Read the market
A 30-day AI visibility experiment, documented
By Ankit Minocha, founding team at Atomz. Updated January 21, 2026.
We took one small Shopify catalog from near-invisible to regularly cited in AI shopping over thirty days. Here is the method, the week-by-week movement, and what worked.
It is one thing to argue that a readable catalog earns AI citations and another to watch it happen on a clock. So we ran a controlled thirty-day experiment on a small skincare catalog of around a dozen products, starting from the position most stores are in, almost never appearing when a shopper asked an assistant for what it sold, and changed only the catalog layer. The point was not the brand, it was to see how fast the structure work moves the result, and in what order.
The starting point
At day zero the catalog behaved like a typical store. Across a fixed set of purchase-intent questions run through the major assistants, it was cited rarely, and when it appeared at all it was for direct brand-name questions rather than for the situations a real shopper describes. The products were good, but the fields a query needed were missing.
What we changed
We changed nothing about the products, the prices, or the design. We mapped every product to the Shopify taxonomy, wrote skin type, concern, ingredient, finish, and free-from attributes to the metafields, gave each product a line of real context about what it is for, made sure the pages rendered that content server-side, and allowed the AI crawlers. That was the whole intervention.
The movement, week by week
| Checkpoint | Citation rate across the question set |
|---|---|
| Day 0 (baseline) | Near zero |
| Day 15 (attributes written, crawlers allowed) | Roughly a fifth of questions |
| Day 30 (full structure, indexed) | Over half of questions |
The shape matters more than the exact figures. The first jump came once the attributes existed and the crawlers were let in, and the second came as the assistants re-read the store and began matching the now-structured products to specific questions. The specific, attribute-heavy questions, the ones that had returned nothing at the start, were where the gains concentrated.
What worked and what did not
The changes that moved the result were unglamorous: structured attributes, server-rendered content, and crawler access. The things that did not move it were the things stores usually reach for first, namely rewriting marketing copy and adding more of it. Better adjectives did nothing, because the assistant was never reading for adjectives.
The honest caveat
This was a small catalog, and your numbers will depend on catalog size, category, and how unstructured you start. The durable lesson is the order: attributes and access first, and the movement follows within weeks, not quarters.
If you want your own baseline before you start a thirty-day clock, the audit is the day-zero measurement, scoring how much of your catalog an assistant can read right now.
Subscribe to The Agentic Operator