Does session replay slow down my website? 10 recorder scripts weighed for founders who build with a coding agent
A little, and how much depends on the recorder. On November 15, 2026 the scripts we downloaded weighed from 8.5 KB compressed (UserProof) to 194.9 KB (PostHog with its recorder). On one test page at a 4x slower CPU, rrweb's full-page recorder added 46 ms of script time and UserProof added 3 ms. Load any recorder with defer.
This page is for founders who build their site with a coding agent and want to know what a recording script costs before they paste it in. We downloaded each tool's script on November 15, 2026 and measured it. Then we ran a lab test on a copy of one of our own guide pages: no recorder, UserProof's snippet, and rrweb, the open-source recorder PostHog's replay is built on. Seven runs each, medians shown. The method and its limits are at the end of each section.
On this page: script weight · main-thread time · vendors' own numbers · why a recorder costs time · how to keep it fast · what UserProof's snippet does · FAQ
How big is each session recorder's script?
Between 8.5 KB and 194.9 KB compressed, a 23x spread. We fetched the files each tool's install snippet loads, then compressed each one ourselves with gzip at level 6 so every row is measured the same way. Tools that fetch more files after they start are listed with every file we could find; where a file needs a real project ID we say so.
| Tool | Files measured | Uncompressed | Compressed (gzip -6) | Note |
|---|---|---|---|---|
| UserProof Ours | ingest.getuserproof.com/sdk.js | 26.9 KB | 8.5 KB | One file. Served at 8.7 KB with gzip |
| LogRocket | cdn.lr-ingest.com/LogRocket.min.js | 51.4 KB | 15.1 KB | First file only; anything it loads after start was not followed |
| Smartlook | recorder.js loader plus the es6/init file it loads | 78.4 KB | 22.6 KB | Loader alone is 1.9 KB compressed |
| Microsoft Clarity | scripts.clarity.ms/0.8.77/clarity.js | 73.9 KB | 26.6 KB | 0.8.77 is the version in Clarity's public repo. The small per-project tag file needs a project ID and was not measured |
| Sentry (replay part only) | CDN bundle 11.12.0 with replay, minus the same bundle without it | 122.6 KB | 38.4 KB | Sentry's docs say "~36KB gzipped". The full bundle with tracing and replay is 91.9 KB compressed |
| OpenReplay | static.openreplay.com/latest/openreplay.js | 194.8 KB | 58.1 KB | The CDN build of its tracker |
| Hotjar | hotjar-<id>.js loader plus the modules file it names | 257.4 KB | 77.9 KB | Loader fetched with a placeholder site ID; it named one modules file, which we fetched |
| rrweb 2.2.2 (open source) | The full UMD bundle from npm | 282.1 KB | 88.0 KB | Record and replay code in one file; the lab test below used this file |
| FullStory | edge.fullstory.com/s/fs.js | 316.3 KB | 99.7 KB | Served already gzipped at 99.5 KB |
| PostHog | array.js plus posthog-recorder.js | 612.3 KB | 194.9 KB | array.js (122.4 KB compressed) also carries analytics, flags and surveys. The recorder file (72.6 KB) loads only when replay is on |
Measured November 15, 2026 with curl from a Mac mini in the US; posthog-js 1.449.2 and @sentry/browser 11.12.0 were the latest versions on npm that day. KB means 1,000 bytes. Mouseflow is left out because its script URL needs a real project ID.
Bytes are only half of it. A compressed file downloads fast on most connections; the browser still has to unpack, parse and run every byte on the visitor's CPU, and that is where a phone feels it.
How much main-thread time does recording add?
On our test page, rrweb's recorder added 46 ms of script time and UserProof's snippet added 3 ms. On a page eight times bigger, those grew to 71 ms and 4 ms. The gap comes from what each one does at start: rrweb copies the whole page so it can be played back later, and UserProof attaches event listeners and a change observer that does not store the changes.
| Test page | Recorder | Script time | All main-thread tasks | Added vs no recorder |
|---|---|---|---|---|
| 387 elements (one guide) | None | 0 ms | 64.8 ms | baseline |
| UserProof sdk.js | 3.1 ms | 70.8 ms | +6.0 ms | |
| rrweb 2.2.2, recording | 46.2 ms | 110.9 ms | +46.1 ms | |
| 2,956 elements (same guide x8) | None | 0.1 ms | 133.2 ms | baseline |
| UserProof sdk.js | 4.2 ms | 141.7 ms | +8.5 ms | |
| rrweb 2.2.2, recording | 70.5 ms | 205.4 ms | +72.2 ms |
Lab test, November 15, 2026: headless Chrome 157 driven by Playwright on a Mac mini, CPU slowed 4x (Chrome's own throttle, a stand-in for a mid-range phone), cache off, files served from the same machine so the network adds nothing. Seven runs per row, medians. Script time and task time are Chrome's ScriptDuration and TaskDuration counters, read 3 seconds after the load event. Long tasks (50 ms or more): the small page had none in the median run with any recorder. The big page had one in the median run with or without a recorder, and its longest was 53 ms with no recorder, 63 ms with UserProof and 57 ms with rrweb, so a 10 ms difference on that page may be noise or may be the recorder.
What the test does not cover: the time a recorder spends on every click and scroll later in the visit, which feeds Interaction to Next Paint. It also leaves out the upload, and pages that change a lot while you watch, like dashboards and animations. We also tried an earlier Clarity build (0.8.43), but without a project tag it does not start recording, so it ran for under 1 ms and would have looked better than it is; that row is left out. The rrweb row is the bare open-source recorder with default settings, not PostHog's own build of it, so it is not PostHog's number.
What do the vendors' own numbers say?
Sentry is the only vendor we found that publishes a before-and-after benchmark. On its own app's busiest data table, Total Blocking Time went from 2,621.67 ms with no Sentry to 3,036.80 ms with Sentry and its replay add-on, about 16% more, while Largest Contentful Paint did not get worse (1,599.19 ms to 1,529.11 ms, inside the spread). Upload during the test went from 21 B to 272.51 KB. Sentry ran it 50 times on an Apple M1 MacBook Pro and last updated the numbers on September 25, 2023.
- Sentry says its replay snapshots the page's DOM once, then sends only the changes it hears from the browser's MutationObserver, and compresses them in a Web Worker so the work happens off the main thread.
- PostHog lists a
lazy_loadingstate in which "the SDK is still fetching the recorder script", and says heavy SVG or Lottie animations "may be throttled or ignored entirely due to the performance overhead in recording and replaying". Its troubleshooting page names the recorder it starts: rrweb. - Microsoft Clarity's FAQ answers "Do heat maps slow down my site?" with "Not at all", because heat maps use data Clarity already collects. For its mobile SDK it says it is "designed to be efficient" and asks you to contact support if you see a significant drop. It publishes no timing numbers.
- Hotjar, FullStory, LogRocket, Mouseflow and OpenReplay: we found no published timing benchmark. Hotjar's help center returned 403 to our fetch.
Why does a session recorder cost time at all?
Four costs, in the order a visitor meets them. First the download, the bytes in the table above. Second, parse and run: the browser compiles the script on the main thread, the same thread that answers taps. Third, the start-up snapshot: a recorder that can play back a video-like replay has to copy the whole page into memory first, which is the 46 to 70 ms rrweb spent in our test and grows with the page. Fourth, the steady cost for the rest of the visit: watching for changes, compressing them and uploading them.
Recorders that only log events (clicks, scrolls, errors) skip the snapshot and most of the steady cost. The trade is what you can watch later: an event recorder tells you where a visitor clicked and whether anything happened, but it cannot show you the page as the visitor saw it.
How do you keep a recorder from slowing your site?
Load it with defer, record fewer sessions, keep heavy parts out, run only one recorder and measure your own page before and after; the six steps are below.
- Load it with
deferorasync. MDN: a deferred script runs after the document has been parsed, so it does not stop the browser from reading your HTML. UserProof's snippet ships withdefer; check that yours does too. - Record fewer sessions. PostHog and Sentry both have a session sample rate (Sentry's default config records 10% of sessions). Fewer recorded visits means less recording work and less upload in total.
- Keep heavy parts out. Mark animated canvases, maps and chat widgets as excluded. That saves CPU and keeps private text out too; our masking snippet generator writes the attribute for 8 tools.
- Run one recorder, not two. Each runs its own listeners and uploads. If you keep Hotjar for heat maps and add another tool, budget for both.
- Load it after consent where the law asks. In the EU you usually need consent first anyway (see is session replay GDPR compliant?), which also means no cost before the visitor agrees.
- Measure your own page, before and after. Run Lighthouse on a phone profile twice and compare Total Blocking Time, then watch Interaction to Next Paint in your field data for a week. web.dev counts an INP of 200 ms or less as good.
Picking a tool and weighing cost as well as speed? The session replay tools compared for coding agents page compares them on price, MCP access and what each records, and the replay cost calculator prices 11 at your traffic. If your site sends a Content Security Policy, the CSP generator prints the hosts each recorder needs; UserProof's are one origin in two directives. The CSP checker shows whether the policy you send already allows them.
What does UserProof's snippet do on a page?
It listens for events and sends them in small batches; it does not copy the page. From the recorder source (public/sdk.js, last changed October 27, 2026):
| What | How |
|---|---|
| Loading | One file, 8.5 KB compressed, installed with <script defer> |
| Clicks, scrolls, errors | Recorded as events with the element and position. Field values are never recorded |
| Page changes | A MutationObserver only marks whether a click changed anything (to find dead clicks); the changes themselves are not stored |
| Mouse movement | Summarised as a count and distance, not stored move by move |
| Sending | Batches every 5 seconds, or at 100 events, with navigator.sendBeacon |
| What you give up | No video-like playback of the page. You get the clicks, the dead and rage clicks, errors and where each visit ended, which is what the coding agent needs to make a fix |
UserProof is free for 3 sites and 5,000 recorded visits a month; Pro is $29 a month for 50,000 and Scale $99 for 250,000 (pricing). How the script tag install works: record real visitor sessions with a script tag.
Sources, read November 15, 2026
- Our measurements: each script fetched with curl and compressed with gzip -6; lab test in headless Chrome 157 with a 4x CPU throttle, 7 runs per row (November 15, 2026)
- Sentry, Replay performance overhead (benchmarks last updated September 25, 2023)
- Sentry, Set up Session Replay (default config
replaysSessionSampleRate: 0.1) - PostHog, Session replay troubleshooting (no date shown)
- Microsoft Clarity, FAQ
- Microsoft Clarity, clarity-js package.json (version 0.8.77)
- rrweb, GitHub repository; version 2.2.2 from npm
- npm latest versions on November 15, 2026: posthog-js 1.449.2 and @sentry/browser 11.12.0
- MDN, the script element (the
deferandasyncattributes) - web.dev, Interaction to Next Paint
- UserProof recorder source (public/sdk.js) and getuserproof.com/pricing
FAQ
Does session replay slow down a website?
A little. In our November 15, 2026 test at a 4x slower CPU, a full-page recorder (rrweb) added 46 ms of script time on a normal page and 71 ms on a page eight times bigger. An event-only recorder (UserProof) added 3 to 4 ms. Loaded with defer, neither stops the browser from reading the HTML.
Does Hotjar slow down my site?
Hotjar's loader plus the modules file it named weighed 77.9 KB compressed on November 15, 2026, about 9 times UserProof's snippet. Hotjar publishes no timing benchmark that we found, and its help center blocked our fetch. Measure your own page with Lighthouse before and after.
Does Microsoft Clarity slow down a website?
Clarity's main script, version 0.8.77, is 26.6 KB compressed. Clarity's FAQ says heat maps do not slow your site because they use data already collected, and publishes no timing numbers. In our lab an earlier build of its script did not start without a project tag, so we did not time it.
Does session replay hurt Core Web Vitals or SEO?
It can move Interaction to Next Paint and Total Blocking Time, not what Google reads in your HTML. Sentry's own benchmark showed Total Blocking Time up about 16% with replay on a heavy table page, with no worse Largest Contentful Paint. Check your INP field data for a week after adding a recorder.
Should a session replay script use async or defer?
Either keeps it off the critical path. defer runs the script after the HTML is parsed, in order; async runs it as soon as it downloads. UserProof's snippet uses defer. Avoid a plain blocking script tag in the head.