Privacy and data retention
SerpKite never logs query text; failed requests keep theirs for a day to debug. Results are cached briefly under a hashed key; usage records are metadata only.
The short version
- Query text is never logged. Not in application logs, not in usage records, not in the request log you see in the dashboard.
- Failed requests are the one exception. A request that fails (HTTP 5xx) or returns no results keeps its query text, options and the steps we tried for 24 hours, so we can reproduce and fix the failure. Successful requests’ query text is never stored.
- Usage records are metadata only (endpoint, status, credits, latency, cache hit, which key) and are deleted after 31 days.
- Results are cached briefly (up to 6 hours by default, 30 minutes for news, 1 hour for web pages and AI assistant answers, and 24 hours for autocomplete) under a hash of the request, batch results are kept for 24 hours and crawl results for 24 hours so you can fetch them. Monitors keep the search you saved until you delete them. Apart from the failed-request records above, nothing else is kept.
- API keys, session tokens and sign-in codes are stored as hashes, never in plain text.
- AI features send text to a model provider (Anthropic or OpenAI) only for the request that uses them, under API terms that exclude training. SerpKite never logs prompts or page text. See AI features.
This page describes how the service is built. The binding terms are in the Privacy Policy and the DPA.
What we store, and for how long
| Data | Contains query text? | Retention |
|---|---|---|
| Usage events (per request) | No. Endpoint, HTTP status, credits, latency, cache hit, batch flag, error code, API key ID, time | 31 days, then deleted |
| Failed-request debug record | Yes, for a request that failed (5xx) or returned no results only: query text or target URL, options, status, the route we tried, account and key ID | In memory, deleted automatically after 24 hours |
| Request log and CSV export in the dashboard | No. Built from usage events | 31 days |
| Sign-up checks | No. Keyed hashes (HMAC) of the network you sign in from and of the random sk_did device cookie, never the IP address itself; used to refuse free credits to repeat sign-ups |
Network 30 days, device 90 days after the last sign-in that used them |
| Daily usage totals and the credit ledger | No. Credits per day, purchases, grants, refunds | Life of the account (needed for billing) |
Result cache (read with max_age) |
The key is a SHA-256 hash of the normalised request, never the query text. The value is the result, with no account or key attached. Every successful live result is written, whether or not you send max_age |
In memory only, expires automatically: up to 6 hours by default, 30 minutes for news, 1 hour for web pages and AI assistant answers, and 24 hours for autocomplete |
| Batch job request | Yes, until the job runs | Cleared as soon as the job is processed |
| Batch job result | Yes (it is the SERP) | 24 hours, then deleted |
Crawl request (/v1/crawl start URL and options) |
Yes, until the crawl finishes | Cleared when the crawl completes, fails or is canceled |
Crawl result (pages read) and its webhook_url |
Yes | 24 hours after completion, then deleted (at once when you delete your account) |
Research request (/v1/research input and options) |
Yes, until the task finishes | Cleared when the task completes, fails or is canceled |
Research result (report, sources, queries) and its webhook_url |
Yes | 24 hours after completion, then deleted |
Prompt sent to an AI assistant (/v1/chatgpt, /v1/gemini, /v1/copilot, /v1/ai_mode, /v1/brave_ai) |
Yes: the prompt, which we handle like a search query (never logged; its answer cached for at most 1 hour under a hashed key) | Not stored by SerpKite. The assistant receives it from a logged-out visitor, without your identity or account |
| Text sent to a model provider (AI features) | Yes: the query or research text and the page text involved | Not stored by SerpKite. Providers process it under their API terms (see AI features) |
Monitor definition (/v1/monitors) |
Yes: the saved query or watched URL, filters, schedule, webhook URL and your metadata | Until you delete the monitor or your account |
Monitor run history (/v1/monitors/{id}/runs) |
What a run found new (search results, or the watched page’s content for page monitors); the run’s status, counts and credits | Results 24 hours, the run record 30 days (or until you delete the monitor) |
| Monitor “already seen” list | No. Truncated SHA-256 hashes of result links (or of the page content), at most 1,000 per monitor | Until you delete the monitor |
| Webhook deliveries | Sent to your URL; we keep only the delivery status | With the job |
| Application and database logs | No. Never API keys, cookies, q, webhook URLs or database query parameters |
Short operational retention |
| Account data | Email, name, avatar, linked OAuth providers, team membership, settings | Until you delete the account |
| MCP app connections (OAuth) | No. The app’s self-declared name, its redirect address and the key it runs as; access and refresh tokens only as SHA-256 hashes | Tokens expire after 1 hour (access) and 60 days (refresh) and are deleted a day later; the connection record until you delete the account |
| Payments | Handled by Paddle, our merchant of record. We store order IDs, pack, amount and credits, not card data | Until you delete the account; Paddle keeps invoices as tax law requires |
| Database backups | Account, billing and usage tables (encrypted) | 30 days |
Query text is never written to a log or a usage record; failed and empty requests keep it in their debug record for 24 hours. The only copies of results are the short-lived ones listed above: cache entries, which deduplicate identical requests and are served only to callers that send max_age, and batch results you haven’t fetched yet.
AI features
Some options use language models: include_answer, /v1/answer, /v1/research, and summary or output_schema on /v1/extract. For those requests, SerpKite sends the query or research text and the text of the result pages involved, never your account details, email or API key to Anthropic, PBC and OpenAI, L.L.C., under the providers' commercial API terms, which do not allow training on API data; the providers may keep it for a limited period for abuse and safety monitoring.
| Provider | What it does | When |
|---|---|---|
| Anthropic (Claude) | Writes answers, page summaries, JSON extractions and research reports | Only when a request asks for one of them |
| OpenAI (Decisions API) | Typed checks: which sources help an answer, whether research has covered the task, boolean or enum extraction schemas, whether a short fetched page is a bot check, login wall or soft 404, whether a monitored page changed in substance | With the AI features above, and on short pages fetched by /v1/webpage, /v1/extract, /v1/crawl and include_content, and on page-monitor changes |
On our side, prompts and page text are never logged by SerpKite; generated answers, summaries and extractions are returned to you and not cached, and research reports are kept like other task results. meta.llm tells you which provider and model handled a response. The decision checks on fetched pages run without any option: they see only short pages (where bot checks and soft 404s hide), only the page you asked for, and nothing about your account. They keep you from being billed for pages that aren’t real content.
AI assistants
The AI assistant endpoints send your prompt to the assistant you call (ChatGPT, Gemini, Copilot Search, Google AI Mode or Brave’s AI answer) the way a logged-out visitor would, through our proxies, exactly as a search query goes to Google. Nothing about your account or key goes with it, and we never sign in to an assistant. On our side the prompt follows the query-text rules above: it is never logged, an answer is cached for at most 1 hour under a hash of the request, and a failed or empty request keeps the prompt in its debug record for 24 hours. Don’t put secrets or personal data in prompts: the assistant’s operator sees them like any visitor’s question.
What we don’t do
- We don’t log the
qparameter, API keys or cookies, and we don’t store the query text of successful requests. - We don’t keep results after the retention windows above, and we don’t keep a searchable archive of queries.
- We don’t scrape anything behind a login. SerpKite fetches the public, logged-out Google results page through proxy networks. See How we collect data.
Choosing the privacy trade-off
- Don’t want a cached copy served to you? Don’t send
max_age(or send0). Every call is then a live fetch andX-CacheisMISS. The fresh result is still written to the cache for the windows above;max_ageonly controls reads. - Don’t want batch results held? Use realtime calls. Batch results always expire after 24 hours, whether you use a webhook or fetch them.
- Sensitive queries? Realtime calls leave metadata in usage records (which endpoint was called, when, by which key, and what it cost) and a cache entry under a one-way hash that expires within hours. A call that fails or returns no results also keeps its query text for 24 hours, to debug it. Raw HTML (
include_html) is never cached.
Secrets
- API keys are stored as SHA-256 hashes. The full key is shown once when created or rotated. See API keys.
- Dashboard sessions and sign-in links and codes are stored as hashes too. You can see and revoke your sessions from the dashboard.
- Webhook signing secrets are shown once; the dashboard shows only their prefix afterwards.
Product analytics
The website and dashboard use privacy-friendly product analytics (PostHog, EU region) to understand which features are used, only after you accept analytics cookies in the cookie banner. Global Privacy Control and Do Not Track count as a refusal. You can change your choice any time under “Cookie settings” in the website footer or Settings → Cookies in the dashboard (see the Cookie Policy). API requests themselves are not sent to analytics. Query strings are stripped from every URL before an event is sent, so playground queries never reach analytics.
Deleting your data
The account owner can delete the account from the dashboard settings. This removes the account, its API keys, remaining credits, ledger, usage events, settings and team, and signs out every session. Copies in encrypted backups roll off within 30 days. Invoices are kept by Paddle, our merchant of record, as tax law requires.
For data requests or questions, email privacy@serpkite.com.
Compliance
- Data processing terms: DPA.
- Who processes data for us: Subprocessors.
- Security practices and how to report a vulnerability: Security.
- SOC 2: planned, not yet certified.
Related
Last updated: 2026-10-28