> ## Documentation Index
> Fetch the complete documentation index at: https://docs.kaireonai.com/llms.txt
> Use this file to discover all available pages before exploring further.
# E-commerce & Retail Media
> Recommend the next product that actually moves the customer — not the one they'd buy anyway — across web, email, and push without over-messaging.
## The problem in this industry
E-commerce has no shortage of signals: browse history, cart contents, purchase frequency, category affinity. The hard part is deciding what to do with them. The default playbook — retarget everyone who abandoned a cart, email everyone a discount, push everyone the new arrival — burns two things at once: discount margin, and the customer's patience. A shopper who was going to reorder their coffee pods anyway doesn't need a 15%-off code; handing them one is pure margin leakage. Meanwhile the genuinely undecided shopper gets the same generic blast and tunes it out.
The winning move is incrementality: spend your best offer on the customers whose behavior the offer actually changes. That means separating the "sure things" (would convert regardless) from the "persuadables" (convert only if nudged) — and staying quiet for the "lost causes" so you don't waste send cost or annoy someone into unsubscribing.
KaireonAI decides per shopper, per moment. It knows what they already own (so it won't re-pitch it), predicts what they're most likely to want next, estimates how much each candidate offer *incrementally* lifts conversion, and coordinates across web, email, and push so the same person isn't hit five times for the same thing.
## What you build in KaireonAI
You translate your merchandising strategy into a handful of platform building blocks. Here is how a typical online-retail setup maps on:
| Platform concept | How you use it in e-commerce |
| ------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Offers** | Free-shipping-threshold nudge, category discount, product bundle, loyalty-tier upgrade, abandoned-cart reminder, replenishment reminder, new-arrival alert, VIP early access |
| **Channels** | On-site web module, email, push, SMS, retargeting audience |
| **Decisioning gates** (Eligibility → Fit → Match) | *Eligibility*: `segment_required` for VIP early access; suppress a bundle whose items the shopper already owns (`offer_attribute`). *Fit*: `recency_check` so a replenishment reminder only fires when it's plausibly time to reorder. *Match*: boost offers in the shopper's top affinity category |
| **Contact policies** | `frequency_cap` per channel; `cross_channel_cap` so the same product isn't pushed on web *and* email *and* SMS in one day; `category_suppression` to avoid three fashion promos in a row; `cooldown` on a dismissed offer; `do_not_contact` for opt-outs |
| **Scoring approach** | [`neural_cf`](/ai-ml/algorithms/neural-cf) shines here — its two-tower embeddings learn "shoppers like this one responded to product Y," which is the essence of a recommendation engine. Pair it with a [`gradient_boosted`](/ai-ml/algorithms/gradient-boosted) model for rich tabular features, and a [`thompson_bandit`](/ai-ml/algorithms/thompson-bandit) for fast-learning promotions |
The decisive addition is an [uplift model](/ai-ml/uplift-modeling). It tags each shopper per offer as **persuadable**, **sure thing**, **lost cause**, or **sleeping dog**, so ranking can prioritize the persuadables and stop wasting discounts on shoppers who'd buy anyway.
## A worked example
**Priya** is a repeat customer. She bought running shoes six weeks ago, browses the `activewear` category weekly, has never used a discount code, and abandoned a cart with a jacket in it two days ago. She's opted into email and push.
Free-shipping nudge, activewear discount, shoe+sock bundle, loyalty upgrade, abandoned-cart reminder, replenishment reminder, new-arrival alert, VIP early access.
The shoe+sock bundle drops (she just bought the shoes — `offer_attribute` ownership check). The replenishment reminder drops (six weeks is too soon for running shoes — `recency_check`). Five survive.
A `cross_channel_cap` notices the new-arrival alert already went out on push this morning; it's suppressed on email so she isn't double-hit for the same thing.
The collaborative-filtering model scores the survivors. The uplift model flags Priya as a **sure thing** for the activewear discount — she reliably buys activewear without a code — so ranking down-ranks it (why give away margin?). The **abandoned-cart reminder** for the jacket ranks first: high propensity, and uplift says the nudge genuinely tips her from "thinking about it" to "bought."
Priya gets one timely, relevant reminder — full price — instead of a scattershot of promos that would have trained her to wait for a coupon.
**A price-sensitive shopper gets a different answer.** Run the same flow for a customer who only ever converts with a code, and the uplift model flags them **persuadable** for the discount — so the activewear discount that KaireonAI *withheld* from Priya is exactly what it surfaces for them. Same offers, same rules; the platform spends the margin only where it changes behavior.
## Measuring success
* **[Business Dashboard](/operations-reporting/dashboards#business-dashboard)** — conversion rate and revenue per offer, and the daily impressions/conversions trend.
* **[Attribution Dashboard](/operations-reporting/dashboards#attribution-dashboard)** — compare Last-Touch vs. First-Touch vs. Linear credit to see which channel introduces vs. closes, and reallocate spend accordingly.
* **Uplift and holdout** — a control group in your [experiment](/ai-ml/uplift-modeling) proves the program lifts *incremental* revenue and quantifies how much margin you save by not discounting the sure-things.
* **Suppression rate** — confirms your cross-channel caps are actually preventing the five-messages-for-one-product problem.
## Where the agentic layer helps
[Decisioning Autopilot](/ai-ml/autopilot) keeps your recommendation and promo models sharp without a standing data-science rota. When a challenger model or a new bundle promo beats the incumbent with significance, Autopilot proposes the swap; drift on a live model triggers a retrain proposal. You choose whether those land as suggestions to review or apply automatically through the audited path.
[Decision Sentinel](/ai-ml/sentinel) catches the config accident that quietly kills revenue — an over-aggressive frequency cap or a broken segment rule that empties your on-site recommendations. It watches the decision stream for suppression spikes and empty decisions and raises an alert (with optional auto-pause) before a merchandising slot goes blank. And the [governed AI assistant](/ai-ml/ai-assistant) lets a merchandiser make a change in plain language — "cap push to twice a week for the sale segment" — with the edit routed through approvals.
## Try it
Re-engage lapsed shoppers with recency-aware eligibility and incrementality measurement.
Recommend the next product without pitching what a shopper already owns.
Separate persuadables from sure-things so your offers create real lift.
Register and build an e-commerce flow end to end.