ecomcloser.com

Shopper unboxing a handbag beside a laptop

How an AI shopping assistant works in WooCommerce

An AI shopping assistant for WooCommerce answers questions about the product a shopper is viewing, uses store data to identify the right variation and can offer to add that variation to the WooCommerce cart used at checkout. For a messenger bag, that means checking whether a laptop fits, identifying the chosen colour, showing the current price and availability, and confirming what actually landed in the cart.

In EcomCloser, product chat and store-wide support chat are controlled separately. A question about the laptop sleeve belongs beside the bag; a question about a parcel already in transit belongs in support. The distinction matters because a product description can answer a fit question, but it cannot tell a shopper where their order is. Here is how to follow one question from the product page to a verified variation, a cart action and, when needed, a support handoff.

Start with the question the product page can answer

Imagine a shopper on the Noircraft messenger bag demo asking, “Will my laptop fit in the black one?” The measurements and conversation below are an illustrative example, not a statement of the demo’s published specifications or a record of a live transaction. The question contains two tasks: establish whether the laptop fits, then identify which purchasable option the shopper means. Black is a preference, not evidence that a particular variation is available or an instruction to add it.

Suppose the laptop measures 33 × 23 × 2 cm on the outside. The example product record gives the sleeve’s inner dimensions as 34 × 24 cm and its opening width as 33 cm. The laptop’s footprint is smaller than the sleeve interior, but its width exactly matches the stated opening. There is no stated clearance, and the two inner dimensions do not tell us whether the opening accommodates the laptop’s thickness. Seams, the shape of the opening or a case on the laptop could make the difference. “It should fit a 15-inch laptop” would not resolve any of those questions: the screen diagonal does not describe the outside of the device.

A useful first reply is: “Your laptop is smaller than the sleeve interior, but its 33 cm width matches the stated opening. I can’t confirm it will pass through. Are those the measurements with any case you’ll keep on it?” If the shopper already supplied bare-device measurements and says they do not use a case, do not ask again. The next useful check is whether the store has verified the opening or tested a device of those dimensions, not whether the shopper can repeat the model name.

To continue the example, suppose a staff member checks the black bag with a rigid sample measuring 33 Ă— 23 Ă— 2 cm and documents that it slides into the sleeve and the bag closes normally. The assistant can now report that a sample of those dimensions fit the black version tested. If the store has no such evidence, the honest answer stays conditional. That need not prevent a shopper from knowingly buying the bag, but it should prevent the assistant from calling the fit verified.

Turn the colour choice into a specific variation

Once fit is settled, the assistant should resolve “the black one” against the options on the product page. If colour is the only choice, this may be simple. If the bag also comes in two sizes, black alone is not enough: the assistant needs to identify the size that matches the sleeve measurements or ask the shopper which size they mean. The storefront label helps the shopper recognise the option; the exact WooCommerce variation identifies what to put in the cart.

Check which facts belong to the parent product and which belong to an individual variation. If the store has established that both colours use the same sleeve construction and dimensions, the assistant can carry that fit information across. A fit test documented only for the black bag should not silently become a test of the brown bag. A different size requires its own measurements. Likewise, price and stock should come from the selected variation, not from a parent product’s starting price or a general indication that some version of the bag is in stock.

Now suppose the shopper changes their mind: “Actually, make it brown.” The assistant should show the brown option with its current price and availability and revisit any fit detail that does not carry over. The laptop measurements are still useful context; the earlier black selection is no longer the proposed purchase. If brown is out of stock, the assistant can explain that and show the black variation if it is still available. It should not swap black back in without asking.

Also test less tidy wording. “The dark one” may refer to black or dark brown; a short clarification is better than guessing. “Can I see the brown one?” asks to see an option, not necessarily to replace the selection. If the shopper asks about a second bag, make the product change explicit before reusing any measurements. Remembering context helps only when the assistant also knows which facts belong to which item.

Keep voice selection, cart acceptance and payment separate

A product preview gives the shopper a chance to catch a mismatch before anything changes. It should name the product, show the complete variation and quantity, and report the selected option’s price and availability as checked. A mock preview is illustrative: it can show what a proposed choice would look like, but it is not proof of live stock, a successful cart update or an order. The same applies to a sample transcript or a polished product card.

If an interface lets a shopper use voice, saying “brown” aloud can select the colour just as typing it can. It does not, by itself, accept an add-to-cart action. By contrast, “Add one brown bag” is an explicit cart request if it identifies a single variation; if two brown sizes are available and the size is unresolved, the assistant still needs to clarify. This distinction is about the meaning of the request, not the input method. The example does not assume that the demo offers voice input.

In the walkthrough, assume brown is available and its fit evidence is sufficient. The assistant can ask, “Add one Noircraft messenger bag in brown to your cart?” Once the shopper accepts, the cart action should use that resolved variation and quantity. A brief “yes” is meaningful when it answers that specific proposal. It is not clear enough if the assistant has just offered three alternatives or asked both whether the shopper wants brown and whether they want it added.

Adding to the cart is not payment. A successful cart update means the selected item is in the WooCommerce cart, not that checkout is complete, an order has been placed or money has been collected. The confirmation should say what was added and make clear that checkout comes next. Avoid wording such as “Your bag is ordered” at this stage. Keeping the language precise also helps support: a shopper who later asks about their “order” may only have an item in their cart.

Use the cart result to verify what happened

The assistant’s reply and the cart are two different records. “I added the brown bag” is a claim in the conversation. The cart item is where the store can check the product, selected variation and quantity. During testing, open the cart the shopper would actually use and compare the line item with the option the shopper accepted. For a store owner investigating a mismatch, the variation identifier is more useful than two similar display names.

Test a shopper who already has the same variation in the cart. If they ask to add one more, the result should reflect the quantity change, and the confirmation should make the new total clear. Then try a cart that already contains the black version. Selecting brown in the chat does not remove black from the cart. If the shopper says “replace black with brown,” that is a different request from adding brown alongside it. Check the final cart rather than assuming that a change in conversational selection changed an existing line.

Stock can change between the preview and the accepted action. Suppose the last brown bag sells after the assistant shows it. If the cart does not accept brown, the correct response is that brown was not added. Recheck availability and present the remaining options as new choices. Do not report success based on the earlier preview, and do not add black as compensation. An alternative is a recommendation until the shopper accepts it.

An uncertain result deserves different wording from a confirmed rejection. If the shopper sees no response after accepting the action, check the current cart before asking them to try again. Otherwise a second attempt may increase the quantity when the first already succeeded. Include this check when testing the store as a guest: close the chat, open the normal cart and see whether it reflects the same shopping session. A chat confirmation is of little use if the shopper reaches checkout with a different cart.

Send the right question to support

Product chat is useful when the answer depends on the item on screen: sleeve dimensions, the difference between two sizes, a colour’s availability or the contents of a proposed cart action. Store-wide support is the better place for an existing delivery, a return request or a shopper who explicitly asks for a person. You can operate the two chats separately. Decide who owns an escalation before turning on product chat, especially if a fit question needs a person to measure the stock rather than look up an answer.

Delivery questions illustrate the dividing line. “When would this arrive?” can be a pre-purchase question, but it needs a destination before the store can give a destination-specific estimate. Ask for the information the store needs to estimate delivery, or explain where the shopper can check it. Do not infer a date from the fact that the variation is in stock. “Where is the bag I ordered?” is different: it belongs with support and the store’s normal order lookup process, not with the product description or the shopper’s current cart.

For a fit escalation, a useful handoff would read: “Shopper is considering the black messenger bag. Laptop is 33 × 23 × 2 cm without a case. Sleeve interior is listed as 34 × 24 cm; the 33 cm opening makes fit uncertain. Please verify the opening or test insertion. No item has been added.” If the shopper later chose brown or a cart attempt failed, update that summary. The point of the handoff is to preserve the question, the evidence and the action status so the shopper does not have to reconstruct the exchange.

Keep the same distinction when using Senoji Baldinė as a reference. Its owner’s description is an account of the workflow; the public product listing can verify only the product information and options it actually displays. The listing does not establish what a shopper said, what a staff member did behind the scenes or whether a cart action succeeded. Treat the owner-described process and the public listing as different kinds of evidence, and treat any mock preview of that process as an illustration rather than a transaction record.

Run a five-question pilot before expanding

Start with products that already generate fit, compatibility or stock questions. Pick at least one variable product with reliable specifications and a clearly identifiable unavailable option. Before opening the chat, write down the expected variation, quantity and cart outcome for each test. Run the pilot through the storefront as a shopper would, and record both what the chat says and what support or the cart actually receives.

  1. “Will my 33 × 23 × 2 cm laptop fit?” Use a bag with published sleeve dimensions and compare the reply with the product record. With a 33 cm opening in this example, a pass means distinguishing interior space from clearance and not claiming certainty without a fit check. Repeat with documented fit evidence and check that the assistant identifies what was tested. If a case changes the dimensions, the answer should account for it rather than falling back to screen size.
  2. “Can I choose black, switch to brown, then add one?” Perform the steps in separate turns so you can inspect the state at each point. Neither colour selection should add an item by itself. After explicitly accepting one brown bag, check the cart for the correct variation and quantity. Repeat with black already present to catch an unintended replacement or a mistaken second addition of black.
  3. “Will it arrive by Friday?” Ask before entering a destination. A pass means requesting the missing destination or explaining how to get an estimate, without guessing a date. Then change the wording to make clear this is an existing order. Check that the conversation moves toward support rather than continuing to calculate a date from the product page.
  4. “Can I add the unavailable variation?” Check that the assistant explains the availability problem and offers only options the store shows as available. Inspect the cart to make sure no substitute was added. If the test setup permits, also make a variation unavailable after its preview but before acceptance. The assistant should report the failed action instead of repeating the old stock status.
  5. “Can I speak to a person?” Ask after discussing fit and selecting a colour. Check that the shopper gets a usable route to support and that the receiving person has the relevant product, dimensions, selected variation and cart status. If your setup does not transfer the conversation automatically, verify the alternative: a concise summary the shopper can carry into support is better than a promise of a handoff that never arrives.

Save the wording for both successful and unsuccessful attempts. The exact message preceding an action can explain why an assistant interpreted “that one” as brown rather than black, or why it mistook a request to see an option for consent to add it. Test a clean cart and a cart with existing items. If voice selection is being evaluated, repeat the colour-switch test by voice and check both the interpreted choice and the absence of any cart change before acceptance.

Measure the decision, not just the conversation

A helpful tone is not proof that the assistant selected the right SKU. In the first review, read the transcript beside the cart result and the product data that was available at the time. Separate failures by where they occurred: missing measurements, a wrong interpretation of the shopper’s preference, a mismatch between the proposed variation and the cart line, or a failed action reported as a success. Those failures call for different fixes.

  • Fit grounding: Of the fit replies reviewed, count how many used the relevant published dimensions and identified any missing clearance or thickness information. Track unsupported yes-or-no answers separately from appropriately conditional answers.
  • Variation accuracy: Of the cases where the shopper settled on an option, compare the proposed variation with the intended colour, size and other attributes. Review colour changes and references such as “the first one” separately from simple selections.
  • Cart accuracy: Of the accepted add requests, count how many produced the intended variation and quantity in the actual cart. Log rejected attempts and uncertain outcomes separately, and flag any item added without an accepted request.
  • Recovery: For unavailable or rejected additions, check whether the assistant reported the actual result and waited for acceptance before acting on an alternative. Look at the next cart state, not just the next reply.
  • Handoff quality: Of the requests that needed support, check how many reached a usable support route with enough context to continue. Note whether the shopper had to repeat dimensions or whether support confused a cart item with an existing order.

Use counts alongside rates in a small pilot. Five successful actions out of five tell you exactly what was observed, but they do not tell you how the assistant handles a stock change or an existing cart line unless those cases were among the five. Do not use checkout completion as a substitute for cart accuracy: a shopper can get the right bag into the cart and decide not to pay, or proceed despite an option being wrong.

Fix the weakest step before adding more products

If the assistant repeatedly stops at the opening measurement, fix the product record first. Record the sleeve interior, the usable opening, and any relevant thickness limit or verified fit test. Say which variations the measurements apply to. If the conversation is accurate but the cart contains the wrong bag, trace the accepted variation through to the cart line. Rewriting the fit answer will not fix a mismatch at that step.

If shoppers keep repeating themselves to support, review what the receiving person actually sees. Carry over the unresolved question and the last confirmed action, not just a product name or an unlabelled transcript. Likewise, if support keeps being asked to track orders that do not exist, inspect the cart confirmation for language that implies payment or order placement.

Expand to another product after the initial set can handle an ordinary choice, a changed choice and a failed action without misleading the shopper. The standard is concrete: the fit answer matches the evidence, the accepted choice matches the cart, and the support handoff says what is known and what still needs checking.

Try the EcomCloser demo

Related guides

Sources