The test needed three more section slots.The template had none left.
A US DTC supplements brand on Shopify. A three-arm placement test in Convert.com, 2025.
A three-arm placement test on a landing page for a US DTC supplements brand on Shopify. The standard way to build it was unavailable, so it got built a different way and shipped in three days.
The hypothesis was the CRO strategist's. The build, the tracking and the delivery were mine.
What the test was meant to do
The CRO team wanted to know where an urgency block performed best on a dedicated product landing page. Three arms: the control, the block above the product section, and the block below it.
I did not write this hypothesis and I do not claim it. My work started at the point where it became a build.
What blocked the standard build
A placement test is normally the cheapest kind of test to build. In Convert you duplicate the section, hide the original, unhide the duplicate, and let the visual editor handle the move.
Two things made that unavailable, both found on opening the template.
| The assumption | What was actually there |
|---|---|
| One section moving | Three separate sections that had to move together and hold their order |
| Room to duplicate them | The template was already at Shopify's limit of 25 sections. Zero slots free. |
Reordering through the visual editor on top of that was unreliable, because the DOM manipulation had to hold across three elements rather than one.
I flagged this the same day I found it, which was a Saturday, rather than discovering it slowly across the following week.
What got built
I proposed a custom section instead of duplicated ones: a single section that renders the urgency block in either position depending on the variant, so the test consumed no additional section slots at all.
Before committing to that I took it to Convert's support team to confirm there was no cleaner path inside the platform I had missed. That took a working day. There wasn't.
Then: the custom section, three variants wired to it, and the goal configuration.
I delivered three preview URLs, one per arm, alongside the goal setup, so the strategist could review the actual rendered variants rather than a written description of them. She approved it and the test went live the same day.
Constraint identified to live three-arm test, on a build the standard method could not produce.
What I can't show you
There is no result on this page, and that is deliberate.
The test ran. What it read lives in a Convert account I no longer have access to, and it was never written into the client's tracking sheet at the time it closed. So I have no primary source for the outcome, and I am not going to publish a number I cannot source.
The process, not the build.
Every step above is backed by a dated, attributable record: the constraint, the escalation, the three preview URLs, the go-live approval. The one thing missing is the two lines someone should have written into the tracking sheet when the test finished. That omission is why this section exists instead of a result, and it is the habit I changed.
How I check a result is real
In August 2026 I went back through every experiment I could still find records for and ran a sample ratio mismatch check on all 56 that reported per-variant traffic. Chi-square, one degree of freedom, p < 0.01.
To be precise about the timing: I ran this audit retrospectively, not while the tests were live. That is part of why I am showing it.
54 of 56 passed. The two that failed both matter.
| What it was logged as | What it actually was |
|---|---|
| An A/B test with above 99.9% confidence | Not an experiment. The arms ran in different windows, 60 days and 73 days, starting six weeks apart. A before-and-after comparison, so the confidence figure is meaningless. |
| A clean 50/50 split | A 49.26 / 50.74 allocation across more than half a million sessions. A 1.5% skew, far outside chance at that volume. The randomiser was not splitting evenly, so the result cannot be trusted. |
Neither is visible from the platform's results screen. Both tests looked fine in the UI. You only find them if you check the split yourself.
The arithmetic sits as live formulas over the raw visitor counts in my own records, so it recomputes and can be checked rather than taken on trust.
Client name, absolute revenue and absolute traffic figures are withheld under NDA. Every claim above traces to a dated primary record: the project management task and its comment thread, the experiment and variation IDs, and the client tracking sheets. Available for verification in a private conversation.