Security · Primetra

What we collect

The collector runs on your stack and forwards requests from bots and agents. It also forwards two kinds of traffic that can be people: visits that arrive from an AI assistant's answer (AI referrals), and calls to your API and MCP paths from anything that isn't a browser, such as your own scripts. For those we send a truncated IP prefix hashed under a secret only you hold; without that secret, no IP prefix or hash is sent. The edge collectors also send the country (and, on Cloudflare, the network) your platform looks up, with or without the secret. We never see a visitor's address or can derive one from the hash.

Ordinary visitors browsing your site, query strings, cookies, auth headers and bodies stay on your stack. The only full IPs we receive belong to crawlers claiming to be a known AI bot, so we can check them against the vendor's published ranges; they are never stored. The install guide lists every field the collector sends. On our side we add a label to each event: which agent or bot it is, and whether we could verify it.

Who can read it

Row-level security in the database: a member of a workspace can only read that workspace's rows.
Our operators, a short list of Primetra staff, can read workspaces to support you. View-as-client sessions, operator changes and deletions are recorded in an audit log; a view-as-client session is logged before any data loads.
Ingest keys are hashed for authentication. A new key is held encrypted until you reveal it once (up to 7 days).

What AI providers see

We send buyer questions about your category to the official APIs of Anthropic, OpenAI, Google, xAI and Perplexity, and, while we run that separate source, to Meta's Model API. To set up your profile, Anthropic also reads the text of your public homepage, pricing and docs pages. Each week Anthropic also writes your memo's headline from that week's aggregated metrics and findings. No AI provider ever gets the raw agent events your collector sends, IP addresses, or anything about the visitors to your site.

Subprocessors

Supabase: Our database and sign-in. Stores workspaces, members' email addresses, the agent events your collector sends, free-check requests (the domain, your email and a hashed IP prefix) and the reports we build.
Cloudflare: Runs the endpoint your collector sends agent events to. Its queue briefly holds each event until we've classified it, including the full IP of a request that claims to be a known vendor bot, until we've checked it. Also runs Turnstile, the bot check on the free-check form, which sees the visitor's browser and IP address.
Vercel: Hosts this site and the app. Every page request passes through it.
Inngest: Runs our background jobs: checks, weekly reports and emails. Job messages carry record ids (and the recipient's address for welcome and invite emails), plus the results of each job step.
Stripe: Payments: Checkout and the billing portal. Stripe collects card and billing details itself; we keep the order, the Stripe customer and subscription ids, and which version of our terms you accepted.
Resend: Sends our emails, sign-in links included. It receives the recipient's address and the message, and tells us about bounces and complaints.
Google Fonts: Our emails load their fonts from Google. When your mail app shows them, Google sees your IP address, unless your mail app blocks remote content.
Sentry: Error reports from the app and our jobs. Each report is scrubbed before it leaves: emails, IP addresses, tokens, keys and query strings are removed, and no request or user details are sent. Reports from a visitor's browser go straight to Sentry, so Sentry sees that browser's IP address.
Anthropic: The buyer questions we ask about your category. When we set up a product profile (a free check or a new workspace), also the text of your public homepage, pricing and docs pages, to propose your category, competitors and questions. Each week, the facts behind your memo's headline: your product and competitor names, the week's aggregated metrics (answer shares, agent request counts, blocked, error and empty rates, the agents that connected, the tools they asked for) and the findings and fixes built from them, which can name pages on your site. It also drafts entries for our internal playbook from public vendor pages. Never the raw events, IP addresses or anything about the visitors to your site.
OpenAI, Google, xAI, Perplexity: The buyer questions we ask about your category, through their official APIs. Nothing from your collector.
Meta (Meta Model API): Only while we run the Meta Model API source, which is shown on its own labelled line and is not counted in Share of Agent: the buyer questions we ask about your category, through Meta's official developer API, in Standard mode (Meta's terms say it does not train its models on Standard-mode requests). Nothing from your collector.
GitHub: Only if you connect a repository: the draft pull requests we open there, with the fix and its evidence.

How long we keep it

Raw agent events: 45 days on Monitoring, 60 on Growth, 90 on done-with-you. Daily rollups: 25 months.

Free checks (the domain, your email, a hashed IP prefix and the result): 12 months.
Email delivery records (who we emailed, which email, whether it arrived): 12 months.
Sign-up and purchase step counts: 13 months. Our job logs: 13 months.
The audit log (view-as-client sessions and operator changes, deletions): 24 months.
Orders and payment records: kept as the law requires. The unsubscribe list: kept, so an unsubscribe keeps working.

Export your metrics any time from the app, as JSON or CSV. Raw events: ask us.
A workspace owner can request deletion any time in Settings. It's final after a 7-day grace period, and you can cancel it until then.
When you request deletion we stop your subscription's renewal right away; if Stripe is unavailable, we retry daily. If you cancel the request, renewal is switched back on.
At the end of the grace period we delete the workspace's data, cancel the subscription, and remove sign-ins that belong to no other workspace. Refunds: [To be completed with counsel].
We keep order records as the law requires, our own cost records, and an audit entry saying the workspace was deleted. Copies held by our providers and in database backups expire on their own schedules.

Report a vulnerability