Rawuser / Questions

Does a personalization script slow down my website?

Short answer

It can.

A script that hides the page, or parts of it, while it decides can hold them for 2 to 4 seconds before it gives up, which are the default timeouts of VWO, Webflow Optimize and the old Google Optimize snippet. In our own test on a shaped phone connection on 26 October 2026, our tag, which hides only the lines it rewrites, moved a small lab page's LCP from 0.60 s to 1.13 to 1.21 s for a ChatGPT visit on a first view, with no layout shift.

Last checked 26 October 2026 · Agshin Rajabov, founder of Rawuser

your-site.com/
The rest of the page
Draws once a 4 KB script is in
The lines Rawuser rewrites
Wait, usually a fraction of a second
A late answer
Your original stays
First view, phone
About 0.3 s to paint, 0.5 s to the headline
What waits while the page loads. A drawing, not a screenshot.

What does a personalization script do while your page loads?

It decides which version this visitor gets, and until it has decided it either holds back what it will change or lets the original show and swaps it a moment later.

These scripts have to load their settings, and often ask a server, before they know which version to show. A script in the <head> without async or defer stops the browser from reading the rest of your HTML until the file has arrived and run. web.dev calls this parser-blocking and advises async or defer for third-party scripts unless the page cannot render without them.

With async, the page paints first and the script changes it afterwards, so the visitor reads the original and then watches it change. Testing tools call that flicker, or a flash of original content.

So vendors either ask for the blocking form or pair async with a hide. Optimizely's docs say not to install its snippet as non-blocking except as a last resort, because loading it asynchronously raises the risk of flicker. VWO's asynchronous SmartCode loads alongside the page and hides the body until it is done.

What is an anti-flicker snippet?

A style that hides the page, or only the parts a test will change, until the tool has applied its change or a timeout runs out.

The timeout is the worst case. On a fast connection the tool usually answers well before it, as Webflow's own help page says. On a slow phone connection it can run to the end, and then the visitor has waited that long for the page they would have seen anyway.

ToolWhat it hides by defaultHow long before it gives upSource, as read
Google Optimize (closed on 30 September 2023)The whole document, in its snippet example4,000 ms in the snippet exampleSpeedCurve, 28 April 2022; Google's closing notice
VWO, asynchronous SmartCode (its help center now says Wingify)The page body, hide_element = 'body'2,000 ms for its settings; 2,500 ms more for its library on SmartCode 2.2 and earlierVWO help, updated 26 August 2026. vwo.com's own tag uses the body and 2,000 ms (our curl check, 25 October 2026)
Optimizely Web ExperimentationNo hide in the default install: the snippet loads synchronously, first in the headNone stated; the docs offer a visibility: hidden mask only for a non-blocking installOptimizely docs, July 2026
Webflow OptimizeThe elements a variation changes4,000 ms; only Enterprise customers can change it, from 0 to 30,000 msWebflow help, updated 9 September 2026

What does hiding the whole page cost?

On one measured homepage, more than three seconds of LCP.

DebugBear tested AdRoll's homepage, which hid the page while Mutiny's personalization decided. With the hide overridden, LCP went from 6 seconds to 2.7 (published 1 February 2022, updated 7 December 2025). Mutiny has since retired that product.

That is one site on one day, measured by a monitoring vendor. With the body hidden, nothing on the page counts as painted until the tool lets go.

Does a script in the head hurt Core Web Vitals?

Only through what your visitors experience: Google scores LCP, INP and CLS from real Chrome visits, not from whether a script is there.

A page passes when 75% of visits have LCP within 2.5 seconds, INP within 200 milliseconds and CLS of 0.1 or less (web.dev). Google's Search docs say good scores line up with what its core ranking systems reward; they do not say how much that weighs against relevance.

PageSpeed Insights shows the same real-visit data, from the Chrome UX Report over the last 28 days at the 75th percentile, when your site has enough Chrome traffic. Its lab score below that is one simulated load.

Whatever a script hides is painted late, and if it is the largest text or image on screen, LCP waits for it. A layout shift counts only when an element that was already visible moves, so a hide with visibility shifts nothing, while a new headline one line longer than the old one pushes everything below it and does count. INP suffers when a heavy script keeps the main thread busy and a tap has to wait.

Is hiding only the changed lines better than hiding the whole page?

For the visitor, yes: the nav, the images and every section the tool leaves alone paint at once, and only the changed lines wait.

Both kinds still wait for the same answer from the server, and both count in LCP when the hidden part is the largest element on screen. With the body hidden, the visitor looks at a blank screen; with the lines hidden, they see the page with an empty space where the headline will be.

The timeout sets the worst case either way: 2 to 4 seconds for the tools in the table above.

What did our own speed test measure?

On a real, shaped phone connection, the Rawuser tag moved our lab page's LCP from 0.60 s to 0.99 s with no signal and to 1.13 to 1.21 s for a ChatGPT visit, on a first view, with a CLS of 0 in every load (our own test, 26 October 2026).

We tested our lab page, rawuser-measure-lab.pages.dev, a 2 KB page with one stylesheet, with DebugBear's free speed test: Chrome on a mobile profile in US East, 12 Mbps, a 70 ms round trip and the CPU slowed 2 times, an empty cache each time. The control is the same page with the tag line removed (/no-tag/). The ChatGPT visit (utm_source=chatgpt.com) gets a rewritten headline and paragraph. We ran each column twice; the free test queues runs, so read these as two runs, not medians.

About half of the cost is the first paint. The tag is a script in the head without async, on our own host, so the browser opens a new connection to it and loads a 4 KB starter before it paints anything: 0.86 to 0.99 s against 0.60 s without the tag. The rest is the rewritten headline, which stays hidden until the answer for this visit comes back; the page's other lines are already on screen. The versions are written before anyone visits, so the answer is a cache read, not a model call.

Our earlier test emulated a slow phone in headless Chromium and found less (0.29 s for a ChatGPT visit after the tag was made faster): emulation slows the download but not the opening of a new connection, which is exactly what a first view pays for.

On a plain page like this one, the script shows its lines by 1.5 seconds at most on a first load. On a server-rendered React site it can wait up to 2.5 seconds more after the version arrives, for React to take over the page, unless the site marks its parts for an early write. If the script stops after hiding, a CSS rule shows the lines anyway. A tag placed after the page has painted hides nothing and leaves that load as built.

These are two runs on one small page. The lab page loads in 0.6 s, which makes the tag's share look large; on your page the cost may differ, and Google scores the 75th percentile of your real visits, not our runs.

Lab page, first view, shaped phone connectionWithout the tagTag, no signalTag, ChatGPT visit
First paint (FCP)0.60, 0.61 s0.82 s0.86, 0.99 s
Largest paint (LCP)0.60, 0.61 s0.99 s1.13, 1.21 s
Layout shift (CLS)000
Runs212

What should you check on any tool before you install it?

Five things, each in the vendor's docs or your page source: how the script loads, what it hides, its timeout, its size, and what the visitor waits for.

Start with how it loads. In the head without async, the script holds the page until its file arrives; with async or through a tag manager, expect a flash of the original. Ask the vendor which one it expects, and look in the snippet for body, opacity: 0 or visibility: hidden to see what it hides.

Then find the timeout and who can change it. The tools above wait 2 to 4 seconds, and Webflow lets only Enterprise customers change its.

Last, check the size of the file your visitor downloads, gzipped, and whether the tool writes a version while the visitor waits or reads one made in advance. A version read from a cache comes back in milliseconds.

How to test any script on your own site, with no tool

  1. Read the tag in your page source.

    Open your page, press Ctrl+U (Option+Cmd+U on a Mac) and search for the vendor's domain. Note whether the script sits in <head>, whether it has async or defer, and whether a style next to it hides body.

  2. Run PageSpeed Insights.

    Paste your URL at pagespeed.web.dev and read the mobile tab. The top part is your real visitors over the last 28 days, if your site has enough Chrome traffic; the part below is one simulated load. Write down LCP and CLS.

  3. Block the script in Chrome and compare.

    Open DevTools, press Cmd+Shift+P (Ctrl+Shift+P on Windows), type block, choose Show Request Blocking, add the vendor's domain and reload. In the Performance panel, record a reload with the network set to Slow 4G and the CPU slowed 4 times, once with the block and once without, and compare the LCP marker.

  4. Test on a real, shaped connection.

    DebugBear's free speed test needs no account and loads your page on a shaped mobile connection, which catches the cost of opening a connection to the vendor that Chrome's own throttling hides. WebPageTest does the same with a free account, and its Block tab lets you run the page without the vendor's domain to compare. Run each side at least twice and compare the filmstrips to see the moment your hero appears.

  5. Compare medians, not single runs.

    Run each side at least five times and compare the middle values. If the gap is smaller than the spread between runs, the script's cost on that page is too small to see this way.

Sources

  1. web.dev, Largest Contentful Paint (updated 4 September 2025) read 26 October 2026
  2. web.dev, Cumulative Layout Shift (updated 12 April 2023) read 26 October 2026
  3. web.dev, Interaction to Next Paint (updated 2 September 2025) read 26 October 2026
  4. web.dev, efficiently load third-party JavaScript (updated 19 February 2024) read 26 October 2026
  5. Google Search Central, Core Web Vitals and Google Search results (updated 10 December 2025) read 26 October 2026
  6. PageSpeed Insights, about (updated 21 October 2024) read 26 October 2026
  7. Google Optimize Help, Optimize and Optimize 360 closing notice read 26 October 2026
  8. SpeedCurve, anti-flicker snippets and web performance (28 April 2022) read 26 October 2026
  9. VWO (Wingify) help, customize your SmartCode (updated 26 August 2026) read 26 October 2026
  10. VWO (Wingify) help, asynchronous SmartCode (updated 23 October 2026) read 26 October 2026
  11. Optimizely docs, install the snippet as a non-blocking resource (July 2026) read 26 October 2026
  12. Optimizely docs, page load speed (July 2026) read 26 October 2026
  13. Webflow help, anti-flicker prevention in Optimize (updated 9 September 2026) read 26 October 2026
  14. DebugBear, A/B testing and anti-flicker body hiding (1 February 2022, updated 7 December 2025) read 26 October 2026
  15. Mutiny blog, introducing the new agentic Mutiny read 26 October 2026
  16. Chrome for Developers, block requests in the Network panel (updated 14 October 2024) read 26 October 2026
  17. Catchpoint docs, WebPageTest instant test (updated 26 August 2026) read 26 October 2026
  18. DebugBear, free website speed test read 26 October 2026

Related questions

Start your pilot

See it on your site.

Free for 30 days. Leave your email and your page; we reply within a day with what your own page would say to a visitor from ChatGPT.

We use your email only to reply. Privacy

What happens next

  1. 01

    Within a day, your own page.

    What it would say to a visitor from ChatGPT, written from your page.

  2. 02

    We set it up with you.

    You see your page and its first version before anything goes live.

  3. 03

    Day 31, you decide.

    Keep it from $99 a month, or stop and owe nothing.