Solutions / Recommendation systems

Personalization that moves revenue

Q: How long does it take to see a lift?

Most clients see measurable A/B results within 6–8 weeks of going live. The first two weeks cover data integration and model warm-up; weeks three to six run the 50/50 holdout. We don't call significance early. The gap has to clear p<0.01 before we ship variant B to everyone.

Q: What data do you need to get started?

At minimum: an event stream (views, add-to-cart, purchases) and a product catalogue. Richer signals like search queries, loyalty data and in-store events improve accuracy but aren't required for a first model. We handle cold-start for new users and new items from day one.

Q: How do you handle cold-start for new users and new products?

New users get a popularity- and context-based fallback that personalises within the first session using implicit signals (category, dwell, device). New products are bootstrapped with content features and promoted into the ranking as early signals accumulate, typically within 48 hours of first impressions.

Q: Can the recommendation system run on-prem or in our cloud?

Yes. We deploy into your cloud (AWS, Azure, GCP) or on-prem Kubernetes. Inference latency sits at p95 <50 ms in all environments. Data never leaves your perimeter. Model artefacts, training pipelines and serving infrastructure all run inside your boundary.

Q: How do you measure success beyond conversion?

We track conversion, average order value, click-through rate and revenue per user as primary KPIs. Secondary metrics include catalogue coverage (to avoid popularity bias), repeat-visit rate and, for loyalty apps, redemption rate. All metrics are reported per cohort against a held-out control.

Q: What happens if the A/B doesn't reach significance?

We don't ship a variant that doesn't clear the bar. Instead we diagnose: check for segment heterogeneity, adjust the model or the ranking policy, and re-run. A null result is information. It tells us where to iterate next. Our contracts include iteration cycles precisely because we know real lift requires honest experimentation.

Recommendations that lift conversion, basket size and click-through across web, app and messaging, proven in A/B on all users

Collaborative filtering Neural LTR Cold-start handled A/B proven

9:41

Personalizationrecommendation engine p95 42 ms

E-commerce / Retail · in production

sess_9f3areturning shopper

retrievaltwo-tower · ANN 1,240 / 84k

rankingGBDT · neural LTR scoring…

re-rankdiversity λ=0.3 · dedup final top-6

online A/B · variant vs control +25% conv

added Merino socks + insoles stat-sig · on all users

iPhone running Zentavor Personalization engine

+25%conversion, A/B on all users

4×click-through rate

+18%average order value

What we do

One storefront for everyone, or a store of one

A shopper opens the app to a cold, generic feed. Behind the screen the engine reads what this shopper actually did, finds the shoppers most like them, and sends back a ranked, personal shelf. That is collaborative filtering, told end to end

This shopper

What they actually did

READING HISTORY

Bought running shoes

Browsed outdoor jackets

Added a yoga mat to cart

Average order value

Loyal customer

Shoppers like them

Ranked by co-buy · top 4

1,240 LOOKALIKES

Helio Jacket

82%

co-buy

Shipped to feed

Aurora Runner

76%

co-buy

Shipped to feed

Nimbus Buds

69%

co-buy

Shipped to feed

Trailmag 30L

64%

co-buy

Shipped to feed

Trail Cap

31%

co-buy

Socks 3-pack

22%

co-buy

0 of 4 ranked picks shipped to the feed

9:41

Camera

Messages

Maps

Music

Photos

Weather

Clock

Notes

Mail

Health

Calendar

Settings

Store

Wallet

Phone

Safari

9:41

For you

store.app

Helio Jacket

$2144.9

Aurora Runner

$1294.8

Nimbus Buds

$1594.7

Trailmag 30L

$894.6

Cobalt Mat

$424.5

Forge Bottle

$284.8

Recommendations · Loyalty app

The same engine, in your customer's pocket

A membership app where three models build each member a personal offer: six discounted picks, refreshed every morning, redeemable online or by QR at the register

9:41

Daily picks ★ 1,240 pts

♥

Your favorites

Built from what you bought this week

How it works ›

Refreshed today

6 picks · −20%

A fresh set every morning, from what this member actually buys, online and in store.

New for youBuy againDiscoveries

Claim discounts

Scan at checkout

Discounts apply automatically

Home · today's offer

9:41

Today's picks −20% × 6

⌖Store pickup · ready in 1 hr›

6 picks todaysave $58

−20%

Shell Jacket

New for you

$79$99

−20%

Road Runner

Buy again

$55$69

−20%

Trail 28L Pack

Discovery

$42$52

−20%

Nimbus Buds

New for you

$36$45

−20%

Flask 0.7L

Buy again

$10$13

−20%

Merino Socks

Discovery

$7$9

Daily picks · −20% on 6

Three models, one offer

Collaborative-filtering, content and propensity models compose each member's six picks: ranked, de-duplicated and explainable.
Refreshed every morning

The set re-scores daily on what they actually buy, online and in store, so the offer is never a stale, one-size catalogue.
Redeem anywhere

Claim online or scan a QR at the register: discounts apply automatically, with no codes to type.
Members come back more

A daily reason to open the app lifts repeat visits and basket size, every cohort measured against a held-out control.

The challenge

Personalization you can't prove

Generic storefront

One generic, one-size-fits-all feed for everyone.

You can't prove the lift is real.

Cold-start and a stale catalogue kill relevance.

"Lift" claimed but never statistically significant.

With Zentavor

A store of one: a personal shelf per shopper.

Every lift proven in a 50/50 A/B at p<0.01.

Picks re-scored daily against live behaviour.

+25% conversion · +18% AOV · ×1.8 CTR, measured.

Proven results

Recommendation systems proven in A/B

Retail recommendation systems

+0%AOV

1,500+ stores. +3% revenue lift on all users in a 6-week A/B test.

Cross-sell ML

+0%basket size

+20–30% conversion · 3–4 months to value.

Smart-push triggers

×0.0CTR

ML-segmented triggers · CTR 3.2% → 5.8%.

ML pricing

+0%revenue

+3–5% revenue · dynamic pricing across categories.

A/B experiment

We don't claim the lift. We prove it

A 50/50 holdout: half the shoppers keep the existing recommendations, half get only the new algorithm. Two months later, every metric that matters has moved, and the gap clears statistical significance

experiments / recsys-v2 · 50/50 holdout draft

Cohort

84,210 eligible shoppers

A · control 50% B · variant 50%

Conversion rate · variant vs control variant control

weeks 1–8

+0%

conversion

+0%

average order value

×0.0

click-through rate

+0%

revenue / user

Significant · p<0.01 · n=84,210 · 8 weeks → ship variant B to everyone

We read the result by statistical significance, not by a flattering chart: if the gap doesn't clear p<0.01, the new algorithm doesn't ship and we iterate. This one cleared it on every metric that matters.

FAQ

Your frequently asked questions

How long does it take to see a lift?

Most clients see measurable A/B results within 6–8 weeks of going live. The first two weeks cover data integration and model warm-up; weeks three to six run the 50/50 holdout. We don't call significance early. The gap has to clear p<0.01 before we ship variant B to everyone.

What data do you need to get started?

At minimum: an event stream (views, add-to-cart, purchases) and a product catalogue. Richer signals like search queries, loyalty data and in-store events improve accuracy but aren't required for a first model. We handle cold-start for new users and new items from day one.

How do you handle cold-start for new users and new products?

New users get a popularity- and context-based fallback that personalises within the first session using implicit signals (category, dwell, device). New products are bootstrapped with content features and promoted into the ranking as early signals accumulate, typically within 48 hours of first impressions.

Can the recommendation system run on-prem or in our cloud?

Yes. We deploy into your cloud (AWS, Azure, GCP) or on-prem Kubernetes. Inference latency sits at p95 <50 ms in all environments. Data never leaves your perimeter. Model artefacts, training pipelines and serving infrastructure all run inside your boundary.

How do you measure success beyond conversion?

We track conversion, average order value, click-through rate and revenue per user as primary KPIs. Secondary metrics include catalogue coverage (to avoid popularity bias), repeat-visit rate and, for loyalty apps, redemption rate. All metrics are reported per cohort against a held-out control.

What happens if the A/B doesn't reach significance?

We don't ship a variant that doesn't clear the bar. Instead we diagnose: check for segment heterogeneity, adjust the model or the ranking policy, and re-run. A null result is information. It tells us where to iterate next. Our contracts include iteration cycles precisely because we know real lift requires honest experimentation.

Let's talk

Turn intent into revenue

Tell us your catalog and channels; we'll scope a recommendation system and an A/B plan to prove the lift

Request a demo All solutions →