Cross-selling in conversation instead of in banners

On nearly every product page there is a box of matching accessories somewhere. It is rarely clicked, and that is usually not down to the range on offer. A banner speaks into a moment in which the visitor has asked no question at all: they are reading the description, checking a measurement, looking for the shipping note. A sentence in conversation does the opposite. It answers something the visitor themselves wanted to know one line earlier. That is precisely the difference between a suggestion that is accepted and one that is read as advertising and skipped. This article describes which add-ons can sensibly be named in chat, where the recommendation has to come from for it to be correct, and where the line runs: no pressure, no invented scarcity, no recommendation without a basis in data.

Enlarge graphic

Labels in the illustration

• Banner on the product page
• See matching accessories
• Sits beside the text, not inside it
• No trigger, no question, no click
• Suggestion in conversation
• Does the mount fit my model?
• Yes, with adapter A2. It belongs.
• Adapter A2, in stock
• Required accessory from the product record
• Required accessory
• Without the part the main
• item cannot be used
• Mandatory catalogue field
• Consumable
• Runs out in use and is
• bought again later
• Linked to the main item
• Compatibility
• Stored as an attribute,
• not guessed from prose
• Otherwise no recommendation
• Where the basis is missing, the assistant names the gap instead of filling it

Add-ons in a banner and in conversation

Why the banner gets overlooked

The accessory box on the product page has a problem that has nothing to do with its content: it sits beside the text rather than inside it. Anyone reading a product page is, at that moment, pursuing a goal they set themselves. They are checking a measurement, looking for the delivery time, comparing two variants. A box at the edge interrupts that process without connecting to it. It is not wrong, it is uninvolved, and uninvolved areas are screened out while reading with remarkable reliability.

On top of that comes a second point that weighs more heavily: the box is the same for every visitor. It does not know the case. Whether someone is seeing the item for the first time or reordering it for the third, whether they own an older device or the current one, whether they already have the required accessory or not: in all four cases the box shows the same tiles. A recommendation that does not know the case can only fit by chance, and a visitor notices that chance. That is why the same item reads as advertising in a banner and as information in a conversation.

The trigger: a question that has just been asked

In a conversation the trigger arises by itself. Someone asks whether a mount fits their model. The complete answer is not “yes” but “yes, with adapter A2, and that is not included”. Here the add-on is not an appendage but the part of the answer without which the information would remain incomplete. Saying only “yes” and leaving out the adapter saves the customer nothing; it merely postpones the disappointment until the day of delivery.

The difference from a banner can be put in one sentence: the banner asserts a need, the conversation answers one that has been voiced. That is why a suggestion in chat needs no advertising language. It manages with the goods themselves: which item, why it belongs, what it does, and whether it is available. This plainness is not a sacrifice of effect but its precondition. How such an item can then be shown as a clickable card is described in Product cards and cart in chat.

Three kinds of add-ons that hold up

Not every item from the same category is suitable as a suggestion in conversation. Three groups hold up, and they have one thing in common: the link to the main item is stored as an attribute in the catalogue rather than being inferred during the conversation. A fourth group appears in practice, the wear and spare part, which behaves technically like a consumable but comes up less often.

Everything else does not belong in chat. The more expensive successor with no discernible reason, the item from another category, the “frequently bought together” collection without a maintained link: such suggestions are at best ineffective and at worst the reason the customer trusts the window less from then on. An assistant that knows three correct items and stays silent on the fourth is more valuable than one that has something to say about everything.

The recommendation comes from the product data

A language model can formulate a plausible set of accessories without knowing a single item from your catalogue. That is precisely the danger. A recommendation in a shop is not a wording task but a lookup: the assistant queries the product record for what is actually linked there and passes on only that. How the assistant picks the right item out of those fields is covered in the article on guided selling in chat.

The difference shows in the wording of the answer. An assistant with a lookup says “adapter A2, in stock, fits your variant” and shows the card alongside. An assistant without a lookup says “a suitable adapter is available in our range”. The second sentence sounds helpful and is worthless: it names no item the customer could order, and it can still be produced when the part does not exist at all. The wiring behind this is a question of tools, described in detail in the article on tool control.

• Accessory link on the item: which parts are linked, and is the link marked as mandatory or optional?
• Compatibility entry per variant: fit belongs on the variant, not on the series, otherwise it is wrong for half the customers.
• Assignment of consumables: which cartridge, which filter, what quantity, and at what interval does the repurchase come up?
• Stock and delivery time of the add-on: a required part that is unavailable is information that belongs in the same answer.
• State of maintenance: a field that is empty on eighty out of a hundred items carries no rule but creates gaps in the conversation.

No recommendation without a basis

The most uncomfortable rule is also the most important: where no link is stored, no add-on is named. Not as a substitute, not approximately, not with the addition “probably a fit”. An assistant that stays silent in this situation, or names the gap, loses a possible additional sale. An assistant that guesses stands to lose the main sale, because the wrong part comes back and the customer doubts everything the window says from then on.

This behaviour can be tested, and it should be tested before an assistant goes live. In our own measurement series 8 out of 8 questions without a basis were declined instead of guessed (XICBOT, own measurement series of 11 September 2026); eight questions deliberately placed outside the available material were tested, and the wording of every answer is in the measurement file. The example questions there concerned contact details that were not on file; for an accessory that is not on file the behaviour is the same. The method and limits of this measurement are set out under measurement series, and the reasoning behind it in declining instead of guessing.

The line: no pressure, no invented scarcity

In chat, pressure has a stronger effect than on a page, because the conversation is read as more personal. A line such as “only two left, act now” is an advertising formula in a banner; in a chat window it reads like information from a member of staff. That is exactly why it does not belong there, unless the stock level really says so. An invented deadline or an invented remaining quantity is not a matter of taste nor venial marketing but expressly prohibited in dealings with consumers.

The Annex to Section 3(3) UWG lists practices that are always prohibited towards consumers, without any balancing in the individual case. For the assistant this means: it may report scarcity only when that comes from stock levels, and deadlines only when a stored campaign supports them. What such a stored campaign list looks like is described in vouchers and discounts in chat. Alongside this, Section 5(1) UWG remains applicable, prohibiting misleading commercial practices in general.

• One suggestion per answer, not three: pushing until something sticks is not selling, it is haggling.
• No follow-up once the customer has said no. A “no thank you” is an answer, not an occasion for a second attempt.
• No urgency without cover: remaining quantities and deadlines only from stock levels or from a stored campaign.
• No suggestion during checkout when the customer is already heading for completion and has asked nothing further.
• No judgement of the customer in the text: neither “others usually take the more expensive model” nor a hint about their choice.

An add-on needs an express decision

As soon as an assistant can not only name an item but include it in a cart it puts together in the chat, advice turns into an act with legal consequences. For this case the German Civil Code states unambiguously how a payment beyond the main service comes about. A silent addition that is already ticked in the cart does not become part of the contract, regardless of how politely the assistant introduced it in conversation.

For building an assistant this yields a simple rule: the add-on only moves into the cart after a separate consent, and that consent belongs visibly in the conversation, not in a default setting in the background. A click on “add” in the product card satisfies this; a pre-set tick does not. How a cart assistant keeps this step properly separate, and what it logs, is among the points that should be signed off before going live.

Before the cart, not after it

Timing plays its part. Across the average of fifty separate cart abandonment studies, 70.22 % of online shopping carts are abandoned (Baymard Institute, as of 22 September 2025). Not every one of these abandonments is down to an add-on, but checkout is the moment when the customer first sees the total in full. A suggestion that appears exactly there adds something to a figure that is being scrutinised. It is therefore accepted less often and read more often as an attempt.

Earlier in the conversation the situation is reversed. Anyone still checking whether the device fits will factor in the required accessory anyway once they learn of it, and is grateful not to have to reorder it. The pointer to a mandatory part therefore belongs in the advisory phase, the pointer to consumables in the answer about usage, and the pointer to an addition where the customer asks for it. How to catch the late phase when it does come to that is described in recovering cart abandonment with chat.

What can be measured

Whether the suggestions work cannot be read off a gut feeling but from four figures that arise anyway: how often a suggestion was made, how often it was accepted, how often the customer completed the main purchase afterwards, and how often the add-on came back. Only the third and fourth figures show whether a suggestion was helpful. A high acceptance rate combined with a high return rate means the assistant persuaded rather than advised.

The reverse direction is useful too: which accessory questions were asked that the assistant could say nothing about? Every such point is a gap in the catalogue and thus a task for maintenance, not a task for the model. Conversation analytics brings these cases together; which evaluations hold up and which merely look good is covered in evaluating conversations. How a recommended add-on appears in the conversation as a card is shown on the product cards page. Which other questions an assistant takes on in an online shop is set out on the page about the assistant for online shops.

Frequent questions