Skip to content

Recommendation and Personalization

Recommendation systems that decide, not just retrieve.

Most engines surface similar items. The architecture that pays off in 2026 decides the next best action for each visitor, in real time and across every channel. JoulesLabs builds recommendation and personalization systems into the stack you already run and takes them to production.

A 30-minute call to find the recommendation problem worth solving first. Free AI Readiness Assessment for qualified companies.

  • 40+ products since 2017
  • 20+ engineers
  • 5.0 on Clutch
  • 4.8 on Google
  • 93% on-time
  • Built in Dhaka, delivered for clients worldwide

A recommendation and personalization capability is the set of models, data pipelines, and real-time services that decide what to show each visitor, so the experience fits their intent and lifts revenue per session. Its two jobs are to recommend the right products and to personalise the experience across every channel, and the payoff shows up as higher average order value, better conversion, and more repeat purchases. JoulesLabs builds this as an engineering partnership that takes a recommender to production, not as a tool you install and run yourself. That scope frame matters up front: JoulesLabs ships commerce and product work (iVendorz, a shipped B2B marketplace build) and AI features inside products (ReviewX and ReviewXpo), and it does not sell a packaged recommendation engine. Personalization itself is no longer the differentiator. 71% of consumers expect companies to deliver personalised interactions, and 76% get frustrated when that does not happen (McKinsey, 1,013 consumers, November 2021). When everyone personalises, the difference is architecture: whether the system decides the next best action or only retrieves similar items.

This page is written for the person who owns revenue per session, a head of e-commerce, a VP of product, or a digital director in the UK, EU, GCC or USA, and for the engineering lead who will own whatever gets built. It covers what the capability does, how product recommendations move revenue, why most engines retrieve when they should decide, how real-time and cross-surface personalization work, the cold-start problem, the data foundation underneath it, readiness for agentic commerce and AI-answer visibility, privacy and transparency, and how recommendations feed engagement and loyalty. It ends with what JoulesLabs has actually shipped, a build-versus-buy start guide, and the questions a buyer asks before a call.

Last updated: November 2026

What is a recommendation and personalization capability, and where does it pay off?

A recommendation and personalization capability decides what to show each visitor, from the products on a page to the offer in an email, so the experience matches their intent and moves revenue per session. It pays off wherever a catalogue is large enough that no one can hand-merchandise it and traffic is high enough that small lifts compound: retail and e-commerce first, then banking and telco, where a next-best-offer per customer does the same job.

The capability has two jobs. It recommends: it puts the item a visitor is most likely to want in front of them. And it personalises: it adapts the wider experience, search results, content, and offers, to the person and the moment rather than to a segment. Both depend on the same engine deciding, in real time, which action is most likely to move this visitor forward.

JoulesLabs treats this as production engineering: the models, the data pipelines, and the real-time serving path that a recommender needs to work on real traffic. The clearest place to see the stakes is AI product recommendations for retail, the demand hub where order value and conversion are won or lost.

  • What it recommends

    Products on the page and in the cart, the next-best-offer in an email or an app, search and discovery results ranked to intent, and the content a visitor sees next. The common thread is a decision per visitor, not a fixed rule for everyone.

  • What personalise means here

    Personalise means the system reads what a visitor is doing now and adjusts within the session, and keeps that context the same whether they are on mobile, on desktop, or in an email. It is not a one-time segment assignment; it is a live read of intent.

“The sign is a merchandiser doing the engine’s job by hand. When someone is rewriting the rules every Monday so the homepage stops showing ten versions of the same shoe, the recommender is retrieving and a person is deciding. That labour is the cost of not having a decisioning layer.”

Mehedi Hasan Nahid, CTO, JoulesLabs

How do product recommendations raise average order value and conversion?

Product recommendations raise revenue by putting the item a visitor is most likely to buy in front of them at the moments they decide: the product page, the cart, and the post-purchase email. The model reads the catalogue and the visitor's behaviour signals, then ranks what to show so the visit is worth more without feeling pushy. Personalization most often drives a 10 to 15% revenue lift, with company results spanning 5 to 25% by sector and execution (McKinsey, November 2021). Generative AI could add USD 240 billion to 390 billion in value for retailers, a margin gain of 1.2 to 1.9 percentage points (McKinsey, 2023). Those are market figures, not JoulesLabs results, and they size the lever rather than promise an outcome.

The ranking and propensity models under recommendations are predictive analytics work: they estimate what a visitor is likely to do next and order the options accordingly.

  1. Where recommendations earn the most

    Product page: complementary and similar items that raise basket size before checkout.
    Cart: last-moment cross-sell that fits what is already in the basket.
    Post-purchase email: the next-best item once you know what they bought.
    Search and discovery: results ordered to intent so the right product surfaces first.

  2. Upsell versus cross-sell

    Upsell offers a better or larger version of what the visitor is already considering; cross-sell offers a complementary item. Both raise order value, and the engine decides which fits the moment rather than showing both at once.

“The cart. A product page suggestion competes with the product; a cart suggestion rides on a decision already made. In every build we have seen, one well-chosen complementary item at the cart moved more revenue than a whole carousel upstream, because the visitor has stopped browsing and started buying.”

— Mehedi Hasan Nahid, CTO, JoulesLabs

Why do most recommendation engines retrieve when they should decide?

Retrieve means the engine surfaces items similar to what a visitor already saw. Decide means the system chooses the single next action most likely to move the visitor forward, which may be a different product, a reassurance, or a wait. Most engines retrieve, and the architecture that works decides the next best action rather than only surfacing similar items, which is where the return on investment is lost or won.

  1. What retrieve looks like

    A visitor views a blue running shoe and sees ten more blue running shoes. The engine matched features; it did not ask what would actually move this person toward a purchase.

  2. What decide looks like

    The same visitor sees a complementary item, a size-and-fit reassurance, or a better-value option, chosen because the system judged it the most useful next step given the session so far.

  3. The two-question test

    Ask any engine two questions before you buy it: does it choose an action, or only rank similar items? And does it react within the current session, or from yesterday's data? Two noes mean you are buying retrieval.

The difference is architectural. Deciding needs a layer that can weigh options and act on live signals, which is agentic systems engineering: an engine that reasons about the next step instead of returning a ranked list of look-alikes. This is the buyer-protective test to apply before signing anything, because a retrieval engine dressed as a decisioning one is the most common gap between a demo and production value.

“Ask where the session state lives. If the answer is a nightly batch job and a feature table, the engine is ranking yesterday’s look-alikes however modern the model is. A system that decides keeps the live session in memory and can change its mind between two clicks.”

— Mehedi Hasan Nahid, CTO, JoulesLabs

How does real-time personalization read intent across channels?

Real-time personalization reads what a visitor is doing in the current session and adjusts what they see within it, then keeps that context the same whether they move to mobile, desktop, or email. Personalization in 2026 is defined by timing: it anticipates intent in real time rather than responding to yesterday's behaviour, and speed sets relevance. A recommendation read from yesterday's clicks is already stale, because the visitor's intent has moved on.

  1. Why timing beats targeting

    Targeting assigns a visitor to a segment and serves that segment's content. Timing reads the live session, so a visitor who switched from browsing to comparing gets a different, more useful response in the moment. The second one wins because intent is a now signal, not a profile.

  2. Keeping context across surfaces

    A visitor browses on a phone, compares on a laptop, and buys from an email. Cross-surface consistency means the system carries one view of that person and one view of their intent across all three, so the recommendations do not contradict each other.

Campaign and channel personalization in the marketing stack is a different lane; that work routes to marketing personalization and GEO. On this page the focus is the in-session, on-site system that decides what a shopper sees next.

“Latency under real traffic. A real-time path that works at ten events a second and falls over at a thousand is not real-time. We know it is working when the recommendation shown in the session reflects the click from two seconds earlier, measured on the busiest hour, not the quietest.”

— Mehedi Hasan Nahid, CTO, JoulesLabs

How do you recommend to first-time visitors with no history?

You recommend to first-time visitors by leaning on what you do know in the moment: the item they are viewing, popular and seasonal lines, and the signals in the current session, until enough behaviour accrues to personalise properly. Recommendation platforms handle cold-start visitors with several fallback strategies, and the problem remains a standing requirement in every enterprise build.

  1. The three fallbacks

    Context: recommend from what the visitor is looking at right now, the item and its category.
    Popularity: fall back to what is selling and what is in season when context is thin.
    In-session signals: read clicks, dwell, and search within the visit to sharpen as the session goes on.

  2. When to switch to personalised

    Switch from fallbacks to personalised recommendations once a visitor has given enough signal, a few meaningful actions or a first purchase, so the system personalises on evidence rather than on a guess.

Cold start is not an edge case; it is most of your traffic on any given day, so a system that only works once it knows someone leaves money on the table at the top of the funnel.

“Context first, popularity second, then in-session signals as soon as there are any. The item a visitor is looking at tells you more than anything a cold model can guess, and three clicks later the session already beats a profile. Guessing from demographics is the fallback we never use.”

— Mehedi Hasan Nahid, CTO, JoulesLabs

What data foundation does a recommendation system need to work?

A recommendation system works only when its data foundation does: connected sources, consistent event flows, and a real-time path, because the model can decide no better than the signals it receives. Gartner forecast in 2025 that 60% of AI projects without AI-ready data will be abandoned through 2026, and a recommender is only as good as the event data behind it (Gartner, 26 February 2025). This is the wedge the off-the-shelf engines do not deliver as a service: they assume the data is already clean, connected, and flowing in real time, which in most businesses it is not.

  • The three foundation checks

    Sources connected: catalogue, behaviour, and transaction data reach one place the model can read.
    Events consistent: the same action is recorded the same way everywhere, so signals are comparable.
    Real-time path present: fresh events reach the model fast enough to act within the session.

  • Why real-time is the hard part

    Batch data tells you what happened yesterday; a decisioning recommender needs what is happening now. Standing up a reliable real-time path, and keeping it reliable under load, is the part teams most often underestimate.

Building that foundation, the pipelines, the event consistency, and the real-time serving, is data engineering work, and it is usually where a recommendation project is actually won or lost.

“The same action recorded three different ways. Add-to-cart from the app, the web and the email each arrive as a different event with different fields, so the model never sees one customer. Fixing the event schema lifts a recommender more than any change to the model.”

Mehedi Hasan Nahid, CTO, JoulesLabs

How do you get products recommended inside AI answers and by shopping agents?

You get products recommended by AI systems by making them legible to those systems: give each product a clear, structured product story and make sure it is present where answer engines and shopping agents pull recommendations from. This is search optimisation reframed for the agentic era. 61% of European consumers already use AI for product discovery and evaluation, and AI could unlock EUR 240 to 320 billion in value across European retail over five years (McKinsey and EuroCommerce, 10 June 2026). As AI systems become the recommendation surface, a brand has to be present where AI pulls recommendations from.

  1. What agentic commerce changes

    When an agent shops on a person's behalf, the buyer is no longer only a human reading a page; it is a system reading structured data. Products that are not machine-legible simply do not get considered, however good they are.

  2. Making products citable by AI

    The same structured product story that feeds the on-site recommender, clear attributes, relationships, and descriptions, is what an answer engine needs to cite the product. One investment serves both surfaces, which is why GEO for commerce and recommendation data quality are the same problem.

The mechanics overlap with how AI answers retrieve any source: products are pulled into answers through retrieval-augmented generation, so a clean, structured product story is what makes a product both recommendable on-site and citable off-site.

“Structure. The visible product is a near twin of its competitor, but one of them has clean attributes, named relationships and a plain description a machine can parse, and the other is a hero image and a PDF. The answer engine cites the one it can read.”

— Mehedi Hasan Nahid, CTO, JoulesLabs

How do you keep personalization within privacy and transparency rules?

You keep personalization compliant by designing consent, data minimisation, and explainability into the system from the start, so it can show why it recommended something and respect the choices a visitor has made. Online platforms in the EU must state in plain language the main parameters of their recommender systems (Regulation (EU) 2022/2065, Article 27), and the EU GDPR and UK GDPR set the consent and data-minimisation duties for everyone (Regulation (EU) 2016/679, Articles 5 and 7). Bolting privacy on after the model is built is both harder and riskier than designing for it at the foundation.

Designing consent, minimisation, and explainability into the data and the model is AI security and governance work, and it is what lets a regulated business personalise at all.

The three governance checks

Consent: the system only uses data a visitor agreed to share, and honours a withdrawal.
Minimisation: it collects and keeps the signals it needs, not everything it could.
Explainability: it can show, for a given recommendation, why that item was chosen.

“Withdrawal. Teams build consent capture carefully and forget that a visitor can take it back. When someone withdraws, their behavioural history has to leave the feature store and the model’s inputs, not just the marketing list, and very few systems can show an auditor that it did.”

— Mehedi Hasan Nahid, CTO, JoulesLabs

How do engagement and loyalty programs use recommendations to keep customers?

Engagement and loyalty programs use recommendations to make each contact relevant: the next-best-offer for that member, timed reminders, and rewards that match what the customer actually buys, which is what turns a first order into a repeat one. Next-best-offer and content personalization are core recommendation applications across retail, banking, and telco, and companies that grow faster drive 40% more of their revenue from personalization than slower-growing peers (McKinsey, November 2021). This is the largest pool of demand around recommendations, because retention compounds in a way that a single conversion does not.

  1. Next-best-offer in loyalty

    A next-best-offer model asks what single offer is most likely to move this specific member, then feeds it into the loyalty touchpoints: the app, the email, the reminder. The program stops sending the same promotion to everyone and starts sending the right one to each person.

  2. Beyond points

    Points mechanics reward spending; a recommendation layer makes the whole relationship relevant, so the member comes back because the offers fit, not only because a balance is accruing.

“Offer acceptance rate, before anything else. The first week a next-best-offer model goes live, the share of members who act on an offer moves before revenue does, because the promotion stops being generic. Repeat purchase rate follows a quarter later. Watch the first number; the second comes.”

— Mehedi Hasan Nahid, CTO, JoulesLabs

What has JoulesLabs built in commerce and personalization?

JoulesLabs has shipped commerce and product work rather than a packaged recommendation engine, and it is clear about what is proven versus what is approach. What is shipped is a B2B marketplace build (iVendorz) and AI features inside products (ReviewX and ReviewXpo, including AI review replies). What is not shipped is a deployed recommendation engine for a client: that work is JoulesLabs' approach, and the first client agents are at proof-of-concept stage. No client-result statistic is asserted here, because none is cleared; the market and vendor figures elsewhere on this page stay attributed to their sources, never to a JoulesLabs build.

That candour is the point. You are told exactly which rows are shipped and which are approach, which is the thing most capability pages in this field do not give a buyer, and it is what de-risks the decision. If you want to talk through where a recommender would pay off for your catalogue, book a call.

  • What has shipped

    A shipped B2B marketplace build (iVendorz) and shipped AI features inside products (ReviewX and ReviewXpo). These are real commerce and product work, framed as what they are, not as a recommendation product.

  • What is still our approach

    Every recommendation and personalization use case on this page is JoulesLabs' approach unless the product-recommendations row is read against iVendorz as a shipped commerce build. A deployed client recommendation engine does not exist yet; it is engineered to production as a partnership, with the build shown before you commit.

“Because the hard part of a recommender is not the model, it is shipping it inside a live commerce system, and that is what we have done. iVendorz is a working B2B marketplace we built end to end. The recommendation model is new to us as a client deliverable; the production engineering is not.”

Mehedi Hasan Nahid, CTO, JoulesLabs

Should you build a recommendation system or buy a platform, and where do you start?

Buy a platform when a default recommender on your stack is enough and you have no unusual data or decisioning needs; build when relevance, real-time decisioning, or AI-answer visibility is where your revenue sits. Then start with the single problem that moves the most money for your catalogue, and measure one outcome, not five. Personalization has moved from differentiator to baseline, and 78% of consumers say personalised content makes them more likely to repurchase (McKinsey, November 2021), so where you focus first decides the gap you open.

Start guide: which recommendation problem to solve first

Your situationWhat to build or buy firstFirst outcome to measure
Basic recs only, limited dataBuy a platform default and connect your dataProduct-page conversion
Strong catalogue and traffic, stale relevanceBuild a decisioning layerAverage order value
Omnichannel, but inconsistent across surfacesBuild cross-surface contextRepeat purchase rate
Preparing for AI-answer and agent demandBuild structured product data and GEOAI-answer presence
  1. When to buy

    Buy when a platform default handles your catalogue, your data is ordinary, and your revenue does not hinge on relevance or timing. A good default beats a half-built custom engine.

  2. When to build

    Build when relevance, real-time decisioning, cross-surface consistency, or AI-answer visibility is where your revenue actually sits, because that is exactly what the off-the-shelf engines do not deliver as a service.

  3. Where to start

    Start with the one problem that moves the most money for your catalogue, ship it, measure a single outcome, and only then widen. If you want a second opinion on which problem that is, contact us.

“Solve the one placement that touches the most money, usually the cart, and watch average order value on that placement alone. One metric on one surface tells you in a week whether the engine decides or retrieves, and only then is it worth spending on the rest of the stack.”

— Mehedi Hasan Nahid, CTO, JoulesLabs

“A recommender that retrieves shows you more of what you just saw; one that decides chooses the next useful step from the live session. Production readiness means that decision arrives in milliseconds under real traffic, with the data path and the consent trail that let you prove why it was made.”

— Mehedi Hasan Nahid, CTO, JoulesLabs, November 2026

Find the recommendation problem worth solving first

A 30-minute call to map your catalogue, your traffic, and where relevance is leaking revenue. Then a free AI Readiness Assessment if you qualify.

Book a call about your recommendation problem

No obligation. You leave with a clear starting point, whether or not you build with us.

FAQ Frequently asked questions

Do you sell a recommendation engine product?

No. JoulesLabs builds recommendation and personalization systems and takes them to production as an engineering partner, rather than licensing a packaged engine you install yourself.

Has JoulesLabs deployed a recommendation engine for a client yet?

Not yet. What is shipped is commerce and product work (iVendorz) and AI features inside products (ReviewX and ReviewXpo). Client recommendation work is JoulesLabs' approach and is at proof-of-concept stage, and you are told which rows are proven versus approach.

How long does a first build take?

A first build is scoped to one high-value problem rather than the whole system, so you see production behaviour on a real workflow before committing to more.

What does an engagement cost?

First contracts are typically USD 20,000 to 50,000, scoped to the work, and a qualified company can start with a Free AI Readiness Assessment (planned) before committing.

Which markets do you serve?

JoulesLabs works with retail, marketplace, and subscription teams in the UK, EU, GCC and USA for this capability, and delivers from Dhaka for clients worldwide.

How do you handle privacy?

Consent, data minimisation, and explainability are designed into the system from the start, so it uses only data a visitor agreed to share and can show why it recommended something, which is what GDPR expects in the UK and EU.

Book a call to scope your highest-value recommendation problem