Product recommendations on autopilot:
what sells next, not just what else exists
Most storefronts show 'related products' picked by category tags, which is why the recommendations rarely match what customers actually buy together. We build a model trained on your own order history, so the recommendation slot earns its shelf space with real co-purchase and behaviour patterns.
The process today
A “you might also like” section on most storefronts is built from category tags or manually curated lists that someone set up once and rarely revisits. It feels like personalisation, but it is really just a static rule: same category, similar price, maybe the same brand. It does not know that customers who buy this protein powder usually buy that shaker bottle, or that buyers of one skincare product tend to come back for a specific second item three weeks later, because nobody has looked at the actual order data closely enough to notice.
The cost is a recommendation slot, often some of the most valuable screen space on a product page, that underperforms its potential. Industry benchmarks put well-tuned recommendation engines at meaningfully higher attach rates than generic related-items widgets, and the gap between the two is pure upside sitting unused on pages that already get traffic.
The second cost is maintenance. A manually curated list goes stale the moment the catalogue changes, a product goes out of stock, or a new line launches, and nobody has the bandwidth to keep hundreds of product pages’ recommendations fresh by hand.
What the agent does
The model trains on your own order history and, where available, browsing behaviour, learning which products are genuinely bought together, bought in sequence, or substituted for each other, specific to your catalogue and your customers, not a generic similarity formula. It places recommendations on product pages, in the cart, and in post-purchase or abandoned-cart emails, with each placement tuned for its context: a cart recommendation favours a quick add-on, a post-purchase email favours a complementary item that fits the delivery timing.
Before replacing whatever you run today, we A/B test the new model against your current logic on a slice of traffic, so the decision to roll it out fully is based on a measured lift, not a promise. New products without enough order history yet get a cold-start fallback based on category and attributes, so a launch is never stuck showing an empty slot or an irrelevant one.
Exclusion rules you set, out of stock, discontinued, below a margin floor, are applied before anything is shown, so the model never recommends something that would hurt margin or disappoint a customer who clicks through to a dead product page. The model refreshes as new orders come in, so it keeps learning rather than freezing at launch-day patterns.
What stays with humans
Merchandising strategy, promotional priorities, and any manual override for a specific campaign, feature this collection this week regardless of what the model says, stay with your team. The model proposes based on behaviour; a merchandiser can always pin, exclude or reorder what shows in a given placement.
Guards
Every recommendation set is logged with the data behind it, and the A/B test result against your previous logic is kept so the rollout decision is documented, not assumed. Exclusion rules are enforced as a hard filter before the model’s suggestions are shown, and a kill switch reverts to your previous related-products logic in one message if something looks off.
Price and timeline
| Option | Price | What it covers | Timeline |
|---|---|---|---|
| Single automation | from $800 | One catalogue, model, product page and cart placement | 7 to 12 days |
| Department package | from $2,500 | Recommendations plus upsell email and post-purchase sequences across your store | 2 to 4 weeks |
Running cost is usually $20 to $70 a month depending on catalogue size and traffic.
Related
Pair this with upsell and cross-sell for the messaging layer that acts on these recommendations, and with RFM scoring so the highest-value customers see recommendations tuned to their buying pattern specifically. The full package breakdown is on the AI agents service page and the automation-everything overview; for a real storefront’s results, see the Thailand D2C rebuild case study and the Balkans supplements store case study.
Ready to see what your own order data actually says customers buy together? Get in touch and we will look at your catalogue in the first call.
Tired of doing this by hand? We can take the whole routine off your team, not just this step: Routine takeover, from $400 →
FAQ
How much does recommendation automation cost?
From $800 for a model covering your main catalogue, placed on product pages and cart, live in 7 to 12 days. A department package adding personalised email and post-purchase recommendations usually starts at $2,500.
How is this different from Shopify's or a plugin's built-in recommendations?
Built-in logic is often rule-based (same category, same tag) and does not learn from your actual conversion data. This model trains on what customers in your store genuinely buy together and keep, and we A/B test it against whatever you run today before asking you to switch.
What happens with a brand-new product with no sales yet?
It falls back to a category and attribute-based recommendation until it has enough of its own order data, so new launches are not stuck showing nothing or showing irrelevant items.
Can we control what never gets recommended?
Yes. Out-of-stock items, discontinued products, and anything below a margin floor you set are excluded automatically, regardless of how strong the behavioural signal is.
Where does it show up for customers?
Typically product pages, the cart, and post-purchase or abandoned-cart emails. We start with one placement, measure it, and expand once it is proven.