Shopify seller tries AI product selection, doubles hit rate in three months — nearly killed the business in the first six weeks
Mark has run a Shopify store for three years — home goods niche, about $1.2M annual revenue.
Not bad. But he told me: "Product selection is gut and doomscrolling TikTok. Two hits out of ten is a good week."
Late 2025, he added an AI selection tool. Logic: scrape trend signals from TikTok/Instagram/Pinterest + Amazon review keywords + Google Trends, then recommend products about to break out.
Hit rate climbed from 12% to 31%. Sounds great, right? But the first six weeks almost shut him down.
Pit one: cold start picked a seasonal dud
Week one, AI pushed an "outdoor portable projector" — TikTok searches up three weeks straight, Pinterest saves doubling.
Mark ordered 600 units at $89 each. Over $50K in inventory.
Problem: it was the tail of the summer-outdoor trend. By the time stock landed, it was September. Temperatures dropped, searches cratered. Sold under 80 units in three months. 500+ units collecting dust in the warehouse.
AI read three weeks of trailing data but didn't know the product was seasonal. Mark didn't ask. Neither side had that string tuned.
Pit two: AI price war to negative margin
Worse. The AI had a dynamic pricing feature — auto-adjusts to competitor prices.
One day a competitor ran a promo. AI matched it. Mark's scented lamp cost $14; AI dropped to $13.20 — selling at a loss. He slept eight hours, woke up to 400+ orders, $300+ in the red.
The AI's logic wasn't wrong: its goal was "protect cart conversion." Conversion did go up. Every sale lost money.
Pit three: patent infringement notice
AI pushed a "magnetic phone stand" that looked almost identical to a major brand's registered design patent. Mark sold 300 units in two weeks, then got a cease-and-desist.
Refunds + delisting — another ten grand gone.
The fix
- Cold-start screening: every AI recommendation passes two filters — Google Trends 12-month curve to rule out seasonal spikes; Amazon BSR history to confirm steady growth, not a pulse
- Dynamic pricing gets a floor: each SKU has a minimum margin line (no lower than cost × 1.3). AI can go up, never below
- Patent pre-check: before listing, run Google Patents for design similarity. Over 70% similarity, auto-reject
Three months later
| Metric | Before AI | After AI |
|---|---|---|
| Product hit rate | 12% | 31% |
| Inventory turnover | 3x/year | 5.5x/year |
| Ad ROAS | 2.1 | 3.8 |
| Select-to-list cycle | 2 weeks | 4 days |
Hit rate more than doubled. Turnover from 3x to 5.5x means half the cash tied up in inventory. ROAS from 2.1 to 3.8 — every ad dollar returns nearly twice as much. Total cost around $12K (monthly subscription + setup), 5-month payback.
Mark's two lines
"AI selection's biggest value isn't finding hits. It's keeping you from stocking duds. One bad pick ties up $50K — dodge two and you've paid for the tool."
"Dynamic pricing needs a floor. It optimizes for conversion, not profit. Leave it unsupervised and it'll happily run you into the ground."
The real value of AI in e-commerce isn't "picking winners" — it's "not picking losers." The cost of a bad pick is dead inventory, thousands gone in a blink. AI earns its keep by using data to block products that look pretty but hide pitfalls. But you've got to set the rules — check seasonality, check patents, set price floors. AI isn't dumb. It just doesn't know where your line is unless you draw it.