Capturing user intent during signup is a common SaaS tactic. Ask users about themselves before they get into the product, then use their answers to tailor the experience when they land.

Shopify has done some version of this for years. When a new merchant signs up, they answer a short questionnaire about their business: what they plan to sell, which channels, and so on. Those answers are what shape the guidance they see when they land in the admin.

The problem with this approach is that identity questions are genuinely hard to answer. Most people don’t think of their business in neat categories, and the options we offered rarely mapped how merchants actually saw themselves.

At Shopify, we saw the results of this firsthand. A merchant signs up for a free trial and either skips the questionnaire or picks the wrong options, rushing through it or over-selecting just to move on. Then they land in the admin and find that the onboarding we built for them doesn’t match what they actually need.

One of the questions merchants answer during signup.

Worst of all, there’s no way to course correct. Once they’re in the admin, they either have to make do with what we gave them or start over from scratch.

The result is a system that fails at both ends. Merchants get guidance that doesn’t fit, and the signals we collect from the questionnaire are unreliable, wrong, or missing entirely.

In the fall of 2025, I started looking at this more closely. By then, the Activation team had already started on a possible fix: instead of relying on the questionnaire, they were tracking admin events as signals of intent and using them to trigger guidance. It was a smarter input, but it worked the same way the questionnaire did, inferring what a merchant wanted rather than asking.

I started working with that team to sharpen those signals, figuring out which combinations of actions pointed most reliably toward what a merchant needed, and how we might instrument event-driven guidance at a larger scale.

But the deeper I got, the more it looked like we were solving the wrong problem. More signals, better signals, smarter signals, none of it changed the fact that we were still trying to infer what merchants wanted instead of letting them tell us directly.

That realization shifted how I thought about the problem entirely. Instead of getting better at guessing, what if we just made it easy for merchants to tell us what they wanted, once they were already in the product and had a reason to?

In late November, I kicked off a sprint to explore the idea. The core question was whether we could build something on Home that surfaced clear, actionable options a merchant could opt into, and let them self-select based on what actually mattered to them.

I paired with a PM and an engineer and spent a week prototyping a proof of concept, enough to get a reaction from leadership. We presented it and got buy-in quickly. The hypothesis was that a clear menu of actions in the admin would do two things: help merchants understand what Shopify can actually do for them, and make it easier to go do those things. If it worked, we’d see it in our activation metrics, the lead-to-converted rates we track at different points after signup, like 7 and 30 days.

The onboarding bento on Home that any new component had to work alongside.

That came with a real question, though. How do we put something new and useful on Home without pulling merchants away from the core onboarding tasks they still needed to finish?

The answer came from work already in motion. The Sidekick team was about to add a Sidekick input to Home. Instead of crowding the page with another component, I saw I could build the discovery experience right around that input. It solved the clutter problem by using something the page was already getting.

So in January 2026, I prototyped a few variations in Cursor to see which held the most promise.

The first put rotating placeholder text inside the Sidekick input. Suggestions like ‘Find products to sell’ or ‘Migrate your store’ cycled through on their own, giving the input some life without taking up extra space. But there was no way to act on a suggestion without retyping it, each one rotated out within a second, and the more suggestions I added, the longer a merchant had to wait to see one that was relevant. The approach made discovery feel incidental, when it needed to be the point.

An early prototype exploring rotating suggestions inside the Sidekick input.

The second used clickable suggestion pills beneath the input, where clicking a pill opened a dropdown of more specific options. The idea was that a few broad pills could cover a lot of ground. But some of the specific options were important enough that hiding them behind a click felt wrong, and not every pill had enough sub-options to justify a dropdown, so the pattern felt inconsistent. The lesson was that the most useful options shouldn’t be buried, they should be the first thing a merchant sees.

A prototype where clicking a pill expanded a drawer of specific options.

The third design was the simplest, and it followed directly from that lesson. Static pills, each one triggering a Sidekick conversation when clicked. No dropdowns, no typing. A merchant clicks a pill, Sidekick opens, and the conversation handles the rest.

The shipped experience; each pill triggers a Sidekick conversation.

With the pattern settled, the work moved to the pills themselves: defining them and designing what happens after the click. Each pill passes context to Sidekick and triggers a conversation tailored to that action, whether that means generating a to-do list, answering a question, or routing the merchant to a specific page in the admin. I worked with engineering to test and refine each one until it felt right.

Then there was the question of which pills to show, and when. I had intuitions about which would land, and even built some segmentation logic to start. But fixing those choices in advance would have put us right back in the business of guessing, and a system that learned from real behavior would beat any rules I could write up front. So we served the pills with a multi-armed bandit model.

A bandit is built for this exact problem. It shows the options we think will work while still testing ones we’re unsure about, and shifts toward whatever merchants actually click. It starts from our best guesses and improves on its own over time. The same idea that shaped the whole project, let behavior tell us instead of guessing, was now running the system itself.

It’s fair to ask why we solved this at the end of the flow, once a merchant was already in the admin, instead of fixing the signup questionnaire itself. Part of it is ownership: the questionnaire belongs to another team, and while we influence what goes into it, a perfectly tuned questionnaire is both hard to build and only a partial fix. It still asks merchants to define themselves before they’ve done anything.

Giving them a way to express and act on intent once they’re in the product felt like the more durable solution. Earlier in 2025 I explored replacing the questionnaire altogether with conversational AI. That work didn’t ship, but it’s part of why I was convinced the real fix lived in the product itself, not in a better set of questions.

When a merchant lands in the admin today, they find a set of pills alongside their onboarding tasks, no matter how they answered the questionnaire, or whether they answered it at all.

What they can do on Shopify is no longer gated behind intent we tried to infer for them. Each pill is something merchants commonly want to do: open a new sales channel, generate product images with Sidekick, source products, migrate a store. Clicking one opens a Sidekick conversation built around that action.

The merchant doesn’t have to know what to ask, or even know what Sidekick is. They see something useful and click.

The bet held. The project drove over $10M in gross profit and a 25% increase in new merchant engagement with Sidekick. Letting merchants tell us what they wanted beat trying to guess it.

The more interesting outcome is what it does for the cold start. One of the hardest problems with any AI tool is that most people don’t know what to ask, or what the tool can even do, so they never start at all. The discovery unit gets around that. A merchant clicks a pill, has a useful conversation, and comes away understanding what Sidekick can do for them. That understanding compounds across the rest of their time on Shopify.

The questionnaire still exists, and event signals still inform guidance. But now there’s a third layer, one that doesn’t ask us to guess at intent, and doesn’t ask merchants to define themselves before they’re ready. They just see something useful and click.