Split URL Testing

Overview

Split URL tests route visitors to entirely different pages (URLs) to compare performance. Convert supports two ways to run them:

1. Split URL Experience Type (recommended) — Handles everything automatically
2. Manual redirects — Using convert.redirect() and convert.refresh() in other experience types

How Split URL Bucketing Works

Unlike standard A/B tests where variations modify the current page, split URL tests require navigation — the visitor must be sent to a different URL. This creates a challenge: how do you count the visitor if the redirect might fail?

Convert solves this with a deferred bucketing mechanism:

sequenceDiagram
    participant V as Visitor
    participant S as Script
    participant C as Cookie
    participant B as Backend

    V->>S: Loads original page
    S->>S: Decides variation
    S->>C: Stores decision in _conv_sptest (15s TTL)
    S->>V: Redirects to variation URL (or refreshes original)
    V->>S: Loads target page
    S->>C: Reads _conv_sptest
    S->>B: Sends bucketing event
    S->>C: Deletes _conv_sptest

The Three Phases

1. Decision — The script decides which variation the visitor gets and stores the decision in the _conv_sptest cookie (15-second TTL)
2. Navigation — The visitor is redirected (variation) or the page is refreshed (original)
3. Tracking — After the page loads, the script reads the split cookie, sends the bucketing event, and deletes the cookie

Key insight: The visitor is only counted after navigation succeeds. If the redirect fails (network issue, JavaScript blocked, visitor closes tab), the visitor is never counted. This prevents Sample Ratio Mismatch (SRM).

Why We Refresh the Original

You might wonder: why does the original variation also require a page refresh? The answer is symmetry.

If only the variation redirects, then:

• Original visitors are counted immediately (no navigation barrier)
• Variation visitors are counted only after successful redirect
• This creates SRM because the original includes visitors who would have abandoned during redirect

By refreshing the original, both variations go through the same flow:

• Both break page execution
• Both require successful navigation to be counted
• Failed navigations are excluded equally from both groups

Manual Redirect API

When using A/B or Multivariate experience types with manual redirects, use these methods:

convert.redirect(url)

Navigates the visitor to the specified URL. Use in Variation JavaScript.

// Variation JS
convert.redirect('/new-landing-page');

convert.refresh()

Reloads the current page. Use in Original Variation JavaScript to pair with the redirect.

// Original Variation JS
convert.refresh();

Correct Pattern

// Original Variation JS
convert.refresh();

// Variation JS
convert.redirect('/new-landing-page');

Common Mistakes

The Split Cookie

The _conv_sptest cookie has a deliberately short lifetime:

• 15-second TTL — Prevents stale bucketing data from persisting
• Always written — Even when cookie consent hasn't been given (it's strictly functional)
• Deleted after use — Removed once the bucketing event fires

If navigation doesn't complete within 15 seconds, the cookie expires and the visitor is bucketed fresh on the next page load.

When to Use the Split URL Experience Type

The dedicated Split URL experience type is recommended because it:

• Handles refresh/redirect pairing automatically
• Prevents the common SRM mistakes listed above
• Supports "Transfer Original URL variables to the variation URL" for query parameter preservation
• Handles all edge cases without custom JavaScript

Use manual redirects only when you need complex conditional logic that the Split URL experience type can't express.

Debugging

• SRM warning in the Convert app — Check if Original has significantly more visitors than expected
• Inspect _conv_sptest — Should appear briefly (15 seconds) after page load
• Network tab — Bucketing events (/track/ requests) should appear only after redirect/refresh, not before
• Quick test — If you can trigger bucketing without a page reload, the implementation is incorrect