Example · Case study
Northstar's edge was expert advice locked in chat logs and PDFs. AI search couldn't cite it. We turned the buying logic into structured criteria and compatibility data agents could quote.
Northstar is a specialty outdoor retailer for ski touring, climbing, and trail — staff averaging eight years in the sports they sell. Online, that expertise lived in chat logs and a few buying guides.
In early 2026, ChatGPT Shopping and Perplexity almost never cited Northstar, even on queries it was perfectly placed to win. Competitors with thinner ranges but cleaner data took the recommendations.
We tested 84 real queries against ChatGPT Shopping, Perplexity, and Claude. Three breaks covered 78% of the misses:
Staff recommended in ninety seconds, but the criteria they used were written down nowhere an agent could read.
Which binding fit which ski geometry sat in attachments, not structured data. Agents wouldn't risk the recommendation.
'Seasonal' and 'in production' didn't tell an agent whether to wait or suggest an alternative — so it suggested the alternative.
"AI search now cites us where competitors used to be. The recommendation logic became part of the product."
Four workstreams over 8 weeks. Turn the advice into data — without flattening what made it good.
We turned the recurring advice patterns from chat and sales into explicit if-then criteria — rider weight, experience, use case → recommended setup — surfaced as 'Best for' and 'Avoid if' on the product page.
Tactic
Pattern extraction from 600 support logs into versioned decision rules.
Compatibility moved from PDF to structured data: this binding fits these ski geometries, this boot sole fits these crampons. Agents could finally make the call without a disclaimer.
Tactic
Schema.org relations across the high-volume pairs, validated with a headless agent crawl.
{ "@context": "https://schema.org", "@type": "Product", "name": "Backcountry Touring Binding", // the decision criteria, made explicit "audience": { "@type": "PeopleAudience", "audienceType": "ski touring, 0-5 days experience" }, "isAccessoryOrSparePartFor": [ { "@type": "Product", "name": "88mm Touring Ski" }, { "@type": "Product", "name": "94mm Touring Ski" } ]}We replaced 'seasonal' with restock dates an agent can reason about: wait, or recommend the in-stock alternative. Either way it can act.
Tactic
Inventory status in clear buckets with explicit dates, synced from the warehouse every 4 hours via Make.com.
For the top recommendation questions, we built comparison surfaces with the criteria and trade-offs made explicit — the thing AI search can actually quote.
Tactic
Comparison pages as structured ItemList with recommended-for properties, plus a weekly coverage check.
Three headline numbers — and the main one breaks down.
+74% answerable product questions
The 74% broke down into four levers we measured separately:
The asset they kept is the criteria. The buying logic is now explicit and structured, so it works for AI search, for on-site comparison, and for the next hire — instead of leaving with whoever knew it. A weekly coverage test keeps new products from lowering it.
Example · Case study
Example · Case study
Want the same on your shop?
Check my shop