Piece II · Trove Recommend · Issue №07

Trove Recommend.

Recommendations modelled on real buyer reasoning, not raw engagement logs. Click logs are deliberately under-weighted. Click-through rate moves slightly down; units-per-transaction, return rate and lifetime value move materially in the right direction.

Apply for an atelier slot Read the method

Architecture

Three training signals. None of them are click logs.

A recommendation system that optimises for clicks will recommend things people click on. That is a tautology, and it is — for a boutique whose customer cares how things were found — a small disaster. Trove Recommend trains on three other signals, each chosen because it correlates better with post-purchase satisfaction.

Signal 01

Curator-labelled reference sets

For each shop, the buyer labels reference sets — pairs and triples of pieces that go together, pieces that emphatically do not, edge cases. The reference set is small (typically three to four hundred items) but rich; the embedding model is grounded against it.

Signal 02

Post-purchase satisfaction

Captured through a one-question follow-up two weeks after delivery: "would you recommend this piece to a friend with the same taste?" Categorical, low-friction, high-signal. The model learns that satisfaction, not the headline conversion, is the optimisation target.

Signal 03

Explicit affinity statements

Where the customer has told you what they like — wishlists, save-for-later collections, requests-to-the-buyer — we treat it as ground truth. These signals are scarce but precise; the recall improvement on the right tail of the catalogue is significant.

What changes

The metrics move in different directions. The right ones move up.

Metric	Direction	Pilot average (mo. 1–3)	Pilot average (mo. 4–12)
Click-through rate	↓	−4.6%	−2.1%
Units per transaction	↑	+3.8%	+11.4%
Average order value	↑	+5.2%	+14.7%
Return rate	↓	−4.1%	−7.8%
Repeat-purchase rate (90-day)	↑	+2.1%	+9.6%
12-month lifetime value	↑	+6.0%	+18.3%

Pilot averages across the first eleven boutiques in the atelier programme. Performance varies by catalogue shape; the directionality is consistent.

The dopamine charts go down before they go up. Operators who can stomach the early dip see the curve invert by month four. The ones who panic and revert to engagement-optimisation see neither curve.

How it integrates

Sits behind your storefront, not in front of it.

Storefront

Shopify, BigCommerce, custom platforms

Drop-in JavaScript widget for the major platforms; REST and GraphQL endpoints for custom storefronts. Latency budget under 120 ms p95 for the embedding fetch; the rest is layout, which you control.

Data flow

Read-only catalogue feed, write-back telemetry

We read the catalogue (SKU, description, taxonomy, imagery) on a fifteen-minute cadence. Telemetry — placement, scroll, click, purchase, return, satisfaction — flows back, hashed and aggregated where appropriate. No customer PII leaves your environment.

Buyer interface

Reference set editor; rejection log

The buyer can mark recommendations as wrong, right, or "yes but not this customer." Every rejection updates the local model overnight; the reference set is your editorial fingerprint and lives with you.

Operations

Weekly calibration review

Every Monday morning the calibration deck is in your inbox. Click vs. satisfaction divergence. Top right-tail pieces. Pieces being recommended too often. Pieces being missed. A real person at our end has read it before it lands.

Apply for an atelier slot