Trading a good result for a $1B+ one inside Amazon search
I bet that a few trusted recommendations, placed right inside the results, would help shoppers stop comparing and choose confidently. The data proved me right, once I got out of its way.
- Company
- Amazon · Search & Discovery
- Role
- Principal UX Designer
- Team
- Solo designer with PM and engineering team
- Read
- 10 min
$1B+
in incremental annual sales
~80%
of search results pages when I left
2
US patents granted from this work
Dozens
of retail categories adapted to a single modular system
The short version
Proprietary internal metrics, return-rate data, and backend ranking weights omitted under NDA- Context
Decision fatigue had been a known problem on Amazon search for years. Broad searches returned hundreds of thousands of near-identical products, and shoppers compared until they gave up. Teams across Amazon were trying to fix it, mine included.
- Problem
Each fix helped, but nothing was a huge winner. And the place that could help most, the search results themselves, was treated as sacred.
- What I did
With my PM, I pitched adding product recommendations directly in the search results and took it to Amazon’s senior leadership. I designed a modular template for the component. I tested my way from my first idea, expert picks, to crowd data, with a placement that floats to wherever it performs best. And while I owned it, I kept ads out.
- Impact
My biggest win at Amazon: more than $1B in incremental annual sales, based on controlled A/B tests, appearing on roughly 80% of search results pages by the time I left.
Business problem
The catalog was the advantage. It was also the problem.
Amazon’s selection of hundreds of millions of items is why people shop there, and it’s also what makes choosing so hard. Search "kitchen knife," and you get hundreds of thousands of options that all look about the same. Years of research kept saying the same thing: shoppers felt overwhelmed by choice.
Amazon had been chipping at this since at least 2015, and it’s still shipping new fixes today. Designers on my team added badges like Amazon’s Choice and Best Seller. I’d already rebuilt search filters and tested smaller ideas, like hiding the total result count because the number alone put people off. Each helped, but nothing really broke through.
The most direct fix was to help people pick right there on the results page. But search results were on a pedestal. You didn’t put anything new in them, and changing that meant convincing the S-team, the SVPs who reported to Bezos.
What it cost
At Amazon’s scale, small search wins move millions of dollars. Every search that ended without a purchase was money left on the table.
The users
Paralyzed by parity
Shoppers said there were too many options. What they were missing was any sign of which one was the right choice.
When fifty products have four-and-a-half stars and nearly identical photos, comparing specs stops helping. Hand shoppers another flashy widget or something that reads like a paid promotion, and they scroll right past it or abandon the search entirely.
"There are 50 pages of results for ‘white running socks’ and they all look identical. I don’t want to read a thousand reviews, I just want to know which one is the ‘good’ one so I can get on with my day."
Solution
The bet, the document, and a test plan
A series of decisions in the order I made them
I bet that a few trusted recommendations, placed right inside the results, would let shoppers stop comparing and choose confidently. The open question was whose recommendation they’d trust. That bet meant working on a page Amazon treated as sacred.
With my PM, I wrote a six-page proposal that made the case and set the direction. It’s how ideas get funded at Amazon. It won the S-team’s buy-in, and Amazon allocated an engineering team to build the supporting data. The doc, test plan, and my mockups shaped the direction.
I owned the test plan. Content, placement, and design were all competing hypotheses, each run through two-week multivariate tests. The data had the final say, even over my own ideas.
Decision 01 of 08
Built the component to flex
Design a modular platform for what might win, not a single widget for a single guess.
Nothing like this existed in the results, and we couldn’t be sure what content would win. So instead of designing one widget, I designed a modular template, assembled from parts search already used. A core of product cards could include optional slots for a brand header, video, an article snippet, or a "sponsored" label. Any slot could turn on or off depending on the content.
Design a bespoke, one-off widget tailored only to our first hypothesis.
We expected third-party expert reviews to be the winner, which needed heavy editorial framing and media slots.
A modular template.
Whatever content won, the template could carry it without a redesign.
Preview: desktop, grid view, showing title, attribution, image, product cards.
The modular slots
- A B
Title and attribution
Optional header to attribute third-party publishers or frame the recommendation source.
- C
Disclosure label
Optional slot for a "sponsored" or editorial label, turned on or off depending on the content.
- D E F
Image, video, and article
Toggleable slots for an image, inline video playback, or expandable editorial excerpts.
- G
Core product cards
Assembled from parts search already used so shoppers could scan prices, ratings, and Prime status instantly.
Decision 02 of 08
Designed the expert version in detail
Give the first hypothesis every chance to win before letting the data judge it.
The first content was curated picks from sources like Wirecutter and America’s Test Kitchen. How do you show a publisher’s brand inside Amazon without it reading as a takeover? Should video autoplay? Does an article snippet open in place or on a new page? I worked through each one, because this was the version we expected to ship.
Treat the first version as a rough test just to validate placement.
With a modular system, it’s easy to under-design the first content model and blame the format if it fails.
Work through every branding, video, and snippet interaction in detail.
If the experts lost, it had to be the idea that lost, not an unfinished design.
Decision 03 of 08
Traded a good result for a much better one
Follow the data when it proves your core bet right, but your favorite idea wrong.
The expert picks tested well. Then we tested crowd-sourced picks (best sellers, trending, "customers also viewed"), and they won by a wide margin. They also scaled to any search with no one curating, so we switched. Every test after that came back bigger.
Shoppers wanted someone to trust.
The expert picks proved that recommendations inside the results helped people choose.
About who they trusted.
They trusted other shoppers. Next to the crowd picks, the experts read like ads.
Decision 04 of 08
Made it look like a result
When trust is the mechanism, the craft is making the design disappear.
The rich-media versions looked like something special, and that was the problem. If trust was the mechanism, anything that looked different looked like an ad. So I switched most of the slots off, and the component showed exactly what an organic result shows.
Ship the version that looked special.
We’d designed it in detail and expected it to win.
Switch most of the slots off.
Because the template was built to flex, that cost nothing.
Decision 05 of 08
Let the algorithm decide where it goes
Don’t argue over page placement when shoppers can rank it in real time.
The obvious move was to pick a spot: top, middle, or bottom. I tested those, but I also tested a "king of the hill" model that let the component float, ranking itself the same way search ranks products. If it was selling better than the result above it, it moved up. If it wasn’t, it moved down.
Floating won. On broad searches like "headphones," it rose near the top. On searches for one exact product, it sank out of the way.
Pick a fixed spot.
Every product team wants guaranteed placement, regardless of what the shopper is searching for.
Let the component float.
No one had to decide where it belonged for each search, because the shoppers decided.
Decision 06 of 08
Gave engineering a shape, not a spec
Steer a remote engineering team with a clear visual target, not a rigid handoff.
An engineering team in Cambridge built the data underneath, testing dozens of recommendation sets per search to find the winners. That was their work, and I didn’t manage them.
A team testing a knowledge graph with no shared target for the interface.
A rigid spec would have slowed down their data experiments.
Steer with the six-page proposal and monthly visits with design explorations.
Once they could see the template’s fields, they knew what data it needed.
Decision 07 of 08
Fit it to every category
Win over skeptical category teams with live test results, not slide decks.
Clothing uses a grid view, electronics a list, and a knife shows different details than a laptop. I worked through dozens of category teams so the component felt native in each.
Force one layout across Amazon, or skip categories that pushed back.
Fashion pushed back hardest, because it wanted clean grids, not widgets.
Test where teams were skeptical.
We tested in Fashion’s categories anyway, and the results brought them around.
Decision 08 of 08
Kept ads out
Protect the organic trust that makes the result possible.
Leadership wanted sponsored placements from early on. I pushed back, because the result rested on these feeling organic. The template had a slot for a sponsored label. While I owned it, that slot stayed off.
Allow sponsored placements early on.
Leadership wanted to monetize the new unit from the start.
Keep the component 100% organic.
The expert picks had already shown what happens to anything that reads like an ad.
Results
Impact of the work
What changed in the business
$1B+ in incremental annual sales. Based on two-week, statistically significant A/B tests. It was my biggest win at Amazon.
On about 80% of search results pages by the time I left in 2019.
What changed in how search worked
A new kind of unit in the organic results. Search results had been reserved for products. This opened them up to something else, for better or worse.
Two US patents. The modular inline recommendation architecture and search discovery work established a new pattern for how e-commerce search results work, resulting in two granted US patents (2017 and 2019).
What survives me
The modular component and the floating placement model. A version of it is still on most Amazon searches today, even though sponsored placements eventually entered the picture after I left.
What I’d do again
Start with the six-page proposal and build the component to flex from day one. Pivoting from rich-media expert picks to stripped-down crowd picks cost nothing in engineering rework.
What I wouldn’t
I’d still design the expert version properly, but I’d test a stripped-down crowd version beside it from the start. That would have gotten us to the winner faster.
