How to use AI for cross-selling and upselling (ChatGPT, Claude, Gemini): a step-by-step e-commerce guide

You asked ChatGPT, Claude, or Gemini to review your catalog and suggest products that belong together. Within seconds, you had sensible pairs: a phone with a case, a dress with a handbag, a coffee maker with coffee beans. You loaded them into a popup, and your strategy was live.

Then came the harder questions. Which pairs actually sell? Did the recommendation add the product, or was the shopper already planning to buy it? What happens when an item sells out, the season changes, or a hundred new SKUs arrive? Should every visitor see the same offer?

Building the first list is the most visible part of the work, but it is only the beginning. A complete strategy also needs data, commercial rules, placement decisions, tracking, order-level attribution, and continuous learning. This guide shows you how to build that system manually—and where automation starts giving you your time back.

What general-purpose AI does really well

A generative model is excellent at organizing ideas, classifying products, spotting semantic complements, and proposing upgrades. It can also help you write offer titles, define margin criteria, review an export, or turn your merchandising knowledge into a consistent checklist.

Use it to accelerate your thinking. Give it context for each product, explain who shops at your store, and ask it to justify every association. Better inputs produce more useful recommendations.

The gap appears in daily operations. By default, a chat does not see your latest orders, live inventory, each visitor's browsing behavior, or the final outcome of every recommendation. It also cannot carry an identifier from impression to order on its own. That requires connections to data, events, and rules that run continuously.

The step-by-step guide

1. Prepare your catalog data

Export product ID, name, category, variants, price, availability, inventory, and relevant attributes. Add cost or margin if available. For orders, you need the order ID, date, products, quantities, discounts, and total—leave out personal data that adds nothing to the analysis.

What usually goes wrong: inconsistent names, variants treated as unrelated products, or discontinued items still marked as sellable.

2. Calculate which products appear together

Before asking for ideas, analyze co-occurrence across your orders. Look at frequency, the probability that B appears when A is purchased, and whether the pair happens more often than product popularity alone would suggest. A spreadsheet works for a small catalog; more volume calls for repeatable queries or a data pipeline.

What usually goes wrong: choosing only the best sellers. Two popular products may overlap often without making a useful recommendation.

3. Generate complements and upgrades

Combine your historical signal with the model's semantic knowledge. Ask for two separate lists: complementary products for cross-selling, and premium versions, second units, or free-shipping threshold options for upselling.

Using this catalog and co-occurrence matrix, suggest three complements and two upgrades per product. Justify each option based on data, compatibility, and use context.

What usually goes wrong: mixing up a complement and a substitute, or recommending incompatible variants.

4. Apply your commercial rules

Filter each suggestion by inventory, margin, price, current discounts, compatibility, and seasonality. Define what you want to prioritize as well: clearing a category, promoting a launch, or protecting items with limited stock.

What usually goes wrong: optimizing for acceptance only to discover that the offer depleted critical inventory or erased your margin.

5. Choose the moment and format

The product page supports discovery; an add-to-cart popup captures confirmed intent; the cart helps shoppers complete their order; begin checkout and checkout need a concise offer; the thank-you page sets up the next purchase. Explore the six conversion moments and see them in CrossUp's experience.

What usually goes wrong: showing the same popup, with the same product, throughout the journey.

6. Implement the recommendation in your store

Every format needs current prices and availability, one-click adding, and correct variant handling. The recommendation also needs a stable identifier that travels from the moment it appears through checkout completion.

What usually goes wrong: building a polished offer that loses its identity when the shopper moves to payment.

7. Measure order-level attribution

Track impressions, clicks, acceptances, additions, removals, and purchases. An attributed order is one where the recommended item was added through the offer and remained in the paid order. Attributed incremental revenue is the value of that item—not the entire cart.

This tells you which sale the module generated. Proving strict causality against what would have happened without the recommendation requires an additional controlled comparison.

What usually goes wrong: treating clicks as sales or crediting the full order to one recommended item.

8. Review results every week

Compare acceptance, incremental revenue, AOV, and products per order for each pair, moment, and format. Keep what adds value, fix weak associations, and determine whether a drop comes from price, relevance, or inventory.

What usually goes wrong: looking at one storewide average that hides both excellent offers and poor experiences.

9. Refresh catalog and seasonal logic

Reprocess new items, removals, prices, and availability. Schedule special reviews around seasonal changes, major sales events, and dates that matter to your category. A winning winter recommendation may stop making sense months later.

What usually goes wrong: leaving old rules active because they worked once.

10. Segment and feed the learning loop

Use signals such as viewed products, cart contents, country, currency, language, and new-versus-returning status. Every impression, acceptance, and rejection can improve the next decision if that feedback flows back into the system.

What usually goes wrong: creating segments too small to build a useful signal, or showing the same pair to everyone forever.

The problems that emerge as the system scales

An initial list can perform well and still degrade over time. These are the problems that start to matter once recommendations are operating every day.

Cold start: making decisions with little history

Cold start means there is not enough history to make a well-supported decision. It appears at three levels: a new product has no sales yet; a new store has not accumulated orders; and a new visitor has not produced browsing signals. This happens because co-purchase analysis and personalization need evidence. Manually, start with category, attributes, and compatibility, add curated rules, and review the first responses quickly. SalesPilot reduces that initial dependency by learning from more than 1,000 stores across every vertical and combining catalog, order, browsing, and seasonal signals.

Freshness: even a good rule gets old

A recommendation becomes stale when the catalog changes, a product sells out, or the season that made it relevant ends. This happens because a static rule preserves the original decision after the context has changed. Manually, review inventory and new products every week, set expiration dates, and maintain a commercial calendar. SalesPilot readjusts as the catalog, inventory, or season changes.

Repetition, diversity, and popularity bias

If recommendations are ranked by sales alone, best sellers take every available space. That creates a loop: what you show sells more, and those sales justify showing it again, while the rest of the catalog loses opportunities. Manually, compare affinity as well as volume, rotate categories and price ranges, and review impression concentration. SalesPilot combines catalog, order, browsing, and seasonal signals instead of relying on a fixed popularity ranking.

Sparse data and long-tail products

Across the long tail of a catalog, many pairs appear only once or twice. That may indicate a real affinity or pure coincidence: two sales are not enough to establish a reliable signal. Manually, require a minimum volume, compare each pair against baseline popularity, and support it with semantic compatibility or curated rules. SalesPilot supplements the store’s own history with patterns learned from more than 1,000 stores across every vertical.

Cannibalization: attribution does not always prove causality

Cannibalization happens when you recommend something the shopper would probably have bought anyway. Order-level attribution is the first filter: it should count the item added through the offer, not credit the entire cart to the recommendation. Even then, tracing the item does not prove what would have happened without the offer. Properly discounting that difference requires a controlled comparison. SalesPilot traces each recommended item to the order containing it; strict causal measurement requires an appropriate comparison design alongside that attribution.

The real cost of running it manually

The initial setup involves cleaning data, analyzing orders, designing offers, implementing events, and testing the full journey. Then comes maintenance: several hours every week to review performance, replace products, monitor inventory, and adapt the strategy to each season.

The workload grows through multiplication, not addition. A catalog with 300 products × 6 shopping moments × 3 segments already creates 5,400 potential product–moment–segment combinations. You will not inspect every one individually, but someone has to identify which ones are outdated. Every new product, stockout, and seasonal change reopens the spreadsheet.

For a small, stable catalog, doing this manually can be an excellent way to learn. As the store grows, that initial exercise becomes a permanent operation: more rules to maintain, more results to compare, and more decisions to revisit every week.

What SalesPilot does with this entire system

SalesPilot, CrossUp's AI engine, turns the ten steps into a continuous process:

• Steps 1 and 2: it analyzes your catalog, order history, and browsing behavior. Its training incorporates millions of interactions from more than 1,000 stores across every vertical.
• Steps 3 and 4: it discovers complements and upgrades, while curated strategies let you set priorities or commercial criteria. Automated and manual approaches work together.
• Step 5: it decides which product to recommend, to whom, at which of the six moments, and in what format.
• Steps 6 and 7: CrossUp implements the modules and traces every dollar of incremental revenue back to the order that generated it.
• Steps 8, 9, and 10: the system learns from each visit's response and adjusts as the catalog, inventory, and seasonality change.

The dashboard shows incremental revenue, AOV, and acceptance. You can explore the thinking behind this combination in AI cross-selling versus manual rules and how AI learns what your customers want.

Manual or automated?

The manual path is useful for learning with a small, stable catalog, a few promotions, and a controlled first implementation. It helps you understand which combinations make sense and which metrics matter.

For most stores already making sales, automation is the natural next step. Install CrossUp, activate SalesPilot, and leave spreadsheet maintenance behind: the AI comes trained on millions of interactions from more than 1,000 stores across every vertical, adapts to your catalog and inventory, and measures from the first order with an accepted recommendation. CrossUp's plans combine a fixed component with a success fee on incremental revenue.

In short

• Generative AI is a powerful tool for analyzing your catalog, proposing pairs, and organizing criteria.
• As the system scales, cold start, stale rules, repetition, sparse data, and possible cannibalization begin to matter.
• Order-level attribution identifies the item added through a recommendation; proving causal incrementality requires a controlled comparison.
• SalesPilot automates the continuous cycle while leaving room for your curated strategies.

The same journey, applied to one industry: how to sell more in an online clothing store and how to sell more in an online toy store.