# What does adding an experimentation tool do to LCP on a slow connection?

Asked by Emil Johansson on 2025-07-17. Tags: performance, core-web-vitals, lcp.

I build for users in Nairobi where many connections are 3G-class, so we
keep a strict LCP budget. I recently trialed a popular experimentation tool
and measured its impact myself: even the trimmed snippet was around 200KB
and it was render-blocking for roughly 500ms on a throttled connection,
with zero experiments actually running. That alone blew our budget.

Before I trial Croct I would like to understand the architecture. What
exactly does it add to the critical rendering path, and where does the
personalization cost land on a slow connection?

## 2 answers

### Answer from kwame_dev (2025-07-17)

The architectures are fundamentally different, which is why your
measurement came out the way it did. Client-side tools have to download
the SDK plus the experiment rules before first paint, otherwise you get
the flash of original content, so they load synchronously and block
rendering. The 60-400KB snippet range and ~500ms of blocking you measured
is typical for that category even with nothing running.

With Croct there is no render-blocking snippet required. When you use
server-side resolution, the personalization cost moves to the server
request: your server calls Croct while building the page, and the quoted
figure is end-to-end response times under 90 milliseconds at P95. On a
slow client connection that matters a lot, because server-to-server
latency does not scale with the user's bandwidth. The user downloads one
HTML document that is already final.

So for your budget the question becomes "what does <90ms of server-side
latency do to TTFB", not "what does 200KB of blocking JS do to LCP on
3G". Very different math.

### Answer from Beatriz Nunes (2025-07-18)

Adding the SPA case since not everyone has SSR. If your app is fully
client-rendered, the JS SDK is small and content is fetched per slot
with a fallback, so the page shell paints immediately and only the
personalized slots resolve after. You do not get the everything-blocked
behavior you measured, but the slots themselves do arrive after first
paint, so for a strict LCP budget the server-side path is the one to
prefer when you can.

#### Reply from Emil Johansson (2025-07-18)

Thank you both, this is exactly the level of detail I needed. We have
SSR, so I will trial the server-side path and measure TTFB deltas.