AI Powered Content at Scale
Led design through three pivots (education, batch Files generation, then GenAI on upload), raising accessible metadata coverage 500%+ while cutting merchant time on image fields by 80%.
Product Design, Accessibility
GenAI alt text at upload for 2M+ merchants: Shopify's first AI feature in admin, with the merchant edit always in the loop.
- Role
- Product Design Lead
- Team
- Tech lead, PM, UX researcher, content designer
- Timeframe
- 6 to 9 months · three pivots
- What I owned
- The three-pivot strategy, Files and product-media UX, model QA, and the accessibility case for keeping override visible.
GenAI metadata at upload for 2M+ merchants.
Three pivots: teach merchants to write alt text, batch-generate in Files, then generate on upload. Close the metadata gap without backfilling catalogs behind merchants.
The gap
Strong photos, empty alt text.
Fewer than 15% of merchants filled in alt text, keywords, or descriptions. In product media and Content → Files, the default was an empty field beside a sharp product photo. An industry pattern, not a Shopify one: Baymard found 55% of e-commerce sites fail accessibility on informational images1.
The stakes
Merchants, SEO, and Shopify's own data quality all paid for the empty field.
Note
- Shoppers using screen readers hit unlabeled images; image search lost the context to rank them
- Roughly 80% of assets carried no metadata, starving analytics and future AI features
- Partner DAM integrations stalled where catalogs had nothing to sync
The squad
One squad, from research to ship.
Note
Design lead, tech lead, PM, UX research, content design. Six to nine months, first pitch through model QA and an Enterprise soft launch.
First pivot
Teaching merchants to write it did not move coverage.
Cut from the build · Educate on the why and the how
What we tried.
Training decks and a Shopify blog guide on writing alt text. The bet: understand the value, fill the field.
Why it lost.
Merchants wanted the field filled, not a course. The blocker was time, not knowledge.
Field research
75% blamed time. GenAI made generating at upload realistic.
- 75%
- Cited time as the blocker
- 60%
- Did not know alt text aids SEO
- 45%
- Managed images in a third-party DAM
Second pivot
Bulk Generate alt text in Content → Files.
Select assets, generate, then review row by row. Closer to the work, but still a step after upload.
01 · Files, mostly empty
The default Content → Files list: an Add alt text placeholder on most assets.
Mobile
The same batch flow, built for thumb reach.
List
Select
Generate
Model QA
A dozen models, then an Enterprise soft launch.
We scored roughly a dozen models on real merchant uploads, tuned on a merchant batch, and validated the override flow with Enterprise before rollout. The bar was high: a confident wrong description misleads a screen-reader user in ways an empty field does not2. Editing had to stay one tap away.
Third pivot
Product signals feed the first draft.
By default
An AI draft by default, on desktop and mobile.
In the loop
Suggestions to weigh, applied only on save.
Field states
Field states, up close.
Results
500%+ coverage. 80% less time. The first GenAI in admin.
Close
Generation belongs in the workflow.
GenAI earned its place once it generated at upload and kept override in plain sight. Not in a training deck, not in an overnight backfill the merchant never saw.
Two more months
Every new upload got alt text. The catalog behind it, years of images, stayed empty. A batched re-review queue was scoped, then slipped under other priorities. Merchants asked for it in every follow-up, and they were right. Shipping the new path was the easy half; the backlog is the work I would finish.
- Tidjane
Shipped with
Sources & notes
- 1
Baymard Institute usability research: 55% of e-commerce sites fail accessibility on informational images. Internal admin audit and partner DAM feedback (2023) corroborate the gap on Shopify.
baymard.com
- 2
WebAIM holds alt text to a strict accuracy bar: it must be an equivalent for the image, since a wrong description still reads as fact to a screen reader.
webaim.org
- 3
Directional program-level metrics from the 90-day rollout window. Time-saved figure from merchant follow-up surveys.
More work
Other projects worth a look.
