PixelMe Product Redesign
Rebuilding an acquired ad-tech product from the foundation up, so people and AI could both build from it.
Context
PixelMe launched in 2017 as a URL shortener that embedded retargeting pixels. Carbon6 acquired it at the start of 2022, specifically for its Amazon Attribution integration, and SPS Commerce then acquired Carbon6 for $210M in February 2025. By the time I picked up the work the product was carrying nine years of accumulated features through two acquisitions, and every new user still landed on the link shortener.
What that added up to was six top-level areas, four competing paths to the same link and attribution job, and two separate top-level sections both built around Amazon Attribution. There was no single artifact anywhere that described what the product actually was, and the public help center documented the retiring product in exhaustive detail while the priority workflows had nothing written about them at all.
The business needed PixelMe to sell itself, meaning a seller could find it and adopt it without getting on a call with anyone. Self-serve onboarding was closed at the time, because the paths through it did not work.
My Role
I led the redesign end to end: discovery, research, information architecture, workflows, the design system, visual design, documentation, stakeholder alignment, the AI layer and handoff. Three months, two PMs, three engineers, and a second designer working full time alongside me while I carried my own IC load.
I worked closely with the CS leads the whole way through. They had the only sustained relationship with both the product and the customer, which made them the closest thing to a user proxy the company had. I presented to the executive team twice and delivered into engineering in stages rather than in one drop.
Discovery & Research
Three tracks running at once. Five user interviews and eight internal ones, on top of hundreds of feedback items CS had already collected through calls and surveys. Ten competitors in adtech and B2B analytics, captured and turned into a pattern reference. And working sessions with product and engineering to reconstruct what the product actually did and which parts of it were on their way out.
Out of that came the first sitemap PixelMe had ever had: 70 nodes across six top-level areas, closer to 90 if you count the multi-step flows collapsed inside them.
Two findings changed the shape of the work. Amazon Attribution was the strategic center and the reason the product had been acquired in the first place, and yet several surfaces still treated it as one option among many. The other was that Link Manager, One Link and Smart Attribution were all on their way out, so redesigning them would have been weeks spent on features scheduled to disappear.
I also ran an audit of the live product, and the first pass gave me findings anyone could have produced. It only got useful once the business context went in on top of it: Attribution as the priority, the retirement list, workspaces being agency-only and capped. Same evidence, read against what the company was actually trying to do, which is where the value of an audit comes from.
What I Built
Information architecture. 70 nodes down to 28, and six top-level areas down to three, a 60% reduction. Two of the six products were retired outright, which accounted for 27 nodes, or 39% of the original map. Another 15 or so were consolidated rather than cut: the two duplicate Amazon Attribution areas merged, and Account and Billing collapsed into one Settings area. Pixels and custom domains survived as settings-level capabilities instead of standalone products.
Design assets. 150 screens and states designed or redesigned across 8–12 core flows, which was the first complete set the product had ever had. It covers zero states, skeleton loaders, validation errors, success and failure, trial expiry and impersonation.
The design system. There was structure there already, so this was a restructuring rather than a build from zero. I rebuilt it into three explicit layers (110 primitives, 75 alias ramps, and 57 semantic mapped tokens with light and dark modes), plus 17 text styles bound to responsive tokens, and 24 component sets with roughly 25 composed patterns built on top of them. All of it is documented down to token semantics, component states and contrast ratios, so the rules are written down somewhere instead of living in people's heads, and that is what let both a designer and an agent build from it correctly.
The AI layer. I rebuilt the system to be correct, and being correct is what made it machine-readable. On top of it there is a skill per document type and per audit, four more covering the design system itself, and CLAUDE.md plus its supporting files acting as the standing contract every build reads from. The documentation is written as machine input rather than as write-ups, so a designer and an agent are reading the same source.
What Was Hard
The design system had structure but it was not holding. Tokens existed and did not get used consistently, and components had drifted from each other. I could have built the redesign straight on top of it and had screens to show inside the first month, and then paid for that every month after, so I stopped and rebuilt the foundation first, which I was doing under real pressure to show screens instead.
I built it for two readers at the same time, the design and engineering teams, and the AI agents I intended to build with. That turned out to be one requirement rather than two, because a system that is correct enough for a person to trust is the same system an agent can read.
Making that argument was the harder part. Foundations before screens is a familiar case and people know the shape of the answer. Foundations before screens so that AI can build from them was still speculative in the room, and I defended it to product and to the executive team on the compounding cost of not doing it.
What Happened
The new IA is live, along with the new navigation. Workspaces and Settings shipped, and Amazon Attribution Campaigns is partially delivered.
Self-serve onboarding opened. It had been closed because of broken paths, billing screens that were not available, and legacy surfaces sitting in the way.
Ticket time fell by nearly half across bugs and features combined, most likely because engineering was now working in a cleaner and more familiar codebase.
Workspaces works for everyone now. It had been broken, capped at a single workspace, and gated to agencies. CS reported back that it resolved long-standing problems for several VIP customers.
Three teams now build on the design system, with 24 component sets and roughly 25 composed patterns in production, and Link Manager, One Link and Smart Attribution are retired from the user-facing product.
Design System
Multi-layered Tokenization / Faster development
Confluence Guide & Documentation
Shared knowledge / Easier maintenance