Every export option converted better once people could find it.

No hierarchy at the highest-traffic moment in the product.

The product is a browser-based PDF editor. When a user finishes editing a document, a modal appears asking "What would you like to do?" — download it, email it, send it for signature, print it, save it somewhere. Hundreds of thousands of new desktop users see it every two weeks. It is, by traffic, the single most consequential decision point in the product.

It is also where most first-time users meet the paywall. Roughly half of new-user revenue is attributable to sessions that touch this dialog. That makes it a monetisation surface as much as a navigation one — and it had grown by accretion. Every new feature over several years had earned itself a tile, and nothing had ever been taken away.

Attention was pointed at the wrong things.

I started with two experiments' worth of event data and ranked every action twice: once by how often people clicked it, once by how well those clicks converted. The two lists came out almost exactly inverted.

The single best-converting action in the product sat dead last by clicks. Email, the second best, sat mid-table. Meanwhile the Word / Excel / PowerPoint cluster took a large share of clicks and returned almost nothing — a near-total drop-off between clicking a format and paying for it, which pointed at a paywall or pipeline problem, not a demand one.

That is a layout problem, not a demand problem. Seventeen tiles at identical visual weight let default scanning behaviour decide the outcome, and default scanning behaviour rewards whatever is nearest and most familiar.

Four frictions, and what each one was doing

Seven products, one pattern.

Before proposing anything, I audited the equivalent moment in SmallPDF, iLovePDF, PDFCandy, Sejda, Canva, Google Docs and Google Slides. The convergence was striking: none of them present a flat grid, and all of them make the same five moves.

Six decisions, each with a reason I could defend.

This surface generates enough revenue that "it looks cleaner" is not an argument anyone should accept. So every change went into the spec paired with the principle it rests on and the behaviour it was meant to change. That pairing is what got the redesign through review.

Two things I proposed did not ship, and I think both calls were right. An AI section (summarise / translate / ask) was cut from this test and split into its own variant, so a latency-and-cost bet wouldn't contaminate a layout result. And renaming "Collect forms" to "Publish as form" was dropped — the team wanted the naming decision made across the whole product, not inside one modal.

A/B/C, desktop, new users without a plan.

The redesign went out as a controlled experiment against the existing modal, with the AI-heavy variant held as a third branch so that layout and AI wouldn't be confounded. Two primary metrics were agreed before launch: the share of viewers who give card details in-session, and the share who complete an export.

Two readings of the same experiment.

At the population level, both primary metrics moved the wrong way — one of them comfortably significant. Read only that far and the redesign looks like a loss.

Read one level down and it isn't. Here is the same experiment, both ways.

So: every action a user could take converted better than it did before — the two biggest movers by roughly a quarter — and two surfaces that hadn't existed at all started producing conversions on their own.

Why the average moved the other way

Because the redesign worked — hardest of all on the action it made most prominent. Download PDF's share of Choice viewers nearly tripled. It became the largest cohort in the experiment by a wide margin — and it was the only action whose conversion stayed effectively flat, with a real drop in registrations.

The mechanism is straightforward once you see it. Download PDF is now the first, most prominent thing in the panel — the path of least resistance. So it absorbs a large volume of low-intent, exploratory clicks from people who previously drifted out of the dialog without clicking anything, or who had to scroll to find the thing they actually wanted. Those clicks are real, they are just not intent. Averaged into a cohort, they dilute it.

What I'd do differently.

• Promote the best converter, not the most popular one. I gave the primary CTA to the action with the most demand. I'd give it to the action with the most demand per click — and treat download as the calm default that doesn't need help being found.
• Fix the download funnel first. A registration drop on the single biggest cohort is its own problem, independent of this layout. Sending more traffic into a weak funnel makes the weakness the headline.
• Personalise the panel to the file. Everyone sees the same list, whatever they just edited. A filled-in form, a signed contract, a scanned receipt and a slide deck imply completely different next steps — the panel should read the document and propose the actions that actually fit it, instead of one fixed set for every user.