Most replay costs are not created by traffic; they are created by a capture policy nobody wrote down. A recorder goes in, every pageview becomes a session, and a month later much of the quota went to crawlers, uptime checks and the same logged-out homepage visit repeated thousands of times. Session replay sampling is the practice of deciding, before a session is stored, which sessions are worth keeping and how much volume each site, team or customer may use.
It has two halves. Sampling picks the sessions: all of them, a fraction, or only those that hit a trigger such as a JS error, a rage click or a failed payment. Quotas divide the volume that remains.
Start from the question replay should answer
A random sample answers aggregate questions well: how often an error appears, which step loses the most people. It answers individual questions badly: why this account stopped working last Tuesday. If your real use is “open the session of the customer who complained”, random sampling deletes exactly the sessions you need. If your use is product research, a broad sample is fine and cheaper. Write down the two or three questions first, then note which need identity (a named user, a named account) and which only need counts.
Gate the recorder with triggers

Record these in full
- JS errors and Web Vitals regressions on key routes.
- Checkout or payment steps that fail, or abandon at the same field.
- Rage clicks, dead clicks and repeated validation errors.
- Accounts you cannot afford to lose: first-day signups, accounts above a revenue threshold, accounts with an open support ticket.
Sample these
Marketing pages, docs, repeated internal admin views, and any route with thousands of near-identical visits a day. They give you layout evidence and rarely need to be complete.
shouldRecord =
hasError ||
(isCheckoutStep && !completed) ||
rageClicks > 2 ||
account.tier === "enterprise" ||
hash(visitorId) % 100 < 5 // 5% of the rest
Start the recorder only when that condition is true. Everything else never spends quota.
Keep bots out of the quota
A crawler that renders JavaScript produces a session that looks like a very fast, very confused user. If it counts against quota you pay twice: in volume, and in noisier heatmaps and funnel counts. User-agent strings are not enough; automated browsers and AI agents increasingly send standard headers. Use a detection layer that reads the whole signal set and drops the session before it is stored — see our bot detection and AI agent detection pages. In Prynt Replay, bot sessions never count against quota: exclusion happens at capture, not in a dashboard after the volume is billed.
Split quota by site, environment and account

Production is not staging
Preview deploys and QA runs teach you nothing about customers and can consume a small plan in days. Prynt Replay counts plans in sites (Free covers one, Pro three), so decide which environment deserves a slot.
Tag every session with the customer
Without identity you cannot search by account or tell an enterprise session from a trial. Pass it once at login with identify(userId, { accountId }); later sessions carry it, and customers can open their own sessions inside your admin, filtered per company or user, in an embedded player. Setup is in the docs.
Reserve a share for the loud paths
A split that holds up in practice: roughly 60% product and checkout, 25% support and account sessions, 15% research. Review after a month; the numbers move.
Sample volume, keep rare sessions whole
Two models. Head-based sampling decides at page load, before you know whether anything interesting happened: cheap and predictable, but it can miss the error you needed. Tail-based decides after the session ends, when you already know it contained a failed payment or an exception: better coverage, more to operate.
A practical mix is head-based for routine navigation and unconditional capture where failures are rare and expensive. If you sample at page load, hash a stable identifier such as visitorId instead of calling Math.random(); the hash keeps one visitor consistently in or out, so monthly numbers stay comparable. Check what “session” means in your recorder too: a 30-minute inactivity timeout splits one long visit into two, and mobile replay on iOS, Android, React Native or Flutter can split differently.
Measure whether sampling is working
Four numbers, once a month:

- Share of recorded sessions that contain a trigger. Mostly routine navigation means the sample is too broad.
- Error coverage: how many error events have a replay attached. Close to full is the goal.
- Sessions opened per hundred recorded. Very low means you record for nobody.
- Quota used per site and environment, so a preview deploy never surprises you at month end.
Py summarizes sessions and sends a daily brief (included with Replay, with a monthly limit per plan), which matters once the sample is smaller than your team’s available hours but larger than what anyone will watch.
FAQ
Does sampling break heatmaps?
Heatmaps need many sessions on the same page to say anything. Sampling marketing pages lightly is safe; sampling a low-traffic checkout page is not — you will read a heatmap built from three visits.
Do bot sessions count against my Prynt Replay quota?
No. Bot sessions are excluded and never count against quota.
How do I show sessions to my own customers?
Call identify(userId, { accountId }) at login, list sessions with the secret key and open them in the embedded player, so each customer sees only their own company or user. The flow is in the docs.
Try Prynt Replay
Start on Free: 1,000 sessions a month, 14-day retention, one site. Pro is $279/month for 10,000 sessions, 30-day retention, three sites and a 14-day trial; volume plans reach 1M sessions. Every form field is masked in the browser and again on the server, with no unmask option, so a sampling decision never becomes a privacy decision. Start at pryntid.com.
Try it free
Prynt is device intelligence with a free plan: 5,000 identifications a month, visitor IDs with confidence and bot detection — no credit card. Upgrade to Pro for VPN, proxy, tampering and behavioral signals.