Conversational design for shopping agents
Plan conversational design around the shopper’s goal. Map questions, catalogue facts, action confirmations and recovery before testing the buying flow.
- Published
- Last updated
Conversational design arranges questions, answers and actions so a customer can complete a task. For a shopping agent, start with the buying goal, then decide what information and permissions each step needs. A friendly greeting cannot repair a flow that recommends the wrong variant or loses the customer’s answer.
Start with the customer’s task
Define the result the customer needs before writing the conversation. “Find a waterproof walking shoe in my size, under $160” gives you a task you can design and assess. “Have an engaging chat” leaves the required outcome unclear.
Conversation design covers the order of questions, the information behind an answer and what happens after a choice. The chat window’s colours and bubble shapes are separate visual decisions. A well-presented window can still ask for a budget twice or mistake a product question for permission to buy.
Write a short task record with the starting request and an observable finish. A useful finish could be a suitable, available variant in the basket, with its price shown. If no product meets the requirements, a useful result could be an explained mismatch and an agreed handoff.
Before scripting, gather your product records and relevant store policies. Identify which basket actions the agent can perform and who receives unresolved questions. Use a simple numbered flow for the task, then add recovery paths at the steps that can fail. Greetings come after those decisions.
Ask only for information that changes the answer
Order clarifying questions by the products they would exclude. Ask about a required use or compatibility constraint before an optional colour preference. Keep information the customer already supplied, even when it arrives earlier than your script expects.
For a walking shoe, size and intended terrain can change the candidate set. A hard spending limit excludes products outside that amount. For an accessory, the exact device model can decide compatibility. Write down which catalogue field each answer changes before adding its question.
An unanswered preference needs a different path from a missing requirement. If colour is optional, show the available colours and let the customer choose later. If device compatibility is unknown, pause the compatibility recommendation and help the customer identify the model. Don’t guess a required value to keep the dialogue moving.
Compare these fictional responses to a customer who has already supplied a $160 limit:
- Incorrect: “What’s your budget?” The agent repeats a resolved question.
- Correct: “I’ll keep the total under $160. Will you use the shoes on paved paths or rough trails?” The answer changes which soles to consider.
Google’s conversation-design principles describe the same problem with a customer who supplies shoe type and size together. The dialogue must retain both answers.
The next best conversation step is the one that removes the uncertainty blocking the customer’s task. It can be a question, a product comparison or an explanation that nothing matches. More questions are useful only when their answers change what happens next.
Ground advice and confirm consequential actions
Tie every factual recommendation to the record that supports it. Product-level availability doesn’t establish that the requested size and colour are available. A general delivery policy doesn’t establish that a particular order will arrive before an event.
Use this source map when drafting the flow:
| Customer needs to know | Supporting record | Conversation consequence |
|---|---|---|
| Whether a shoe is waterproof | Product specification for that model | Use the recorded property and avoid upgrading water-resistant to waterproof |
| Whether the selected shoe is available | Stock for the exact size and colour variant | Offer only the available variant or explain the mismatch |
| Whether the purchase fits the budget | Product price and applicable delivery charge | Compare the amount the customer means, including delivery when required |
| Whether an item can be returned | Store returns policy and relevant product conditions | Explain the applicable rule without promising approval of a particular return |
Keep advice separate from action. “That looks suitable” can accept a recommendation without authorising a basket change. Ask which action the customer wants when their intent is ambiguous. A clear request to add one specified item can supply that permission without another yes-or-no loop.
Before a consequential action, make its scope visible. Show the exact variant and quantity for a basket change. Before payment, show the payable total and delivery choice, then obtain approval through the authorised checkout flow. A basket instruction alone doesn’t authorise payment or a change to an existing order.
Google’s confirmation guidance distinguishes acknowledging understood details from asking for explicit approval. Reflect a stated budget as you proceed. Reserve an approval question for an uncertain instruction or a consequence that needs agreement. Report a completed action only after its result confirms success.
Design recovery and human handoff
Give each failure a response that preserves the customer’s goal. Google’s error guidance calls for responses suited to the turn where the problem occurs. Repeating a generic request to try again loses the detail needed to recover.
Write these branches alongside the main flow:
- If the selected variant is unavailable, retain the required size and budget. Explain which constraint failed, then ask whether the customer wants an available colour or another model.
- If constraints conflict, name the conflict. A $160 maximum and a $180 required model cannot both hold. Ask which requirement the customer wants to change before substituting a product.
- If order access fails, keep the order question unresolved. Use the store’s supported identity or order-access path, or pass the case to staff. Don’t turn a failed lookup into a claim that the order doesn’t exist.
- If the customer rejects a recommendation, ask what ruled it out. Apply that answer to the next search instead of returning the same item with different wording.
A handoff needs the customer’s task, constraints and the unresolved question. Include the products considered, the reason a recommendation failed and any basket action that actually succeeded. Pass only the customer information needed for the case through the retailer’s authorised channel.
Tell the customer what was sent and what remains unresolved. Promise a response time only when the store has a supported commitment. Waiting for staff must not become a promise of stock, a refund or an order change that nobody has approved.
Test one complete shopping conversation
Run a complete task against known catalogue records and inspect the final state. The following is a hypothetical design example. Every product, price, stock position and delivery rule is fictional. It illustrates expected dialogue, not a tested agent result.
Assume the fictional Harbour store has a waterproof Path shoe at $140. Its black size 9 variant is available, while navy size 9 is unavailable. Standard delivery costs $10 to the example customer’s destination. The customer has already confirmed size 9 using this fictional brand’s size guide.
- The customer says, “I need waterproof walking shoes for paved paths. Size 9, under $160 including delivery. Navy if possible.” The agent records each constraint and treats navy as a preference.
- The agent checks the product and variant records. It replies, “The Path shoe is waterproof and costs $150 with standard delivery. Navy size 9 is unavailable. Would black work?” It retains the size and spending limit.
- The customer says, “Yes, add one pair in black.” The agent adds the black size 9 variant once. After a successful basket result, it reports, “One pair of black Path shoes, size 9, is in your basket. The total with standard delivery is $150.” Payment remains a separate step.
- For the failure branch, change stock so black size 9 becomes unavailable before the add attempt. The agent reports the failed addition and leaves the basket unchanged. It asks whether the customer wants another model or help from staff.
- If the customer chooses staff, the handoff carries the paved-path use, waterproof requirement, size 9 and $160 delivered limit. It includes the rejected navy option and the failed black addition. The expected result is a handed-over question with no added pair.
Assess the conversation against the task, using a short acceptance record:
- Completion: the basket contains exactly one requested variant, or the unresolved case reaches staff with the customer’s agreement.
- Unnecessary questions: the agent never asks again for the supplied size, use or spending limit.
- Factual support: the product claim matches its specification, and the $150 total matches $140 plus $10 delivery.
- Recovery: a failed addition leaves the basket unchanged and preserves the constraints for the next step.
Tone and message count can help you edit the dialogue after these checks. Before releasing the flow, repeat the stock-change branch against your own store’s records and confirm the saved basket matches the agent’s reply.
Give your store its own shopping agent
Customers ask in their own words. Your agent finds, compares and checks out from your catalogue, right on your site.