How AI Product Recommendations Actually Work in Ecommerce

Three Generations of Product Recommendations

Every recommendation system you've ever interacted with falls into one of three categories. They showed up roughly in order, and each one solved problems the last one couldn't.

Most Shopify stores are running generation one. A handful have generation two. Almost nobody has generation three done well. And honestly? Generation one is fine for a lot of stores. But if you're leaving money on the table and want to understand why, it's worth knowing what's possible.

Collaborative Filtering: The "People Like You" Approach

This is the one everyone's heard of. Amazon made it famous. "Customers who bought this also bought that." The basic idea is dead simple.

You build a giant matrix of users and products. Each cell represents an interaction — a purchase, a view, a rating. Then you find users whose matrices look similar and recommend products that one user bought but the other hasn't seen yet. That's it. No understanding of the products themselves, no reasoning about why someone might want something. Just pure pattern matching across behavior.

It works surprisingly well when you have enough data. The problem? You need a lot of data. A store doing 50 orders a day will have sparse, noisy patterns. A store doing 5,000 orders a day has a goldmine. This is the cold start problem — new stores, new products, and new visitors all break collaborative filtering because there's no history to filter on.

The other issue: it's backward-looking. Collaborative filtering tells you what people like this visitor did in the past. It can't react to what this specific visitor is doing right now.

Content-Based Filtering: Matching Product DNA

Content-based filtering flips the approach. Instead of looking at what other people bought, it looks at the products themselves. What are the attributes? Color, material, price range, category, brand, style.

If someone views three blue cotton t-shirts in the $25-35 range, a content-based system says "here are more blue cotton t-shirts in that price range." It's matching product features to inferred preferences.

The advantage: no cold start problem. It works from day one because it only needs your product catalog, not purchase history. The disadvantage: it's narrow. It'll never recommend jeans to someone browsing t-shirts, even if people frequently buy them together. It also depends heavily on how well your products are tagged — garbage metadata in, garbage recommendations out.

Most AI Shopify apps use some version of content-based filtering as a fallback, even if their primary engine is collaborative.

Real-Time Behavioral Analysis: What's Happening Right Now

This is where things get interesting. Instead of relying solely on historical data (what other people did) or product attributes (what the item is), behavioral analysis focuses on what this visitor is doing in this session.

Think about what a great salesperson in a physical store does. They watch. They notice you picked up two jackets and compared the price tags. They see you lingering on the wool section. They register that you walked past the clearance rack without stopping. All of that is signal. And they use it to make a recommendation that's specific to you, right now, in this moment.

Real-time behavioral AI does the same thing digitally. The signals it tracks:

Session-Level Behavioral Signals

• Product view sequences — not just what they viewed, but in what order. Viewing A then B then back to A tells you something different than viewing A, B, C, D linearly.
• Time spent per product — 3 seconds means skimming. 45 seconds means genuine interest. 2 minutes means they're reading reviews and seriously considering.
• Scroll depth — did they scroll past the fold? Did they reach the reviews section? Did they check the size chart?
• Cart activity — adding and removing items is a massive signal. It means they're weighing options. A static recommendation widget misses this entirely.
• Comparison behavior — toggling between two product pages is the digital equivalent of holding two shirts up side by side. The visitor is close to a decision and needs a nudge.

This is the approach that gets the biggest lifts because it's acting on intent, not just history. Someone toggling between two products is in a completely different mindset than someone casually browsing your homepage. They deserve completely different recommendations. For a deeper look at how this tracking works, see behavioral tracking in ecommerce.

The Tiered Approach: Why Smart Systems Don't AI Everything

Here's something that took us a while to figure out when building Maevn. Running AI on every single visitor interaction is expensive and slow. If you send every page view to a language model, you're burning money on decisions that don't need intelligence.

The answer is tiers. Most decisions should be fast and cheap. Only the hard ones need the expensive stuff.

Tier 1 handles maybe 70% of recommendation decisions. Someone lands on a product page? Show related products from the same collection. That's a lookup, not an AI problem. Tier 2 handles another 25% — cases where you need session context, like knowing what they've already viewed. Tier 3 — actual AI reasoning — fires for maybe 5% of interactions. The moments where a visitor is comparing two products, or has a high-value cart and is showing exit intent, or has a browsing pattern that doesn't match any common template.

This is how Maevn's recommendation engine works in practice. The AI layer uses Claude to actually reason about visitor behavior — not just pattern match, but think about what the visitor might need. But it only does this when the cheaper tiers can't handle the situation. And every interaction feeds back into the system, so the tier 1 and tier 2 models get smarter over time as they absorb patterns from tier 3 decisions.

Why "Frequently Bought Together" Isn't Enough

The standard "frequently bought together" widget is collaborative filtering in its most basic form. And it has real limitations that become obvious once you think about them.

First, it treats all visitors the same. The bundle suggestion doesn't change based on who's looking at it. A first-time visitor and a loyal repeat customer see identical recommendations. That's a missed opportunity.

Second, it's only triggered on product pages. What about the visitor who's been browsing for 10 minutes, has three items in cart, and is now stalling? A static widget on a product page doesn't help them. You need a system that can intervene at the right moment with the right offer — and that requires understanding the full session, not just the current page.

Third, it can't reason about why products go together. It knows that people buy a phone case with a screen protector, but it doesn't understand that someone looking at a high-end camera probably needs a memory card, a carrying case, and maybe a lens cleaning kit. That kind of reasoning — understanding the use case behind the purchase — requires actual intelligence, not just co-occurrence data.

Good recommendation systems use "frequently bought together" as one signal among many, not the whole strategy. Think of it as tier 1 — the cheap fast answer that's right 60% of the time. For the other 40%, you need something smarter. The same logic applies to AI sales associates — the best ones combine multiple data sources rather than relying on a single signal.

Measuring Recommendation Effectiveness

If you're not measuring, you're guessing. Here are the metrics that actually matter for product recommendations.

Click-Through Rate (CTR)

What percentage of visitors who see a recommendation click on it? Industry average for recommendation widgets is 3-8%. Below 2% means your recommendations aren't relevant. Above 10% means they're genuinely useful. If you're in the 3-5% range, there's room to improve but things aren't broken.

Conversion Lift

This is the big one. Compare conversion rate for visitors who interact with recommendations versus those who don't. But be careful with this metric — visitors who click recommendations might just be higher-intent shoppers to begin with. The gold standard is an A/B test: recommendations on vs. recommendations off, random assignment, same time period.

AOV Impact

Good recommendations don't just convert more visitors — they increase average order value. Track AOV for orders that include a recommended product versus orders that don't. A well-tuned recommendation engine should lift AOV 10-20% by surfacing relevant complementary products or higher-value alternatives.

Revenue Per Visitor

This is the metric that combines everything — conversion rate, AOV, and traffic quality — into a single number. It's the most honest measure of whether your recommendations are actually making money. Track it weekly and look for trends, not daily fluctuations.

One last thing. Don't optimize for CTR alone. I've seen stores boost recommendation clicks by showing steep discounts on popular items — the CTR looks great, but you're just cannibalizing full-price sales. Always tie recommendation performance back to actual revenue impact.

The gap between basic and great recommendations is huge, and it's only getting wider as real-time behavioral AI becomes more accessible. If you're still running static "you might also like" widgets, you're competing with one hand tied behind your back against stores that are personalizing in real time. The good news? You don't have to build this yourself — the tooling exists. You just need to pick the right tier of sophistication for your store and actually measure whether it's working.