The ranking doesn't just look at what you type. It looks at what customers actually buy.
When a customer searches for something in chat, Aisly doesn't just match keywords against your catalog: it also weighs what real visitors to your store have actually added to their cart. Below is the real formula, and a simulation of what it can do once behavioral data starts coming in.
The ranking corrects itself
25 simulated days · 40 visitors/day · top 4 positions shown each day
No rewrite for this page
It's the same SQL function running in production, not a simplified version
similarity is the real cosine similarity between the product's embedding and the search's (pgvector). Bayesian smoothing — a weak prior, 1 cart add per 5 "imaginary" impressions — keeps a product with no history from sitting at zero, and one with very little data from dominating on statistical noise. It runs in sql/002_personalized_ranking.sql, the same function that powers the advanced analytics dashboard where you see what actually converts.
What we assumed for the scenario
For every product: similarity is real. The "true" add-to-cart rate is a declared assumption — not observed data — chosen to represent a common e-commerce phenomenon: a product that's mid-pack by semantic similarity can convert far more (or less) than the text alone would suggest.
| Product | Real similarity | Assumed conversion rate |
|---|---|---|
| The Multi-managed Snowboard | 0.9450 | 4.0% |
| The Multi-location Snowboard | 0.9387 | 3.5% |
| The 3p Fulfilled Snowboard | 0.9205 | 1.0% |
| The Out of Stock Snowboard | 0.9077 | 0.5% |
| The Collection Snowboard: Oxygen | 0.9030 | 3.0% |
| The Compare at Price Snowboard | 0.9000 | 2.0% |
| The Inventory Not Tracked Snowboard | 0.8923 | 2.0% |
| The Collection Snowboard: Hydrogen | 0.8919 | 9.0% |
| The Collection Snowboard: Liquid | 0.8906 | 3.0% |
| The Videographer Snowboard | 0.8362 | 1.5% |
| Selling Plans Ski Wax | 0.7952 | 0.5% |
| Gift Card | 0.6941 | 0.2% |
Day-by-day log
What the engine would have shown in 1st place that day, and the cumulative count under both scenarios
| Day | 1st place (adaptive) | Hydrogen rank | Adds/day adaptive | Adds/day static | Cum. adaptive | Cum. static |
|---|---|---|---|---|---|---|
| 1 | The Multi-managed Snowboard | 8 | 5 | 3 | 5 | 3 |
| 2 | The Collection Snowboard: Oxygen | 4 | 8 | 7 | 13 | 10 |
| 3 | The Collection Snowboard: Liquid | 3 | 2 | 2 | 15 | 12 |
| 4 | The Multi-location Snowboard | 3 | 7 | 6 | 22 | 18 |
| 5 | The Multi-managed Snowboard | 2 | 6 | 2 | 28 | 20 |
| 6 | The Multi-managed Snowboard | 2 | 6 | 1 | 34 | 21 |
| 7 | The Collection Snowboard: Hydrogen | 1 | 9 | 6 | 43 | 27 |
| 8 | The Multi-managed Snowboard | 2 | 3 | 4 | 46 | 31 |
| 9 | The Multi-managed Snowboard | 2 | 11 | 6 | 57 | 37 |
| 10 | The Collection Snowboard: Hydrogen | 1 | 7 | 1 | 64 | 38 |
| 11 | The Collection Snowboard: Hydrogen | 1 | 8 | 4 | 72 | 42 |
| 12 | The Collection Snowboard: Hydrogen | 1 | 3 | 2 | 75 | 44 |
| 13 | The Collection Snowboard: Hydrogen | 1 | 8 | 3 | 83 | 47 |
| 14 | The Collection Snowboard: Hydrogen | 1 | 6 | 3 | 89 | 50 |
| 15 | The Collection Snowboard: Hydrogen | 1 | 8 | 3 | 97 | 53 |
| 16 | The Collection Snowboard: Hydrogen | 1 | 5 | 6 | 102 | 59 |
| 17 | The Multi-managed Snowboard | 2 | 6 | 1 | 108 | 60 |
| 18 | The Multi-managed Snowboard | 2 | 2 | 7 | 110 | 67 |
| 19 | The Multi-managed Snowboard | 2 | 5 | 3 | 115 | 70 |
| 20 | The Multi-managed Snowboard | 2 | 3 | 2 | 118 | 72 |
| 21 | The Multi-managed Snowboard | 2 | 3 | 2 | 121 | 74 |
| 22 | The Multi-managed Snowboard | 2 | 5 | 3 | 126 | 77 |
| 23 | The Multi-managed Snowboard | 2 | 7 | 5 | 133 | 82 |
| 24 | The Multi-managed Snowboard | 2 | 6 | 3 | 139 | 85 |
| 25 | The Multi-managed Snowboard | 2 | 7 | 3 | 146 | 88 |
What this shows, and what it doesn't. It shows that the real scoring function, without any modification, shifts the ranking toward what actually converts as behavioral events come in — with genuine similarity and conversion assumptions stated in full above.
It doesn't show a result for your store: traffic, catalog, and actual customer behavior vary, and none of the numbers above are a performance promise.