We Loaded 200K Products Into AI Memory. Conversion Jumped to 40%.
Former Co-Founder of Nansen.ai ($80M+ raised from a16z, Accel, Tiger Global)
When we analyzed pre-purchase calls to our AI voice agent, product questions dominated. Not "do you have this in stock?" questions — consultative conversations spanning 6-8 decision points. The kind that turn browsers into buyers.
The numbers: 61% of pre-purchase calls are product questions. They convert at 40%.
That 40% didn't happen on day one. It took a fundamental shift in how AI product knowledge works. Two and a half months after publishing the original measurement, the 40% has held steady — and the harder question turned out to be how the AI uses that knowledge in real time, not how much of it gets loaded.
The Breakdown
Pre-purchase calls make up about half of all inbound calls. When we categorized what customers were actually asking:
- Product questions: 61%
- Insurance questions: 29% (covered here)
- Issues: 10%
Product questions convert at 40% — the highest of any pre-purchase category. But we had to earn that number.
The Problem: Search Wasn't Enough
First attempt: 25% conversion rate.
The AI searched products via the website API. Response time averaged 4 seconds per query. The AI was limited by search logic — it could find products, but couldn't truly understand them.
Customers asked nuanced questions. AI gave generic answers.
"What's the difference between 52mm and 54mm lens size?"
"Which frame works for a 4-year-old?"
"Is this the right bridge width for my face?"
The AI could locate products. It couldn't educate customers about them.
The Fix: Full Product Knowledge
We loaded 200,000 products with complete details directly into AI memory. Every spec, every variant, every sizing detail — organized around Shopify's standard product taxonomy so categories and attributes were consistent across the catalog.
Response time dropped to under 1 second. More importantly, the AI could now think about products — not just search for them.
Result: 25% → 40% conversion rate.
The difference wasn't speed alone. It was depth. The AI became a product expert who could guide customers through complex decisions.
What Consultative Product Conversations Sound Like
Here's a real conversation. Customer looking at Gucci GG0131O eyeglasses:
Customer: "I don't know what lens size means."
AI Agent: "No problem. The lens size is simply how wide each lens is — 52mm is slightly narrower; 54mm is a bit wider. If your current glasses feel a touch small, go 54; if you prefer a tighter fit, 52. Would you like me to suggest a best-fit based on average sizing, or keep it compact with 52mm?"
Customer: "Keep it compact."
AI Agent: "Compact fit — 52mm, noted."
The AI then walked through lens type (single-vision vs progressive), material (polycarbonate vs high-index), and coating levels (standard vs premium anti-glare). Each answer led to the next question. Each question moved closer to purchase.
Customer converted.
More Than Specs: Real Product Guidance
Another example — a parent buying kids' glasses:
Customer: "I just wanna make sure those glasses will fit my toddler. She's four years old."
AI Agent: "For a 4-year-old, the Dilli Dalli Tutti-Frutti 44-13-130 is a good toddler fit: Lens 44mm, bridge 13mm, temple 130mm — typical for ages 3-5. Flexible, kid-safe material; works well with the strap for secure wear. Prefer a smaller fit? We also have the 42-13-125."
Customer: "I'll stay with the 44-13-130. If that one works for her age, I'll give it a try."
AI Agent: "Great choice. Frames have a 90-day money-back guarantee with no restocking fee."
Customer: "Thank you so much."
The AI didn't just confirm stock. It validated the parent's choice, offered an alternative, and removed purchase anxiety with the return policy. That's consultative selling.
Pattern: Answer, Then Advance
Every successful product conversation follows the same structure:
- Customer asks a question
- AI answers with specifics (not "it depends")
- AI offers a choice or recommendation
- Customer decides
- AI advances to next decision point
Here's another real example — customer checking stock on Ferrari sunglasses:
Customer: "Confirming an item is in stock. Ferrari FH1017, men's blue."
AI Agent: "Ferrari FH1017 Blue/Blue Mirror for men. We have it in stock at $464.95. Would you like to order it now online, or hear a quick size note before checkout?"
Customer: "Hear a quick size note."
The AI didn't just confirm "yes, in stock." It anticipated the next question and offered value. That's initiative capture — answering what was asked, then guiding forward.
The Revenue Math
Let's scale this:
- 100 pre-purchase calls per month
- 61% are product questions = 61 product calls
- 40% conversion rate = 24 orders
- $350 average order value = $8,400/month
Annual revenue: $100,800 from product knowledge conversations.
And 75% of these calls happen after hours — when most stores can't answer them.
What We Got Wrong The First Time
When we first published this analysis, we framed the upgrade as "load more product data." That was half right. Loading the full catalog into memory was necessary, but it wasn't sufficient.
The real lift came from how the AI retrieves product context during a call — not how much it has access to. Three things mattered more than the catalog size:
- Recency of retrieval. The AI re-reads relevant product data at the point of each customer turn, not just at call start. A customer can change subject from frames to lenses to insurance and back without the AI losing context.
- Variant-level granularity. "Do you have this in 54mm?" requires the AI to know not just the product but every variant SKU. Catalog feeds that flatten variants into a single product break this entirely.
- Knowing what it doesn't know. The AI is now better at saying "let me check" or transferring to a human when a question falls outside the catalog — instead of confidently inventing a spec. This is the single biggest driver of the 4.8/5 satisfaction rating.
The pattern transfers: any vertical with deep variant complexity (eyewear, jewelry, custom furniture, specialty footwear) gets the same lift. Verticals with flat catalogs (apparel basics, accessories) see smaller gains because the consultative conversation isn't there to begin with.
Why Product Questions Convert
Product questions aren't about finding products. Customers can browse your website for that.
They're about understanding products:
- What lens terminology means
- What options apply to their situation
- What the right choice is for them
A customer who calls with product questions has already narrowed down their options. They're not browsing — they're deciding. They need someone to confirm they're making the right choice.
AI loaded with deep product knowledge does exactly that. Without judgment. With patience. At 2am on a Sunday.
That's consultative selling at scale.
We break down every product question in detail: The 10 Eyewear Questions Customers Ask Before Buying. Product questions aren't the only high-conversion category — stuck customers convert at 60%. And for the "why voice outperforms text" math, we compared AI phone support vs chatbots. See how the full stack comes together on our how it works page or estimate impact with the ROI calculator.