Split-URL A/B tests add a redirect, and on Android it cost a median 3.2 seconds

Split-URL testing is popular because it's easy to understand. Half your traffic goes to URL A, half goes to URL B, and you see which one wins. What's easy to miss is how the visitor actually gets to URL B. In most client-side setups the browser loads the original page first, a testing script picks the variant, and then the browser gets sent on to the variant's URL. That extra hop is part of what you're testing.

We looked at a DTC store that ran a lot of split-URL tests on its paid landing pages. The redirect added a median 2.7 seconds on iOS and 3.2 seconds on Android. On Android, one redirect in ten took more than 20 seconds and 27.6% took more than 10. Only 15% to 21% of redirected visits could be measured end to end, because a lot of visitors left before the original page's analytics fired. So what you end up comparing is the variant plus a redirect against a control that loads directly.

How a client-side split-URL test works

Ad click
  → the original URL starts loading (HTML, then the testing script)
  → the script assigns the visitor to a variant
  → the browser navigates to the variant URL
  → the variant starts loading from scratch

Everything before that second navigation, including the original page's server response, its scripts and any flash of the original content, gets charged to the variant.

What we measured

We used the store's warehouse data from July and August 2026, plus lab runs on the same pages.

I'd treat the lab number as a worst case. The warehouse medians are what real visitors got.

Your numbers will be different. But any client-side redirect that happens after the first page starts loading adds a round trip, a script download and a second page load, and mobile networks make each of those slower and less predictable.

Why more traffic doesn't fix it

A lot of people assume you can just run the test longer. I don't think that works here, for a few reasons.

1. The delay biases the result, and more data doesn't remove bias. Extra load time changes conversion, so a bigger sample just measures that skewed difference more precisely.
2. The data goes missing unevenly. If most redirected visitors who give up never make it into your analytics, the variant's conversion rate gets measured on the patient ones. That can make a slow variant look better per recorded visit than it really is.
3. The two arms aren't the same. If the control loads at the original URL and only the variant redirects, you're comparing a page against a slower copy of a different page. If both arms redirect, both pay the cost and you still haven't isolated the page.
4. The worst delays hit the traffic you care about most. Android on mobile networks had the longest delays, and Android phones carry a big share of paid social traffic.

How to audit your own tests in ten minutes

1. List your live ad destinations. Open your active ads and write down every final URL, including the parameters.
2. Load each one on a throttled phone profile, three times. In Chrome DevTools, pick a mid-range phone and Slow 4G, and turn off the cache.
3. Watch the Network panel for a second document. If the first HTML response is followed by another HTML document for a different URL, the test has a redirect.
4. Look for testing scripts in the page source. Most testing tools leave a recognizable script or network call on the original page.
5. Sort each destination into one of four buckets: same-URL test, split-URL with a redirect, assignment at the edge or server before the page is sent, or no test running.

If step 3 shows a hop, your result is really "page plus redirect", and I'd report it that way.

What to do instead

If you run split-URL tests on pages that search engines crawl, Google's guidance is to use temporary (302) redirects and a rel="canonical" link from the variant back to the original URL, and to take the test down once it ends (Google Search Central). That's about protecting your rankings. It doesn't do anything about the delay visitors see.

At Lutiq we assign each paid click at the edge, before the first byte. We call it one decision, one load: the page being tested loads without a redirect in front of it.

A few caveats

These numbers come from one store's data and lab runs in July and August 2026, not from a controlled comparison of testing tools. How many visits were measurable depends on where that store's analytics fired. And the cost of a redirect depends on the size of the original page, the testing script, the visitor's network and whether things are already cached.

FAQ

Do A/B testing redirects slow down the page?

A client-side redirect adds a full second page load after the original page has already started. In our data that was about 3 seconds at the median on mobile, with a long tail above 10 seconds on Android. Same-URL tests and assignment at the edge or on the server avoid the second navigation.

Is split-URL testing bad for SEO?

Not if it's set up the way Google recommends: temporary 302 redirects, a canonical link from each variant to the original, and removing the test when it ends. That guidance doesn't cover the delay visitors experience, though.

Why don't my split-URL test results match my analytics?

Visitors who leave during the redirect often never trigger analytics on either page. In our data only 15% to 21% of redirected visits could be followed end to end, so your testing tool and your analytics can easily be counting different groups of people.

What's the fastest way to A/B test landing pages?

Decide the variant before the page is sent, at a CDN edge worker or on the server, and serve the chosen page as the first response. The visitor loads one page, and both arms start from the same place.

Related

• This tracking tag didn't fire until 8 to 58 seconds into the visit
• Lutiq Holdout Protocol
• What is a holdout test

---

I'm Colin Behr, co-founder of Lutiq. Before Lutiq I spent years in mobile ad tech at AppLovin, Vungle and Branch. Lutiq predicts which on-brand landing page to show each paid click and measures the lift against a live holdout. The figures in this post are our own measurements from one store in July and August 2026 unless noted, and the brand isn't named.