How to use AI for cross-selling and upselling (ChatGPT, Claude, Gemini): a step-by-step e-commerce guide
A step-by-step guide to cross-selling and upselling with AI: data, pairs, rules, moments and order-level measurement, and what to automate on Tiendanube.
In short
To do cross-selling and upselling with AI you need five pieces: clean catalog and order data, pairs that make sense, business rules, the right moment and order-level measurement. ChatGPT, Claude or Gemini help propose pairs; running the system takes live data. CrossUp automates it on Tiendanube with SalesPilot, CrossUp’s AI, combined with your rules.
Cross-selling adds a complementary product to what the customer chose; upselling offers a better version, a bigger pack or more units. To do both with AI you need five pieces: clean data, pairs that make sense, business rules, the right moment and order-level measurement. This guide shows how to build each piece by hand with ChatGPT, Claude or Gemini, and how CrossUp automates them on Tiendanube (Nuvemshop in Brazil) with SalesPilot, CrossUp's AI, combined with your own rules.
Asking a general-purpose AI to review your catalog is a great first step: in seconds you get reasonable pairs, like a coffee maker with coffee, a dress with a handbag or a phone with a case. Then come the questions that decide whether the strategy pays off: which of those pairs actually sell? What happens when a product runs out or the season changes? Should every visitor see the same offer? This guide answers those questions.
Which cross-selling and upselling strategies can you run with AI?
Before the steps, it helps to separate the strategies, because each one has its moment and its way of being automated (examples are illustrative):
| Strategy | Goal | Example | Moment | How CrossUp automates it |
|---|---|---|---|---|
| Complementary cross-sell | Add a product to the order | Coffee maker → filters | Add to cart, cart, checkout | SalesPilot picks the complement; you pin the key pairs |
| Upgrade upsell | Raise the value of the chosen product | 128 GB phone → 256 GB | Product page | Rules you set; SalesPilot adds alternatives when similarity signals exist |
| Quantity upsell | Sell more units | Socks on a 3-for-2 | Product page | Volume discount campaigns (2-for-1, 3-for-2, packs) |
| Bundle or kit | Sell a complete solution | Skincare routine | Product page, add to cart, home page | Manual bundles or bundles suggested by SalesPilot |
| Post-purchase recommendation | Set up the next purchase | A complement for what they just bought | Thank-you page | SalesPilot chooses what to show |
Quantity upselling has its own guide in volume discounts: 2-for-1, 3-for-2 and packs, and complements in related products on Tiendanube.
What does a general-purpose AI do really well?
A generative model is excellent at organizing ideas, classifying products, spotting complements and suggesting upgrades. It can also help you write offer titles, define margin criteria, review an exported file or turn your business knowledge into a consistent checklist.
Use it to speed up your reasoning. Give it context on each product, explain who buys from your store and ask it to justify each association. The better the input, the more useful the suggestions.
Day-to-day operation asks for more. By default, a conversation doesn't see your new orders, live stock, each visitor's cart or the final outcome of each recommendation, and it doesn't keep an identifier from the moment the offer appears until the order is paid. For that you need to connect it to data, events and rules that run all the time.
How do you do cross-selling and upselling with AI, step by step?
These are the ten steps of the full system, with the most common mistake in each one:
| Step | What to do | What usually goes wrong |
|---|---|---|
| 1. Prepare the data | Export ID, name, category, variants, price, stock and, if you have it, margin. For orders, ID, date, products, quantities, discounts and total are enough, without personal data. | Inconsistent names or variants loaded as separate products |
| 2. Find what's bought together | Look at frequency, the probability that B appears when A is bought and whether the pair occurs more than popularity alone would explain. | Sticking to best sellers only |
| 3. Generate complements and upgrades | Combine that signal with the model's knowledge and ask for two separate lists: complements and better versions. | Mixing a complement with a substitute |
| 4. Apply your business rules | Filter by stock, margin, price, promotions, compatibility and season. Define what to prioritize. | Optimizing acceptance and draining critical stock or margin |
| 5. Choose the moment and format | The product page is for exploring; add to cart confirms the decision; the cart completes; checkout calls for something short; the thank-you page sets up the next purchase. | Showing the same recommendation across the whole journey |
| 6. Implement it in the store | Up-to-date price and stock, one-click add, variants respected and a stable identifier all the way to checkout. | Losing the offer's identification when the shopper moves to payment |
| 7. Measure order-level attribution | Track impression, click, acceptance, add, removal and purchase. Attributed revenue is the value of the recommended item, not the whole cart. | Counting clicks as sales or crediting the whole order |
| 8. Review results every week | Compare acceptance, attributed revenue, average order value and products per order by pair, moment and format. | Looking only at an overall average |
| 9. Update catalog and seasonality | Reprocess new items, removals, prices and stock. Plan reviews for Black Friday, Cyber Monday and your industry's key dates. | Leaving rules on because they worked once |
| 10. Segment and keep learning | Use signals like viewed products, cart contents, country, currency and new versus returning customer, and feed every result back into the system. | Segments so small they never gather enough signal |
For step 3, a prompt that works well:
Using this catalog and this co-occurrence matrix, suggest three complements and two upgrades per product. Justify each option with data, compatibility and usage context.
On step 7: order-level attribution tells you which products came in through each recommendation. If you also want to measure how much of those sales wouldn't have happened without the recommendation, you need a controlled comparison. We go deeper into the moments in step 5 in the 5 moments to recommend products.
Which problems show up as the system grows?
An initial list can work very well and lose strength over time. These are the five challenges that appear once recommendations run every day.
Cold start: deciding with little history
Cold start is the lack of enough history to decide well, and it shows up at three levels: a new product with no sales, a new store with no orders and a new visitor with no browsing. By hand, start with category, attributes and compatibility, add curated rules and review the first responses quickly. With SalesPilot, your pinned pairs and bundles work from day one, product similarity adds alternatives when it's available and co-purchase data grows stronger as orders come in. And because it decides from the product the shopper is viewing and their cart, a first-time visitor gets a recommendation as thoughtful as a returning one.
Freshness: good rules age too
A recommendation loses meaning when the catalog changes, a product sells out or the season ends. By hand, review stock and new items every week, set expiration dates and build a commercial calendar. SalesPilot always works with your synced catalog and stock, so a sold-out product stops being recommended on its own.
Repetition and popularity bias
If you rank only by sales, best sellers take every slot: what you show most sells most, and those sales justify showing it again. By hand, compare affinity as well as volume and rotate categories and price ranges. SalesPilot starts from what gets bought together with the product the shopper is viewing, not from a general best-seller ranking.
Sparse data in the long tail
Among slow-selling products, many pairs appear once or twice, which may be a real affinity or pure coincidence. By hand, require a minimum volume and back the pair with compatibility or curated rules. SalesPilot cross-checks those pairs with your catalog signals, like variants and product similarity when it's available, and always respects the pairs you pin.
Attribution and causality
Sometimes you recommend something the shopper would have bought anyway. Order-level attribution is the first filter: it counts the item that came in through the offer, not the whole cart. SalesPilot traces every recommended item to the order that contains it; measuring strict causal impact calls for adding a controlled measurement design.
How much does it cost to run by hand?
The first setup requires cleaning data, analyzing orders, designing offers, implementing events and testing the full journey. Then comes maintenance: several hours a week to review results, replace products, monitor stock and adapt the strategy to each season.
The work grows by multiplication. A catalog of 300 products × 5 moments × 3 segments already opens 4,500 combinations of product, placement and audience. You won't review them one by one, but someone has to catch the ones that went stale. With a small, stable catalog, doing it by hand is a great way to learn; as the store grows, it becomes a permanent operation.
What does SalesPilot do with this whole system?
SalesPilot, CrossUp's AI, turns the ten steps into a continuous process:
- Steps 1 and 2: it analyzes your catalog and your store's order history. SalesPilot connects catalog, order, variant and co-purchase signals to understand relationships between products: what we call your store's product map.
- Steps 3 and 4: it discovers complements, and your rules pin pairs, order and exclusions. Automated and manual work side by side.
- Step 5: it decides which product to recommend at each of the five moments and can suggest bundles that you review before publishing.
- Steps 6 and 7: CrossUp implements the modules in your store and traces every accepted product to the order as attributed revenue.
- Steps 8, 9 and 10: the dashboard shows attributed revenue by moment and by product so you can decide what to adjust, SalesPilot always works with up-to-date catalog and stock and, instead of hand-built segments, it decides from the product the shopper is viewing and their cart.
For quantity upselling, CrossUp adds volume discount campaigns: 2-for-1, 3-for-2 and packs of up to 6 units, configured in the dashboard. The logic of combining AI and rules is explained in AI cross-selling vs. manual rules and in what SalesPilot is and how it decides.
Manual or automated?
The manual path is a good way to learn with a small, stable catalog, few promotions and a first controlled implementation. It helps you understand which combinations make sense and which metrics to watch.
For most stores that are already selling, automating is the natural next step. You install CrossUp on Tiendanube, turn on SalesPilot and the AI adapts to your catalog and stock from day one, with a 15-day free trial to see attributed revenue on your own data. CrossUp also supports VTEX integrations through a guided implementation. CrossUp's plans combine a fixed component with a success fee on attributed revenue.
Key takeaways
- Generative AI is a great tool for analyzing the catalog, proposing pairs and organizing criteria.
- A complete system adds live data, business rules, moments, implementation and order-level measurement.
- As it grows, cold start, aging rules, repetition, sparse data and the question of causality come up.
- SalesPilot automates the continuous cycle, works alongside your rules and measures every attributed sale all the way to the order.
The same playbook, applied to one industry: how to sell more in an online clothing store and how to sell more in an online toy store.
Frequently asked questions
What is the difference between cross-selling and upselling?
Cross-selling adds a complement to what the customer chose; upselling offers a better version, a bigger pack or more units.
Can ChatGPT do cross-selling for my store?
It helps analyze the catalog and propose pairs. Running it every day requires connecting it to stock, orders, events and order-level measurement.
What can be automated with AI?
Choosing complements, the moment to show them, updates for stock and catalog, and the measurement of every attributed sale.
How do I start with little data?
With category, attributes and curated rules. With SalesPilot, your pinned pairs and bundles work from day one, and co-purchase data grows stronger as orders come in.
When should I use manual rules?
For launches, clearances, supplier agreements or hard exclusions; the AI covers the rest of the catalog.
How do I measure whether it works?
With order-level attributed revenue: the value of the item that came in through the recommendation and stayed in the paid order.