---
title: "How Amazon Drives Add-Ons Using Embedded UI Components"
description: "Learn how Amazon uses embedded recommendations, FBT bundles, cart upsells, and placement psychology to drive add-on purchases without interruptions."
publishedAt: "2026-05-29T18:14:00.000Z"
updatedAt: "2026-05-29T18:14:00.000Z"
author: "Premansh Tomar"
categories: ["App Engagement", "Mobile App Architecture"]
canonical: "https://www.digia.tech/post/amazon-embedded-upsell-strategy-ui-patterns-that-convert"
---

# How Amazon Drives Add-Ons Using Embedded UI Components

> **TL;DR:** Amazon's add-on conversion engine is not a campaign. It is not a notification. It is not a modal. It is a set of embedded UI components that live inside the natural decision flow of a product page, each one placed at a specific moment of established purchase intent and designed to feel like content rather than promotion. This breakdown maps every major embedded component Amazon uses, the placement logic behind each one, the psychological mechanism it activates, and what the model means for mobile engagement teams who are still treating upsell as an interruption problem. **Sourcing note:** Amazon does not publish component-level conversion data. The widely cited "35% of revenue from recommendations" originates from a 2013 McKinsey estimate, later challenged by independent academic research. All data in this article is attributed to its source.

### The Upsell That Does Not Feel Like One

There is a predictable moment in an Amazon shopping session where a user who came to buy one thing ends up buying three. They did not plan to. They were not interrupted by a promotion. They did not receive a push notification. They were simply browsing a product page, scrolling through information they genuinely needed, and found something else sitting in the flow that made complete sense given what they were already about to do.

This is not accidental. It is the result of two decades of deliberate embedded UI architecture, a system that Amazon has built, iterated, and refined since Greg Linden, Brent Smith, and Jeremy York published their foundational paper on item-to-item collaborative filtering in 2003. That paper, which won IEEE Internet Computing's "test of time" award in 2017, described the algorithmic foundation for matching each item a user has purchased or rated to similar items based on what other customers tend to buy together. The "Customers who bought this item also bought" widget was not a marketing decision. It was an engineering decision that happened to produce marketing outcomes.

What Amazon built on top of that algorithm over the following twenty years is the subject of this article: a complete embedded UI system that turns product discovery, cross-sell, and upsell into components that are indistinguishable, in user experience terms, from the core product page itself.

This is article six in Digia's [Engagement and Lifecycle series](https://www.digia.tech/blog/categories/app-engagement/). The previous articles established the format framework ([bottom sheets vs modals](https://www.digia.tech/post/bottom-sheets-vs-modals-interruption-layer/)), the repeat order mechanics ([Swiggy](https://www.swiggy.com/)), and the restraint-first approach ([CRED](https://cred.club/)). Amazon is the counterpoint and the synthesis: aggressive at scale, but embedded so deep in the decision flow that the aggression does not register as such.

### Why "Embedded" Is the Operative Word

Before the component breakdown, the distinction that structures everything else in this article.


![Amazon shopping cart page showing a pacifier recently added to the cart, with a "Top picks for you" recommendation carousel displaying related baby products such as pacifiers, onesies, pacifier clips, diaper organizers, and grooming kits, alongside checkout and cart summary options.](https://cdn.sanity.io/images/53loe8pn/production/954c2404588989b722d2e54d1b5a8900e547d689-2560x1191.png?w=1200&fit=max&auto=format)


There are two fundamentally different ways to surface upsell and cross-sell content in a digital product. One is **triggered interruption**: a modal appears, a bottom sheet slides up, a push notification fires. The user is in some state, browsing, completing a task, navigating, and the engagement campaign inserts itself into that state. The user is asked to switch contexts.

The other is **embedded continuity**: the recommendation exists as part of the page the user is already reading. It does not interrupt anything. It does not require a context switch. The user's scroll naturally carries them through core product information and into recommendation content as a continuation of the same surface.

Amazon has almost entirely committed to the second approach for its add-on conversion system. The product page, the cart page, and the checkout flow all contain components that surface related products inline, within the user's current context, at moments of established decision readiness.

This is not a philosophy stance. It is a conversion finding. Amazon's own 2003 research explicitly noted that "the click-through and conversion rates of recommendations vastly exceed those of untargeted content such as banner advertisements and top-seller lists." Banner advertisements and pop-ups are interruption formats. Item-to-item recommendations embedded in the purchase flow are continuity formats. The conversion differential has driven the architecture.

A 2017 Salesforce study across 150 million shopping sessions found that shoppers who clicked product recommendation links accounted for 26% of revenue on ecommerce sites, despite representing only 7% of visitors. The mechanism: the 7% who click recommendations are already in high purchase-intent states. Meeting them within their existing decision flow, rather than interrupting from outside it, is the structural advantage of embedded components over triggered campaigns.

### The Product Page as a Layered Conversion Architecture

Amazon's product detail page on mobile is often critiqued for being long, scroll depth is extensive, information density is high, the page can feel overwhelming to first-time users. That critique misreads the intent.

The Amazon product page is not designed as a linear document where users scroll from top to bottom. It is designed as a layered architecture with multiple independent decision surfaces, each positioned to capture users at different stages of purchase intent within a single session.


![Amazon homepage featuring a personalized shopping experience with recommendation sections such as "Pick up where you left off," "Continue shopping deals," "Customer's most loved picks for you," and "Deals related to items you've saved." The page also includes a promotional banner for discounted books and curated product suggestions based on browsing and purchase history.](https://cdn.sanity.io/images/53loe8pn/production/b7ec7f7575a218f88bc6bead64d6a199f3a36de2-3600x2338.png?w=1200&fit=max&auto=format)


**Zone 1: Above the fold (core decision content)** Title, primary image, price, ratings summary, delivery promise, and the buy box. This zone is optimised entirely for primary conversion and contains no upsell components. Amazon does not interrupt the primary purchase decision with add-on offers.

**Zone 2: Below the buy box, above reviews (add-on zone)** Frequently Bought Together, Subscribe & Save toggle, and protection plan offers. The add-on conversation begins here, placed at the specific moment when the user has processed primary product information and is in a state of near-committed purchase intent.

**Zone 3: Mid-page (confidence building)** Customer reviews, A+ content, product specifications. This zone serves confidence-building rather than add-on conversion.

**Zone 4: Post-review (discovery zone)** "Customers who bought this item also bought," "Products related to this item," sponsored product rows, and "Customers who viewed this item also viewed" carousels. These components target users who have finished their core evaluation and are in an exploratory mindset.

The strategic insight: **add-on components are placed in Zone 2 (near-committed intent) and Zone 4 (post-evaluation exploration), never in Zone 1 (primary decision).** Amazon does not interrupt the purchase decision. It extends the session after the decision has essentially been made.

### Five Embedded Components That Drive Add-On Conversion

#### 1. Frequently Bought Together (FBT): The Bundle That Looks Like a Fact

FBT is Amazon's most direct add-on conversion component. It appears in Zone 2, directly below the buy box, presenting two or three items alongside the primary product with a single checkbox UI and a combined "Add all to cart" button.


![Amazon product page displaying a "Frequently Bought Together" section beneath a decorative vase listing. The bundle includes the featured vase set, wooden bead garlands, and artificial eucalyptus stems, with checkboxes for each item, a combined price, and a single "Add all three to Cart" button designed to encourage bundled purchases.](https://cdn.sanity.io/images/53loe8pn/production/447319e5aacb7ddd8fa04ba9ca53441beb3d8c8c-2048x1287.png?w=1200&fit=max&auto=format)


The design is deceptively simple. Each item has a checkbox, pre-selected or not. The combined price is shown. The user can deselect individual items before adding everything to cart. The entire interaction takes three to five seconds.

What makes FBT exceptionally effective is its framing. It does not say "you might also like." It says "frequently bought together," which is a statement of observed behaviour, presented as a fact rather than a pitch. The widget's authority derives from its apparent objectivity. It is not a seller recommendation. It is a data-derived observation about how people actually shop. The psychological mechanism is social proof applied to purchase bundling.

The positioning in Zone 2 is critical. The user who has just absorbed the product title, images, price, and reviews is in a state of near-committed purchase intent, having essentially decided to buy and now in confirmation mode. Introducing a bundle at this moment means the additional items are evaluated against a baseline of already-committed spend, not against zero. The anchoring effect makes the add-on items feel inexpensive relative to the primary purchase already in progress.

The one-click "Add all to cart" removes the last friction point. Accepting the bundle requires no navigation, no separate product page visit, no additional checkout steps. It is a single button press at the moment of maximum receptivity.

#### 2. Subscribe & Save Toggle: The Default That Converts


![Amazon Subscribe & Save mobile page showing a cancelled recurring subscription. A highlighted confirmation message states that the subscription for a cat litter product has been cancelled, with an option to reactivate it later. The screen also includes navigation tabs for deliveries, subscriptions, and account settings.](https://cdn.sanity.io/images/53loe8pn/production/463212d300d43cf596e06ce6ae1faad80054de65-1536x864.png?w=1200&fit=max&auto=format)


The Subscribe & Save component offers a recurring delivery discount, typically 5–15%, in exchange for the user setting up a recurring order schedule. On the product page, it appears as a toggle within the buy box area, often pre-selected as the default choice.

The default selection deserves direct examination. Multiple UX analyses have identified this as a persuasion design choice that sits at the edge of what is considered ethical. An Amazon UX analysis noted: "Amazon sets Subscribe & Save as the default not a one-time purchase. I've almost bought things on subscription without realizing it." The mechanism is what behavioural economists call "default bias," where people tend to accept default options when the alternative requires additional cognitive effort.

By defaulting to Subscribe & Save, Amazon converts a one-time purchase into a recurring revenue stream unless the user actively intervenes. This is worth naming directly rather than presenting uncritically. The FBT component, the recommendation carousels, and the zone placement strategy described in this article are all legitimately embedded engagement design. The pre-selected Subscribe & Save default is a darker pattern that achieves conversion by reducing the user's attention to their own decision, not by making the option more genuinely relevant.

The lesson both approaches offer is the same: defaults have enormous power. The ethical and commercial question is whether your default serves user intent or overrides it.

#### 3. Product Recommendation Carousels: Managed Intent Continuation

Amazon deploys multiple horizontally scrollable carousels throughout the product page, each carrying a different intent signal.


![Amazon product page recommendation carousel titled "Customers who bought this item also bought", displaying a PlayStation DualSense controller alongside related PlayStation 5 games and accessories. The horizontally scrollable row includes popular titles such as EA Sports College Football 25, Grand Theft Auto V, Call of Duty, Star Wars Jedi: Survivor, and NBA 2K24, illustrating Amazon's collaborative filtering recommendation system.](https://cdn.sanity.io/images/53loe8pn/production/51bc1f05881e31c1f4db607e95a592ccec074c9a-2954x1020.png?w=1200&fit=max&auto=format)


"Customers who bought this also bought" is pure collaborative filtering, surfacing items commonly purchased in the same transaction and adjacent products the user likely needs but has not yet thought to search for.

"Customers who viewed this item also viewed" is a browsing continuation signal. It targets users in evaluation mode, still comparing and not yet committed, and surfaces the consideration set they are likely moving through.

"Products related to this item" is a broader category exploration signal, often including sponsored placements alongside organic recommendations.

"Buy it again" appears for returning users, surfacing previously purchased consumables. This is the Amazon equivalent of the Swiggy reorder prompt, meeting established habit at the session open.

The horizontal scroll format is not incidental. Vertical scroll communicates "this is content to read through." Horizontal scroll communicates "this is a collection to browse selectively." The format signals to the user that they are in discovery mode, which is the correct frame for a recommendation carousel. Each item is skippable. The user is not expected to evaluate them sequentially. They scan until something catches their attention.

Each carousel is an independent conversion surface. A user who ignores the FBT bundle and scrolls past the reviews may still tap an item in the "Customers who viewed" carousel. The layered architecture means multiple components get independent shots at the same user within the same session, without any single component feeling like an interruption, because none of them demand attention.

#### 4. The Cart's Inline Recommendation Strip

When a user opens their cart view, Amazon surfaces a recommendation row directly within the cart page, typically "Customers who bought items in your cart also bought," presenting items relevant to current cart contents.


![Amazon shopping cart page showing multiple retro gaming controllers in the cart, with a checkout summary panel on the right and an inline recommendation section featuring additional products such as games, collectibles, cleaning accessories, and toys. The recommendations appear alongside the cart contents, encouraging add-on purchases without interrupting the checkout flow.](https://cdn.sanity.io/images/53loe8pn/production/b3044276206199d838a0db8225a50ab68e8b9577-1500x1130.png?w=1200&fit=max&auto=format)


The placement in the cart is strategically distinct from the product page. A user viewing their cart is in completion mode. Their primary focus is reviewing what they have, confirming the total, and proceeding to checkout. They are not in discovery mode.

Amazon's inline cart recommendation sits at the edge of that completion focus, present and accessible at a high-intent moment, but not forcing itself into the user's attention. A user moving purposefully toward checkout will not register it as an obstacle.

The conversion mechanism here is completion bias: users who have already built a cart want to feel the cart is complete. A component that says "people who bought what you have also got this" activates that feeling, making them wonder whether they are missing something obvious. The tap that results is not persuaded. It is satisfied.

#### 5. The Protection Plan and Warranty Offer: Inline Insurance at the Buy Decision

For electronics and other high-value items, Amazon surfaces protection plan options directly in the buy box area, inline, not as a modal, not as a popup after add-to-cart, but as a selectable option within the purchase UI itself.


![Screenshot of an Amazon India product page for a DMR 7.5 kg top-load washing machine. The product image is displayed on the left, while the center section shows the product title, ratings, price of ₹6,499, EMI options, and promotional offers. On the right, the purchase panel includes delivery details, quantity selection, and "Add to Cart" and "Buy Now" buttons. A yellow circle highlights the optional protection plans section, featuring a 1-year extended warranty for ₹199 and a 2-year total protection plan for ₹349.](https://cdn.sanity.io/images/53loe8pn/production/a36c848acf5479ad2864509eb8a62ad5d27e19ad-3600x2338.png?w=1200&fit=max&auto=format)


The user evaluating a ₹35,000 laptop is in a heightened loss-aversion state. The purchase is large enough that damage or failure would be painful. A protection plan offer placed at the moment of primary purchase decision, rather than after checkout confirmation, intercepts the user exactly when they are actively thinking about risk.

The inline placement matters here more than anywhere else in the Amazon system. A modal offering a protection plan after checkout completion would feel like an afterthought at best. The user has already completed their primary decision; asking them to reconsider their risk exposure after the fact creates cognitive dissonance. The inline offer is part of the primary decision surface, presenting the product, the extended warranty, and the total for each together so the user makes one decision, not two.

The psychological mechanism is prospect theory: users weight losses more heavily than equivalent gains. Framing the protection plan as protection against loss (your device fails and you are unprotected) is more effective than framing it as a feature gain. Amazon's protection plan copy consistently uses protection framing rather than feature framing.

### The Algorithm Is Not the UX

One important distinction this breakdown requires: the quality of Amazon's embedded components is not just a function of the algorithmic quality of the recommendations. It is a function of the UI architecture that surfaces those recommendations.

Amazon's item-to-item collaborative filtering has been replicated by hundreds of competitors, many of whom have built recommendation engines of comparable quality. What they have not replicated with the same precision is the placement logic, the decision about where in the page hierarchy to surface which component, at what scroll depth, using which interaction model.

A competitor who surfaces a "Customers also bought" carousel above the buy box, or who opens a modal at the moment of add-to-cart to surface related products, will get worse conversion outcomes from the same algorithmic quality, because the placement violates the intent state the user is in at that scroll depth.

The algorithm tells Amazon what to recommend. The embedded UI architecture tells Amazon when and how to recommend it. **Solving the first without solving the second produces systematic underperformance.**

### What the McKinsey 35% Number Actually Means

The figure that appears in nearly every article about Amazon's recommendation engine, that 35% of Amazon's revenue comes from product recommendations, traces to a 2013 McKinsey study. It has also been challenged. Anuj Kumar, a professor at the University of Florida's Warrington College of Business, published research finding that product recommendations boosted product sales by approximately 11%, and that recommendations actually reduced sales of some less-popular items by making more popular items more visible.

The honest read: Amazon's embedded components drive meaningful additional revenue. The exact magnitude is not reliably established in the public record. The 35% figure is a McKinsey attribution from 2013, not a conversion experiment result. Anyone citing specific format-level conversion rates attributable to Amazon without a primary source should be treated with scepticism.

What is not in dispute: Amazon's product page generates add-on purchases at scale. The mechanism is embedded UI placement at moments of established purchase intent.

### Why Format Determines Conversion Outcome

The [bottom sheets vs modals framework](https://www.digia.tech/post/bottom-sheets-vs-modals-interruption-layer/) established a general principle: format is a structural signal about whether engagement continues the user's task or interrupts it.

Amazon's embedded component system is the most comprehensive implementation of the continuation principle in ecommerce. Every add-on component lives within the user's existing scroll context. None require the user to stop their current task. None demand acknowledgement before the user can proceed. All are skippable with zero effort.

The conversion this system generates is not despite this permissiveness but because of it. Users who are not interrupted stay in the browsing mindset longer. Users who are not anxious about closing a modal are more receptive to content in their peripheral attention. Users who have been treated as having agency are more likely to exercise that agency by tapping something that genuinely interests them.

The modal upsell produces a binary outcome: the user engages or dismisses with a mildly negative experience. The embedded component produces a probabilistic outcome: across a large enough user base, the percentage of users whose scrolling attention lands on a genuinely relevant recommendation and results in a tap is consistently positive, without producing any negative experience for the majority who scroll past.

### The Infrastructure Question: Why Most Apps Cannot Do This Yet

Amazon's embedded component architecture is not difficult to understand. It is difficult to replicate for a specific reason that has nothing to do with algorithmic quality or design sophistication.

The embedded components, the FBT bundle, the recommendation carousels, the protection plan inline offer, are server-configured. Amazon can change what appears in each zone, for which user segments, based on which product, without a release cycle. The product page is not a static template. It is a configurable surface where the content of every recommendation zone is dynamically controlled.

This is the same architectural advantage described in the [first article in this series](https://www.digia.tech/post/server-driven-ui-for-engagement/), a backend-driven UI system where the team can modify what components appear, where they appear, and to which users, without engineering sprint costs for each change.

Most mobile apps that want to implement Amazon-style embedded engagement cannot do it because their UI is hardcoded. The product page or home screen has fixed sections. Adding a "customers also bought" equivalent requires a release. Testing whether it should appear above or below reviews requires two releases and a four-week experiment cycle.

The apps that can replicate Amazon's embedded engagement model at pace are the ones whose UI is configurable at the component level, where growth teams can add, move, test, and remove inline components from a dashboard rather than a sprint.

[Digia Widgets](https://www.digia.tech/products/widgets/) is built for exactly this gap. Carousels, grids, embedded cards, recommendation strips, all configurable from a dashboard, renderable inline within any screen, without requiring a release. The component architecture Amazon built into its product page can be replicated inside any mobile app's core flow. What it requires is the ability to place and iterate inline components at the speed that makes the placement logic meaningful.

### What Growth and Product Teams Can Extract

**The intent-zone mapping principle.** Before placing any add-on or recommendation component, identify which intent state the user is in: primary decision mode (evaluating the main action), near-committed mode (essentially decided), or post-evaluation mode (finished primary task, browsing residually). Place components only in near-committed or post-evaluation zones. Never in primary decision zones. Amazon does not upsell above the buy box. Neither should you.

**The skippability principle.** Every embedded component should be skippable with zero effort. The user who can skip frictionlessly will not be frustrated by the component's existence. The user who cannot skip will carry that frustration into their evaluation of every subsequent component. Zero-friction skippability is the prerequisite for ambient attention capture.

**The relevance-before-volume principle.** Amazon surfaces many recommendation components, but the relevance of each to the user's current context is high. For teams without collaborative filtering infrastructure, a smaller number of highly relevant embedded components will convert better than a large number of generic ones. The FBT bundle with three well-chosen items outperforms a carousel of thirty loosely related ones.

### Key Takeaways

- Amazon's add-on conversion system is an embedded UI architecture where recommendation components live inside the natural decision flow, each placed at a specific intent-state zone, each skippable without effort.
- The product page has four intent zones. Add-on components appear in Zone 2 (near-committed intent, below buy box) and Zone 4 (post-evaluation exploration, post-reviews). Zone 1 is kept clean. This zone placement architecture is as important as recommendation quality.
- Frequently Bought Together converts via social proof framing combined with anchoring (add-on cost evaluated relative to already-committed primary purchase) and a one-click bundle action that removes the last friction point.
- The Subscribe & Save default pre-selection is a documented dark pattern in Amazon's otherwise well-structured embedded system. Defaults are powerful. The ethical distinction is whether the default serves user intent or overrides it.
- The "35% of Amazon revenue from recommendations" is a 2013 McKinsey attribution, challenged by independent research finding approximately 11% actual lift. The directional case for embedded over interrupted is not in dispute. The specific figure should not be cited without the caveat.
- The algorithm (what to recommend) and the UI architecture (where and how) are separate problems. Amazon's advantage is not just recommendation quality but precision of placement architecture.
- Most apps cannot replicate this model at pace because their UI is hardcoded. Server-driven component architecture, configurable from a dashboard, is the infrastructure prerequisite.

### Further Reading

#### From Digia

- [What is Server-Driven UI for Engagement (And Why It Matters)](https://www.digia.tech/post/server-driven-ui-for-engagement/)
- [Eliminating Mobile App Release Dependency for Engagement Experiments](https://www.digia.tech/post/eliminating-app-release-dependency/)
- [Bottom Sheets vs Modals: Choosing the Right Interruption Layer](https://www.digia.tech/post/bottom-sheets-vs-modals-interruption-layer/)
- [How Swiggy Uses Bottom Sheets to Drive Repeat Orders](https://www.digia.tech/post/swiggy-bottom-sheets-repeat-orders/)
- [Breaking Down CRED's Subtle In-App Nudges](https://www.digia.tech/post/cred-in-app-nudges-breakdown/)
- Spotify's Upgrade Nudges and Timing Strategy _(up next)_

#### External Sources — All Claims Attributed

- [Amazon.com Recommendations: Item-to-Item Collaborative Filtering — History](https://www.amazon.science/the-history-of-amazons-recommendation-algorithm) - Amazon Science (Linden/Smith/York 2003 paper; IEEE "test of time" award 2017)
- [Amazon Product Detail Page: Section Breakdown](https://litcommerce.com/blog/amazon-product-detail-page-guide/) - LitCommerce
- [UF Research Challenging the 35% Figure — 11% Actual Lift](https://warrington.ufl.edu/news/how-valuable-are-online-product-recommendations/) - University of Florida / Warrington College of Business
- [Salesforce: 26% Revenue from 7% of Recommendation Clickers](https://www.practicalecommerce.com/study-personalized-recommendations-produce-4-times-conversions) - Practical Ecommerce (Salesforce 2017 study)
- [Amazon AOV ~$52 on US Marketplace](https://redstagfulfillment.com/amazons-average-order-value/) - Red Stag Fulfillment
- [Subscribe & Save Default Pattern Analysis](https://medium.com/design-bootcamp/amazons-ux-a-masterclass-in-psychology-or-manipulation-3ced8a9ea6f4) - Medium / Kanvi Makwana
- [Cross-selling Contributes Up to 30% of Ecommerce Revenue](https://www.getkard.com/blog/10-proven-ways-to-increase-average-order-value-for-ecommerce-growth) - Kard
- [Cart Upsells: 10–30% AOV Increase](https://upsell.com/blog/cart-upsell-examples) - Upsell.com

_This article is part of Digia's Engagement and Lifecycle series. Next: Spotify's Upgrade Nudges and Timing Strategy — how timing and constraint-triggered moments drive free-to-paid conversion better than any discount._

_Want to add inline recommendation components to your app without a release cycle?_ [See how Digia Widgets work](https://www.digia.tech/products/widgets) or [book a demo](https://calendly.com/anupamsingh-digia/connect).