Personalizing Your Store with AI

AI personalization means adapting the experience and the products shown to each visitor, in real time, based on their behavior and intent. Instead of an identical journey for everyone, each customer sees a store that feels made for them. On paper, it's the most obvious conversion lever in e-commerce. In the audits we run at Stellar, it's also the one most often installed and least often fed: the block is there, it's actually running on your best-sellers, and nobody has ever measured it.

There's one nuance few articles own up to: on Shopify, what's native in 2026 isn't individual personalization, it's product-to-product recommendation. The difference isn't semantic, it determines your roadmap. This guide sorts out what the platform already does for free, what you need to buy, the volume of data you really need, and the French legal framework you have to comply with. For the big picture of AI on the platform, see our guide to AI on Shopify.

Personalization or recommendation: the costly confusion

Personalization adapts the entire experience to a given visitor: content, product order, messages, offers. Recommendation is one building block of that personalization, the one that suggests the right products. Recommendation is part of personalization, never the other way around. Confusing the two means buying a recommendation engine and believing you've solved personalization.

A second distinction, just as fundamental: rules versus learning. A rule is fixed, along the lines of "if the visitor comes from this campaign, show this banner". An AI model learns from behavior and predicts what will be relevant, without you having to program everything in advance. That's where the whole gap lies, provided you have enough to feed the model.

A third distinction, the most useful in practice. There are three levels of personalization, and they have neither the same cost nor the same prerequisites:

• The product level: "customers who bought this also bought that". The starting point depends not on the visitor but on the page being viewed. This is what Shopify's native engine does.
• The segment level: one group of customers gets different treatment from another. New visitors versus loyal customers, buyers from one category versus buyers from another.
• The individual level: the experience is recalculated for one person, based on their own history. This is true one-to-one, and it's the most demanding in terms of both data and compliance.

What AI personalization really changes

A visitor who sees relevant products buys more, and more often. Personalization acts on conversion rate, average order value and loyalty, the three engines of a store's profitability. It turns a passive catalog into an advisor that guides. So much for the principle. The numbers, however, deserve a close reading.

The most solid data available today is measured, not self-reported. Salesforce, which observes more than 1.5 billion shoppers in more than 89 countries, put figures on the 2025 holiday season: $262 billion in sales influenced by AI and agents, out of $1.29 trillion in global online sales, or about 20% of online sales for the period (data published on January 8, 2026). The previous year, AI and agents had influenced $229 billion, or 19%. The growth is real, but it isn't explosive.

On the expectations side, the 2025 Twilio survey, conducted among 7,640 consumers and 637 business leaders in 18 countries, gives the real figure to remember: 88% of consumers say they're more likely to buy when the relationship is personalized in real time, but only 44% of the brands surveyed say they manage to do it. The gap isn't a technology problem, it's an execution and data problem.

In France, the benchmark comes from Fevad and Odoxa (February 2026): 31% of online shoppers already use generative AI when shopping online, and 49% among 15- to 24-year-olds. On the merchant side, Fevad notes in the 2026 edition of its Key Figures that 94% of online merchants now use generative AI solutions.

An honest counterpoint is in order. The famous "+40% revenue from personalization" attributed to McKinsey is a misreading: the 2021 study says that fast-growing companies derive 40% more of their revenue from personalization than slow-growing companies. It's a correlation, not a gain. The orders of magnitude McKinsey publishes for retail are more modest: a 1 to 2% increase in total sales for grocery players, and more for other retailers.

Another counterpoint, less well known. A randomized experiment by Lee and Hosanagar, published in Information Systems Research and covering 82,290 products and 1,138,238 users, shows that classic collaborative filtering reduces aggregate sales diversity: it concentrates the catalog on products that are already popular. Personalization isn't magic, it's a lever you have to manage.

This is the paradigm shift summed up by Daniel Zyss, PhD in bioinformatics (Mines ParisTech) and co-founder of Radar, interviewed in our book Créer sa marque à l'ère de l'IA.

We're moving from a model where everyone follows the same generic journey to a model where the experience adapts to each person, in the right context, at the right time. It's a question of relevance.

The five forms of AI personalization

AI personalization takes several forms on a store. The five most useful, in order of ease of implementation:

• AI product recommendations: cross-selling, up-selling, "complete the look", similar products. This is the entry point, and the only form partly covered by Shopify's native features.
• Smart search: results ranked by affinity rather than by simple word matching. Often the best return on investment, because a visitor who uses search already has strong intent.
• Personalized email: messages triggered by behavior, through a tool like Klaviyo. On our own brands, Superbon and Cabania, email accounts for 28 to 35% of revenue, as we detail in our Klaviyo guide for Shopify. It's the channel where personalization pays off fastest, because the customer is identified there.
• Adaptive merchandising: highlighting the right product to the right segment, taking stock and margin into account. Since September 2025, Shopify lets you segment customers by the product categories they've viewed or bought, based on its standard taxonomy of more than 16,000 categories.
• Dynamic content: a homepage and collections that reorder themselves according to the profile. It's the most spectacular, the most expensive, and the one that requires the most traffic to be reliable.

The most telling illustration remains intent. Daniel Zyss describes it like this:

You're chatting with an AI agent and you tell it you have to go to a wedding in Italy in July. A linen suit brand detects that intent and instantly generates an experience within the conversation.

Shopify's native recommendation engine, and its four limits

Natively, Shopify does product-to-product recommendation and nothing else: automatically generated related products, manually entered complementary products, and semantic search reserved for paid plans. None of these building blocks adapts the experience visitor by visitor. Here is the exact state of native features as of August 19, 2026, before you install anything.

The four limits of the native engine

Four limits deserve to be spelled out in black and white, because they change the buying decision.

First limit, and the most important for a French store. Shopify's automatic recommendation algorithm relies on three strategies. The official documentation specifies that the one based on product description similarity is only available to merchants whose storefront is in English. On a French-language store, that leaves only purchase history and related collections. In practice: if your catalog is young and your collections are poorly organized, the native engine has almost nothing to work with.

Second limit. The recommendations API takes a product ID as input, not a visitor ID. It answers the question "which products go with this one", never "what should we show this person". The only documented native individual personalization goes through Shop identity: the AI sales assistant introduced in the Spring '26 Edition, built on Shopify Inbox, recommends products based on the customer's history only when they're logged in with Shop.

Third limit. Orders imported from another platform aren't taken into account by the algorithm. After a migration, the engine starts from scratch on purchase history, even if you've brought over ten years of orders.

Fourth limit. Out-of-stock products, products priced at zero, gift cards and products already in the visitor's cart are excluded from recommendations. That's good default behavior, but you need to know it before concluding that the block "doesn't work". Note: you can't configure the algorithm to ignore a product, but the documentation does allow you to filter, on the theme side in JavaScript, which of the returned products you display.

The three data thresholds you need to know

Yes, you need volume, and here are the orders of magnitude nobody gives: about 1,000 interactions for a model to exist, about 50,000 for it to beat your best-sellers. In between, you need to distinguish three thresholds that app vendors have every interest in blurring.

The gap between the training threshold and the quality threshold is a factor of fifty. That's exactly the range in which most personalization apps are sold, and it's where the disappointment happens: an engine serving disguised best-sellers, billed at the price of a predictive engine.

Two useful clarifications. First, training time is counted in days, not hours: Klaviyo cites two days to a week to train its recommendation models, then retraining every two to seven days, with a 90-day observation window on products viewed and added to cart. Second, contrary to a common belief, a longer history helps rather than hurts: Google's documentation states that model quality improves markedly with more data, and that a full year of purchases lets models make better use of seasonality and trends.

Translated into merchant language: below a hundred or so orders per month and a few thousand product views, the most profitable investment isn't a recommendation engine, it's structuring your catalog and improving the quality of your collections.

Our roadmap for deploying on Shopify

There's no point deploying everything at once. Here's the sequence we apply with our clients, in this order, because each step is a prerequisite for the next.

1. Structure the data first. Shopify taxonomy categories filled in, complete metafields and metaobjects, coherent collections, real descriptions. On a French store, related collections are one of only two signals the native engine can use: treating them carelessly amounts to sabotaging the engine before you switch it on.
2. Activate the native features and measure them. Install Search & Discovery, fill in complementary products for your twenty to thirty best-selling products (nobody does it, it's manual), and take care of your filters within the 25 available. Then let it run for three to four weeks and record a baseline.
3. Tackle internal search. It's almost always the most profitable opportunity, because a visitor who searches has already made up their mind. If your plan includes semantic search, turn it on and measure the rate of searches with no results before buying a third-party solution.
4. Personalize email. Behavioral segments and triggered messages at the right moment. It's the only channel where the customer is identified by design, so it's where individual personalization is both the simplest and the most legally sound.
5. Buy an app, only now. Once the data is clean, the native features are put to use and a baseline is established, you know what you're buying and you'll know whether it pays off.
6. Measure the right indicators. Click-through rate on recommendation blocks is a vanity metric. What counts is average order value, items per order and overall conversion rate, compared against your baseline. Since the Winter '26 Edition, Shopify offers native A/B testing on the online store, which makes honest comparison accessible without development.

Which recommendation app should you choose in 2026?

Below 50 orders per month, the right answer is to stick with Search & Discovery and your email tool: no paid app pays for itself at that volume. Above that, the market has split between accessible tools billed by order volume and platforms built for Shopify Plus. Data collected on August 19, 2026 from the Shopify App Store.

The three pricing traps to check before you sign

You never see them on the app's home page:

• The entry price doesn't always include personalization. At Boost, personalized search and AI-personalized recommendations are only available on the Accelerate tier, starting at $699 per month. The $29 displayed buys you search and filters, not a predictive engine.
• Shopify Plus compatibility isn't a given. LimeSpot's entry-level plans are explicitly labeled "not compatible with Shopify Plus". Only the Max plan, the most expensive, isn't.
• Watch out for billing indexed on impressions or attributed revenue. Glood includes a quota of widget displays and then bills for overages: your cost follows your traffic, not your sales. Others charge a percentage of the revenue they attribute to themselves, with an attribution method you don't control. Have the bill simulated on your real volumes, not on the advertised tier.

GDPR, the AI Act and personalized pricing: the French framework

Personalizing means processing data. The topic is often dispatched in two lines at the bottom of the page, even though it determines the technical architecture. The right way to approach it is to build two distinct levels.

The two consent levels to build

The no-consent level. Recommendations calculated solely from customer account data and order history, along with strictly necessary trackers (cart, authentication, language). This scope can rely on legitimate interest, provided you document the balancing test and offer a simple way to object.

The mandatory-consent level. Anything that reads from or writes to the visitor's device for behavioral personalization purposes, any enrichment with third-party data, any sharing with an ad network. In its recommendation on cookies and other trackers, adopted on September 17, 2020 and amended by the deliberation of December 18, 2025, the CNIL explicitly classifies content personalization among the purposes subject to consent. When it comes to personalization, only interface personalization (the choice of language or layout) is exempt, provided it is an intrinsic and expected part of the service.

Sanctions are no longer theoretical. On September 1, 2025, the CNIL imposed a €150 million fine on Shein and €325 million on Google, notably for trackers placed before consent was collected and kept after a refusal. On December 30, 2025, a €3.5 million fine sanctioned the transfer of the contact details of more than 10.5 million loyalty program members to a social network for advertising purposes, without a valid legal basis. The operational rule that follows is simple: the personalization engine must not technically start before the consent signal.

What the European AI regulation really says

Two clarifications help avoid overreacting. A product recommendation engine isn't classified as high risk: it isn't listed in Annex III of Regulation (EU) 2024/1689. The transparency obligation in its Article 50, applicable since August 2, 2026, targets systems that interact directly with a person: a conversational agent is covered, a recommendation block isn't. However, the ban on manipulative techniques and the exploitation of vulnerabilities has applied since February 2, 2025: no AI-generated false scarcity, no targeted sales pressure on profiles detected as vulnerable.

The special case of personalized pricing

This is the point most often overlooked. Article L. 221-5, I, 11° of the French Consumer Code requires informing the consumer, before the order, that the price has been personalized on the basis of automated decision-making. Failure to do so is punishable by an administrative fine of up to €75,000 for a legal entity, under Article L. 242-10. Our position, which we systematically recommend in audits: personalize recommendations, offers and promo codes, never the catalog price. You keep the commercial benefit and you fall outside the scope of the obligation.

The six mistakes we see most often in audits

These are, in order, the six situations that come up most often when a store tells us that "personalization isn't delivering anything".

• Installing the app before structuring the data. A recommendation engine plugged into empty product pages and catch-all collections can't produce anything interesting. This is by far the number one case.
• Leaving complementary products empty. They're manual in Shopify, so nobody fills them in, so the highest-quality block on the product page never gets used.
• Stacking three recommendation blocks on the same page. "Similar products", "you may also like" and "frequently bought together" one under the other cannibalize each other and lengthen the mobile page. Shopify's documentation actually recommends four related products per product page and two to three complementary products, no more.
• Measuring by clicks. Click-through rate on a recommendation block always goes up. Average order value only goes up if the suggestions are genuinely relevant. That's the only real arbiter, along with items per order.
• Starting the engine before consent. Technically invisible, legally expensive, and it's precisely the violation cited in recent sanctions.
• Over-personalizing. Locking a customer into what they already know kills discovery, and the catalog concentration effect is documented. Keep a share of novelty and editorial curation: personalized content is still your content, and it has to carry your voice, as we explain in our method for keeping your brand consistent.

And a common-sense reminder worth more than any algorithm: no personalization saves a bad product. It amplifies a desirable offer, it doesn't create one.

Toward the adaptive experience

Today's personalization, built on recommendation blocks, is only one stage. What comes next is a fully adaptive experience, where the store reconfigures itself for each visitor, all the way to the agent that converses and advises.

The shift is already visible in Shopify's product trajectory: most of the platform's investment in 2025 and 2026 went into agentic commerce. What interests us here is the consequence for personalization: it no longer plays out only on your site, it plays out in the conversation where your product is mentioned, and it relies on the identity carried by the agent, not on your cookies.

The practical consequence is almost reassuring: the work that prepares for one prepares for the other. Rich product data, clean categories, complete attributes and coherent collections are exactly what feeds a recommendation engine and what makes a catalog readable by an agent. Personalization is the first step on that staircase, and it's the only one that delivers measurable results today.

AI-driven e-commerce personalization is only as good as three things: clean data, sufficient volume and a desirable product. The rest is configuration. Want a personalization engine tailored to your store and connected to Shopify? Discover our AI Studio or talk to our Shopify agency about it.