.serpkite.com/v1)

  • POST /v1/map: a site's URLs from the sitemaps in robots.txt (XML, gzip, RSS/Atom, plain text, indexes followed) and the start page's links, cleaned, kept on the host and de-duplicated. include_paths / exclude_paths (regular expressions), include_subdomains, ignore_query_parameters and sitemap (include, only, skip) narrow it; search ranks and filters by relevance. 1 credit, free when nothing is found.
  • The remote MCP server has a map tool. The SDKs have map (TypeScript, Python) and Map (Go).
v1.2.0

Search filters: domains, date ranges, highlights

Restrict, exclude or boost domains, bound results by publish date, read typed published_at dates and get query-ranked highlights instead of whole pages. None of it costs extra.

Search API (api.serpkite.com/v1)

  • include_domains and exclude_domains on /v1/search, /v1/news, /v1/images and /v1/videos: up to 20 domains, path prefixes (github.com/org) or TLDs (.gov), compiled into site: operators and enforced on the results. Not with engine: "consensus".
  • boost_domains on /v1/search and /v1/news moves matching results to the top without dropping the rest.
  • start_date and end_date (YYYY-MM-DD) on the same four endpoints set a custom date range. They can't be combined with time or tbs; fallback engines map the range or are skipped.
  • published_at (ISO 8601) next to date on organic results, news, top stories and videos.
  • highlights: true with include_content returns up to 3 query-ranked passages per page in results[].highlights instead of content.
  • Other endpoints reject these parameters with 400 invalid_request. The SDKs and the MCP tools take them.
v1.1.0

PDFs, links and images on /v1/webpage

/v1/webpage reads PDFs, can return a page's links and images, and accepts pages up to 8 MiB.

Search API (api.serpkite.com/v1)

  • /v1/webpage extracts the text layer of PDFs (application/pdf, or sniffed from the file), with <!-- page N --> markers in the Markdown and the page count in metadata.pages. metadata.content_type names the media type.
  • include_links: true adds links (absolute outbound links with their anchor text, deduplicated, up to 500); include_images: true adds image_links (up to 100). Neither costs extra.
  • Pages and PDFs are read up to 8 MiB (was 2 MiB). A PDF that can't be read is a 503 upstream_error and is not billed.
  • The SDKs' webpage methods take include_links and include_images and type the new fields.
v1.0.5

Exit country for webpages

/v1/webpage takes country to fetch geo-dependent pages through an exit in that country.

Search API (api.serpkite.com/v1)

  • country on /v1/webpage picks the country of the exit that fetches the page (default us). It was a 400 since 1.0.4.
  • The SDKs' webpage methods and the MCP webpage tool take country.
v1.0.4

Stricter, more accurate geo and time filters

Filters are applied or rejected, never silently dropped. ll now works, long location names encode correctly, and fallback engines honour the time window or step aside.

Search API (api.serpkite.com/v1)

  • ll (@lat,lng,zoomz) now searches around that point on /v1/maps and /v1/places. It was accepted but had no effect; on other endpoints it is now a 400.
  • location names of 64 characters or more are encoded correctly for Google.
  • Without country, a location that ends in a country name ("Munich, Bavaria, Germany") sets it, so the search runs from that country.
  • country must be an ISO 3166-1 alpha-2 code (uk is accepted as gb). language accepts regions and _ (pt_BR → pt-br, es-419).
  • A parameter an endpoint would ignore is a 400: for example time on /v1/scholar, country on /v1/patents, or sort on /v1/search. Invalid booleans in query strings are a 400 too.
  • Fallback engines map time and raw tbs windows (qdr:w2, cdr date ranges) where they can. One without the window is skipped instead of answering unfiltered; with an explicit engine that is a 400.
  • Brave, Bing, DuckDuckGo, Mojeek, Wikipedia and Yandex now map gb/uk, Hebrew (iw), Traditional Chinese and Norwegian correctly; Bing picks its market from country.
  • /customsearch/v1 honours cr=countryXX.
v1.0.3

Lens endpoint removed

Google refuses logged-out Lens requests, so /v1/lens could not return results. The endpoint is gone; every other endpoint and every credit price is unchanged.

Search API (api.serpkite.com/v1)

  • POST /v1/lens is removed. Existing calls now get 404.
  • POST /v1/batches with endpoint: lens is rejected with 400 invalid_request.
  • The SDKs no longer have a lens method.
v1.0.2

AI Overview and AI Mode removed

SerpKite no longer parses Google's AI Overview or offers the AI Mode endpoint. Web search returns organic results and the other SERP features as before.

Search API (api.serpkite.com/v1)

  • /v1/search drops the ai_overview request parameter and the ai_overview response field.
  • POST /v1/ai-mode is removed.
  • The remote MCP server (/v1/mcp) no longer lists the ai_overview tool.
v1.0.1

Upstream errors are 503

Upstream failures now return 503 with the usual JSON error body and a Retry-After header. The API never returns 502 or 504, because proxies such as Cloudflare replace those bodies and clients could not read the error.

Search API (api.serpkite.com/v1)

  • upstream_error, upstream_blocked and upstream_timeout are 503 (were 502, 502 and 504). The code values are unchanged, the calls are still not billed and safe to retry.
  • Upstream 503s carry Retry-After (a few seconds). The official SDKs already retry 5xx and honour it.
  • Dashboard API: a payment provider outage or an invoice that isn't ready yet is 503 upstream_error with Retry-After (was 502).
v1.0.0

The v1 API and official SDKs

SerpKite's own versioned API: every route under /v1, Bearer auth, one snake_case response envelope for every vertical, a dedicated batches endpoint and official SDKs. The OpenAPI contract is v1.0.0; breaking changes from here on get a new version and a migration note.

Search API (api.serpkite.com/v1)

  • All routes live under /v1: /v1/search, /v1/images, /v1/videos, /v1/news, /v1/maps, /v1/places, /v1/reviews, /v1/shopping, /v1/scholar, /v1/patents, /v1/autocomplete, /v1/lens, /v1/ai-mode, /v1/webpage, /v1/rank, /v1/account, /v1/status. GET /customsearch/v1 is unchanged.
  • Authentication is Authorization: Bearer skt_live_… (GET requests may use ?api_key=). The legacy key header is no longer accepted.
  • One response envelope for every vertical: request (the normalised request), results (the vertical's primary list), vertical extras and meta (request_id, credits_used, cached, latency_ms, parse_quality…).
  • All keys are snake_case: ai_overview, knowledge_graph, people_also_ask, related_searches, displayed_link, image_url, rating_count…
  • Own parameter names: country and language for geo and interface language; reviews take place_id (or cid/fid), sort and page_token (from next_page_token).
  • Strict validation: unknown parameters return 400 invalid_request with a message that names the replacement, for example unknown parameter "type": use the endpoint path.
  • POST /v1/batches queues 1 to 100 requests for one endpoint at half price and returns one job per request. Poll GET /v1/batches/{id} or receive a signed batch.completed webhook. It replaces the batch mode flag and JSON array bodies.
  • POST /v1/rank: the position of a domain for a keyword in the top 10 to 100.
  • Remote MCP server moved to https://api.serpkite.com/v1/mcp (Bearer auth).

Billing

  • The $1,000 and $3,000 credit packs are no longer sold. Self-serve packs run from $10 to $300; larger volumes are quoted on request. Credits already bought never expire, and accounts keep the rate limit of the largest pack they bought.

Official SDKs

  • TypeScript: npm install serpkite (zero dependencies; Node 18+, Bun, Deno, edge).
  • Python: pip install serpkite, with SerpKite and AsyncSerpKite.
  • Go: go get github.com/serpkite/serpkite-go.
  • LangChain: pip install langchain-serpkite (search tools, retriever, webpage loader).
  • CrewAI: pip install "serpkite[crewai]" (SerpKiteSearchTool).
v0.9.0

Preview release

The pre-v1 preview of the SerpKite API, dashboard and billing. Its unversioned routes are superseded by the v1 API.

Search API (api.serpkite.com)

  • JSON search endpoints for Google Search, Images, Videos, News, Maps, Places, Reviews, Shopping, Scholar, Patents, Autocomplete, Lens, AI Mode and a webpage-to-Markdown fetcher.
  • GET /customsearch/v1: a Google Custom Search JSON API–compatible endpoint (key, cx, q, start, num, siteSearch, dateRestrict, searchType=image…).
  • AI Overview in web search at no extra credit.
  • Output formats: format=json|compact|markdown, plus fields= projection.
  • include_content=0–5 fetches the top result pages as Markdown (+1 credit per page).
  • Resolved destination URLs and a canonical domain on every organic result; meta.resolved_urls and meta.parse_quality in every response.
  • num=100 depth bundle (7 credits), max_age cache acceptance (0.5× on a hit).
  • Batch lane at 0.5× credits with polling and signed webhook delivery.
  • Account endpoint: balance, rate limit, plan and this month's usage.
  • Response headers: X-Credits-Used, X-Credits-Remaining, X-Cost-USD, X-Cache, X-Latency-Ms, X-Tokens-Estimate, X-Request-Id.
  • Remote MCP server (streamable HTTP) with search, news, maps, scholar, webpage and AI Overview tools.
  • No-signup playground endpoint, rate limited per IP.

Billing

  • Credit packs from $10 (10,000 credits) to $3,000 (12.5M credits). Credits never expire.
  • Failed, empty and blocked requests are refunded automatically.
  • Free tier: 1,000 credits on signup, 1,500 more after linking GitHub or Google, 1,000 every month.
  • Payments through Polar as merchant of record, with invoices.

Dashboard (app.serpkite.com)

  • Sign in with a magic link (link or 6-digit code), GitHub or Google.
  • Multiple API keys with names, rotation and per-key monthly credit limits. Only a hash of each key is stored.
  • Usage by day, endpoint and key; request log (metadata only); CSV export.
  • Monthly spend cap, low-balance and usage alerts, auto-recharge links, webhook secret rotation.
  • Session management: see and revoke active sessions.

Start building

Get your API key in 30 seconds

2,500 free credits, then 1,000 every month. No credit card.