ecomcloser.com

Phone held next to a microphone and headphones

An accessibility checklist for voice shopping on mobile

An accessible voice chatbot for ecommerce lets mobile shoppers complete the same shopping task with or without speaking or listening. Give the microphone, stop control, transcript and cart action clear labels and a keyboard path. A shopper should be able to read a recommendation, check the selected item and quantity, and add it to the cart with the device muted. If microphone permission is denied or the prompt goes unanswered, the existing text path should still work.

The simplest test is to ask for two blue units by voice, then repeat the request by typing. Both routes should show the same item, variant and quantity and offer the same cart control. Keep three events distinct: the shopper selects something, the store accepts or rejects the cart action, and the shopper proceeds through checkout to payment. Hearing “two blue units” tells the shopper what the assistant understood. It does not tell them that the cart changed or that they paid.

1. Build a keyboard path for the four essential actions

Start with the mobile layout and attach a physical keyboard. Open the chat from its launcher, follow the visible focus indicator through the controls, and work both forward and backward. Give each action a visible label and a programmatically available name that matches it. A microphone icon can help people find the control, but it should not be the only clue about what selecting it will do.

  • Microphone: Use a control labeled “Start microphone” or similarly direct wording. Reach it with Tab and activate it with Enter or Space. Verify that the interface distinguishes requesting permission from actually listening.
  • Stop: Provide a keyboard-reachable “Stop listening” control while the microphone is active. If the start button changes into the stop button, verify that focus remains useful and the new name is exposed. Use a separate “Stop readout” control when the assistant is speaking: stopping playback and stopping microphone capture are different actions.
  • Transcript: If the transcript is collapsible, provide a labeled “Show transcript” button. Check that a keyboard user can open it and a screen reader user can read the messages in order. If it is already visible, make it navigable as text rather than making every line a Tab stop.
  • Cart: Put an “Add to cart” button alongside the readable selection. Use the keyboard to reach quantity and variant controls, submit the action and reach “View cart.” Repeat after starting with typed input and check that neither route has fewer controls.

Check where focus goes on open and close. If chat opens as a modal, keep keyboard focus within it while it is open and give the shopper a reachable close control. If it is a nonmodal panel, do not trap focus inside it. On close, return focus to the launcher or to a sensible nearby location if the launcher is no longer present.

Test focus after a message arrives, too. A reply should not pull focus from a shopper who is editing a quantity or reading an earlier message. Look for controls hidden behind a sticky cart bar, a chat header or the on-screen keyboard. A focused button that can be activated but not seen is still a problem.

2. Make captions and text fallback useful on their own

Turn the media volume down before opening chat. Read the assistant’s first recommendation without listening to it. If you cannot tell which item was recommended, why a choice is required, or what to do next, the text is not yet an equivalent path. A generic “See recommendation above” message is no help if the recommendation exists only in audio.

Distinguish the shopper’s recognized words from the assistant’s response and from the current selection. The transcript may say “two blue ones,” but a separate selection summary should name the item, show the blue variant and give the quantity. If the shopper changes blue to green, update that summary. Leave earlier conversation readable without making it look like the old choice still applies.

If you display captions alongside the spoken reply, compare the two for shopping details that affect a decision: item name, variant, quantity, limits and the next available action. Then stop playback and check that the shopper can still read the relevant information. If captions disappear with the audio, keep a persistent text response or summary so the shopper can review it at their own pace.

Example: A commuter asks for a blue insulated bottle, then mutes the phone when boarding a bus. The visible response names the bottle and shows two available sizes in this illustrative flow. The shopper can choose a size, inspect the summary and add the item without hearing “Which size would you like?” A text version that shows the bottle but omits the size choice leaves the transaction unfinished.

Keep the text input in the same conversation. A shopper who types “make that two” after speaking should see which selection the edit applied to. If chat cannot resolve what “that” means, ask a short question in text as well as in audio. Do not send the shopper to a separate help flow just to finish a choice they already started.

3. Test denied, ignored and later-granted microphone permission

Run permission tests with a fresh site permission state. Start with denial: activate the microphone, decline access, and look for a short explanation near the microphone control. Something like “Microphone access wasn’t granted. You can type instead” describes what happened and offers a next step. Keep the existing text input reachable, and preserve any draft or selection the shopper already entered.

Next, leave the permission request unanswered. Depending on the device and browser, the shopper may have to dismiss or leave a browser-controlled prompt before interacting with the page again. Test what happens as soon as the page is available: does chat remain in a waiting state, or can the shopper focus the text input and continue? Do not require the shopper to answer the permission question to regain the shopping flow. Do not display “Listening” while access is still pending.

Check the transition back to text with the keyboard. If the shopper chooses a “Type instead” control, moving focus to the input makes sense. If they have already moved somewhere else, an eventual permission response should not steal focus or erase their draft. Verify that typing can lead to the same item choices and cart action as speech, not just a message explaining why voice failed.

Finally, try granting access after switching to text, where the browser allows that sequence. Define what the interface will do if the original voice interaction is no longer active. A useful behavior to test is requiring a fresh start rather than unexpectedly opening the microphone while the shopper is typing. Check that closing chat or canceling the request leaves no stale “Listening” label when chat reopens.

Record the device, browser, permission state and point at which the text path became available. Separate behavior imposed by the browser prompt from what your page does afterward. A report that says “permission failed” is too vague to tell whether you need to fix the status label, recover focus or restore the input.

4. Use background noise to test corrections, not just recognition

A quiet room is a useful baseline, but it does not show what happens on a noisy trip or in a busy home. Repeat a short shopping request with repeatable background sound at a comfortable level. Keep the intended selection fixed across runs. The aim is not to assign the speech system an accuracy score from a few phrases; it is to find out whether a shopper can spot and repair a consequential mistake.

Choose requests where a plausible mishearing would affect the cart. “Two blue units” could come back as one unit or the wrong color. A phrase such as “not the large one” can reveal whether the interface exposes the negative instruction or silently picks a size. Use products with known variants in a test catalog so you can distinguish a recognition problem from a variant that simply is not available.

When the request is ambiguous, show the alternatives or ask for the missing detail. When the assistant thinks it understood, show its proposed selection and let the shopper correct it. The correction should update the same summary the cart button refers to. A transcript that changes to “two” while the summary still shows one is not sufficient.

Example: In a mock preview, the shopper says “two blue mugs,” and the selection shows two black mugs. They change the variant with a text or keyboard control. The preview should show blue before the shopper adds it. If blue is unavailable in the test setup, report that beside the variant choice and let the shopper decide what to do. Do not silently substitute black.

Track what happened after the mishearing: Was the mistake visible before submission? Could the shopper fix it without restarting voice? Did the accepted cart line match the corrected selection? Record recognition errors and recovery failures separately. Repeating the same noisy request may not produce the same transcription, so include the words actually displayed in your test note.

5. Interrupt a readout and check what survives

Start a spoken recommendation long enough to interrupt. Stop it partway through with the visible playback control, then try the same action with the keyboard. If the experience also allows a shopper to interrupt by speaking, test that separately. Voice interruption is an additional route, not a reason to omit a stop button.

When playback stops, inspect what is still on screen. The recommendation or current selection should remain readable, the cart should not change just because audio stopped, and the shopper should know whether the microphone is listening. If the spoken reply was mid-sentence, do not treat the unfinished audio as the only place where an important option or warning appeared.

Try an interruption while the assistant is reading back a quantity. Say or type “Actually, make it one,” then verify that the selection shows one and that any subsequent confirmation refers to one. If an earlier response finishes processing after the correction, it should not overwrite the newer choice or restart the old narration. Look for a short stale readout as well as a stale visible summary.

Also interrupt when a cart action is in progress. Stopping playback is not the same as canceling a request already sent to the store. The interface should tell the shopper whether the add is pending, succeeded or failed, even if it never finishes speaking. If your interface offers a separate cancellation action, test its actual result rather than assuming it can undo a completed add.

Run one version with a screen reader speaking. Listen for overlapping chat audio and screen reader announcements, then turn off the chat readout and repeat. The shopper should still be able to find the next action and discover the result of an add. Avoid announcing each partial transcript update if it makes the selected item or cart status harder to hear.

6. Confirm the selection, the cart result and payment separately

Use a consistent vocabulary for the three events. “Selected: Blue mug, quantity two” describes a proposed choice. “Added two blue mugs to your cart” describes an accepted cart action. Neither means “Order placed” or “Payment received.” Match the spoken and written confirmation to what actually happened, and keep the written result available when playback is off.

Before the add: show what the shopper is acting on

Place the item name, variant and quantity near the control that submits the cart action. Check that the shopper can edit those details without navigating back through the whole conversation. In a mixed-channel test, speak the initial request, use keys to change the quantity, and submit from the button. The final request should reflect the visible quantity, not the original voice transcript.

After the add: check the store response

For a WooCommerce implementation, use the cart-items interface as a reference for how your integration adds items and reads their quantity and variation. Test against the actual cart, not just the assistant reply. If the add is rejected, explain the problem in text beside the affected control and preserve the selection where it is still valid. If the add succeeds, compare the resulting line item with the requested variant and quantity.

Include a test with one matching item already in the cart. If two more are accepted and the line now contains three, wording such as “Added two; three in cart” distinguishes the change from the new total. Test a slow response and a repeated activation too. If the outcome is unclear, check the cart before encouraging a retry that could add a second set of items.

At checkout: make the handoff explicit

Give the shopper a clear route to view the cart and proceed to checkout. Verify that the interface does not present either navigation step as payment. A keyboard or screen reader test can follow the selected item into the cart and on to the checkout entry without submitting payment. If you need to test a completed order separately, use a test environment and record that as a distinct scenario.

7. Treat product listings, described workflows and mock previews as different evidence

When discussing Senoji BaldinÄ—, keep the owner-described workflow attributed to the owner. Separately, use the public product listing to verify only what it actually displays. For example, copy a product name or an option into a test note only after checking the listing. Do not treat the existence of a listing as proof that voice selection works, that a cart request is accepted, or that a shopper can complete payment.

If you build a mock preview to explain the workflow, label it illustrative. For example, an illustrative furniture selection could show a finish choice and a quantity field so a reader can see where a spoken request becomes an editable selection. That does not establish that Senoji BaldinÄ— offers that finish, that quantity control or the illustrated chat behavior. Use the preview to explain what to test, not as evidence that the live flow passed.

Apply the same distinction to your own release notes. A design preview can show that a stop button was planned. An observed session can show that it stopped playback on a particular browser. A cart check can show that interrupting playback did not alter a specific accepted item. Those are different findings; none should stand in for the others.

8. Run a compact test matrix and measure what shoppers can finish

Use one stable product and a known starting cart for the first pass. Write down the intended item, variant and quantity before each run. Reset the cart and permission state where relevant so one attempt does not change the next. Use a second product only when it exposes something new, such as a required option not present on the first item.

  1. Keyboard only: Open chat, reach the microphone and stop controls, review the transcript, edit a variant or quantity, add the item, view the cart and close chat. Check visible focus at each transition.
  2. Muted: Read the recommendation, make any required choice, and verify the cart result without audio. Check that captions or the persistent text response include the detail needed for the decision.
  3. Permission denied and ignored: Run these separately. Continue in the same text input when the page is available, and note any reload, lost draft or unnecessary return to the microphone prompt.
  4. Noisy: Speak the request, inspect what the assistant understood, correct an error if one appears and compare the accepted cart line with the intended selection.
  5. Interrupted: Stop a readout, change the quantity and complete the add. Repeat with an add already pending and check that a delayed result remains visible.
  6. Screen reader: Check control names, focus on open and close, access to the conversation and selection summary, the cart announcement and the route to checkout. Note where the spoken output was missing or competing.

Measure completion at two checkpoints: whether the shopper reaches the intended selection and whether the correct line appears in the cart. Count wrong variants, incorrect quantities and unintended repeat adds separately. For fallback, record how many actions it takes to reach a usable text input after the page is available. For interrupted audio, note whether the shopper can continue without restarting chat or replaying the recommendation.

Write findings so another tester can repeat them. “With the prompt dismissed on this browser, two Tabs reached the text field but the earlier draft was gone” is more useful than “fallback was confusing.” For a focus issue, identify the control that had focus before the transition and the control that received it afterward. Retest the exact sequence after a fix, then check the other input route for regressions.

9. Use observations to prioritize fixes, not to claim certification

WCAG 2.2 provides criteria you can use to review keyboard operation, focus order and visibility, labels, errors and status messages. Use those criteria to frame review questions, but keep the results of this checklist specific to the flow tested. A muted session that reaches the cart is evidence about that session; it is not a determination of conformance for the site or checkout, and this checklist does not provide legal certification.

Prioritize failures by their effect on the shopper. An unanswered prompt that blocks text, an unreachable stop button, or a wrong variant added despite a correct visible summary can prevent the shopper from controlling the interaction. Fix those before polishing caption timing or rewording a message that already communicates the outcome. Keep a record of untested devices and assistive technology instead of implying that a pass covers them.

Before release, run one final mixed-channel task: start with speech, stop the readout, correct the selection in text, add it with the keyboard and inspect the cart with audio off. Close chat and check where focus lands. If the shopper can keep control through those handoffs and see exactly what the store accepted, you have a concrete usability result to document and a repeatable check for the next release.

Try the EcomCloser demo

Related guides

Sources