Storefront Builder

Context

Yandex.Market is one of Russia’s largest e-commerce platforms: 18 million active buyers, 80,000 sellers, more than 80 million SKUs.

At the time, only our largest enterprise clients could customise their shop pages, built manually through a clunky internal CMS that nobody wanted to touch. Research showed the demand shifting: mid-sized and small sellers were building their own brands and driving traffic to their own websites, and competitors like Ozon and AliExpress were already offering storefront tools for SMEs.

We needed to act fast and to design a self-service solution for SMEs that was affordable, scalable, and kept sellers inside our ecosystem instead of losing them to external platforms.

Challenge

How do you empower thousands of sellers to create branded, high-quality storefronts — fast, at scale, and without rebuilding the entire tech stack or hiring a content team to review everything they publish?

Initial metrics

Adoption: maintained storefronts, sellers who built one and kept it alive.

Efficiency was the kill criterion: if this wasn't dramatically cheaper than the manual baseline, the whole initiative failed.

Retention and support escalations were there to catch experience problems before they became platform trust problems.

GMV made the list but couldn't be isolated cleanly as storefront traffic overlaps with other acquisition channels, so we tracked it directionally, not as a KPI.

Research

2 tracks of benchmarking: competitive (Ozon, Wildberries, Tilda) and inspirational (Readymag, Framer, Shopify), plus the external websites our sellers already ran. The pattern was consistent — hero banner, themed carousels, simple navigation and it separated the non-negotiable blocks from the nice-to-haves.

Seller interviews were built around one question: what job is your external website doing that Yandex.Market isn’t? Sellers wanted to tell a brand story, not just list products. First impressions mattered, and they had no control over them.

The most useful research: our own CMS

Then I audited the internal CMS and found every component sellers cared about already built: hero banners, carousels, groups. No new blocks needed. The job was a simpler interface on top of what existed, and that finding saved weeks of engineering work.

Constraints that shaped the design: first attempt

I put together a first version of the IA — Storefront Builder as a simplified UI on top of our CMS, publishing through a microservice to the Partner app. Technically clean. Straightforward to build.

Then came the feedback that stopped everything: moderation. Anything sellers published needed review before going live — illegal material, misleading offers, trademark violations. No content team existed, hiring one would price the product out of the SME market, and AI moderation wasn't reliable enough to run unsupervised.

Redesigned around the constraint

Then I looked at Toloka, Yandex’s own crowdsourcing platform for micro-tasks, similar to Amazon Mechanical Turk. If the moderation workflow could run on Toloka, we’d get human review at a fraction of the cost of a permanent team. I redesigned the architecture around three parts: the microservice as orchestrator, the Builder as the seller-facing editor, and Toloka reviewing visual content before anything went live. Confirmed feasibility with the dev team, got sign-off, moved on.

I build the dataset to train crowdworkers

To train Toloka moderators, I built a dataset of ~200 banner examples: 100 approved, 100 rejected, covering realistic violations — misleading offers, illegal content, low-quality visuals. Approved examples came from our real B2C Figma files.

The rejection cases we brainstormed as a team, which became an unexpectedly enjoyable session.

My estimate, from comparable image-review tasks: ~4 hours per storefront. The actual number: 10 minutes. Disputed decisions were near zero. No full-time moderators, ever.

The pipeline outlasted me: the crowdsourced model held until AI moderation matured, starting from the dataset and rejection rules I built — Yandex has since publicly described automating banner review with neural networks.

Designed UI of storefront builder

The goal was a tool sellers could open and use without reading anything first. I designed a lightweight WYSIWYG editor: a control header for key actions, an interactive workspace with real-time preview. Sellers assembled pages visually; the system generated a JSON configuration behind the scenes, sent it to the CMS, and triggered a moderation task automatically.

Every interaction was validated with engineers before it became a design decision — what triggers a request, how long processing takes, what to show during loading, how to handle connection drops. These conversations shaped the UX in ways that desk research couldn't have.

I ran around a dozen usability sessions with sellers before launch. Feedback was largely positive. One consistent request: short descriptions for each content block. I added them. One seller, mid-session, even asked: «Are you going to raise our fees now that it looks this good?» That's the bar I was aiming for.

Results

The tiles tell most of it. What they don't show: zero full-time moderators hired — the Toloka model held, with near-zero disputed decisions.

Storefronts went from a luxury agency-style work, weeks of waiting, hundreds of thousands of rubles to a free default for every seller.

Takeaways

Next time, let the machine do the boring part. I built 200 moderation examples by hand — half approved, half rejected — brainstorming fake violations with the team like a very niche game show. It worked. It was also a terrible use of designer time. A generative model could do that in an afternoon now. I'd use it.

Design for the ecosystem on day one. Sellers were building moderated, ready-to-go banners inside the Builder. One obvious next step: push that same banner into Yandex.Direct, Search, Maps, 85M+ users. I saw it. I moved too slowly on pulling in the right teams, and the window closed. Cross-surface distribution belongs in the initial IA, not the roadmap backlog.

Templates beat blank canvas. My hunch: give sellers a pre-assembled layout instead of individual blocks, and creation time drops to two or three clicks. Adoption follows. I had the designs. The project was cut short before I could test it. I still think I was right — and the retention data would have told us within weeks.