Milly Chat in July 2026: Your Store's Search Becomes a Conversation, and the Widget Works for Everyone

July had two headline items and a lot of quiet correction. The headlines are an opt-in setting that turns your storefront's own search into a conversation, and a systematic accessibility pass that found thirteen shopper-facing defects in our own widget. The quiet part is a long run of work on answers the assistant could not back up — less visible, and probably the more valuable half.

For full details on any one feature, see its dedicated post on /writing.

Your store's search becomes a conversation

Turn on one setting and the search bar or search icon your shoppers already reach for opens the assistant instead of a static results page. A typed query becomes a guided conversation that recommends products and answers questions about them.

The reason to care about this one is intent. Search traffic is already the highest-intent behaviour on a storefront — someone typing "waterproof case for a 17 Pro Max" has told you exactly what they want. Meeting that with a conversation instead of a grid is a better use of a moment you are already getting for free.

It is opt-in and fully reversible. Leave it off and your native search behaves exactly as it always has.

The widget now works for every shopper

We put the shopper-facing side of the widget through a systematic accessibility pass — every format, at every viewport, with a keyboard and a screen reader — and found thirteen defects. A keyboard shopper could not finish the quiz. Adding a recommended product to the cart threw focus to the top of the storefront, twice. Replies were announced to screen readers not at all, because the widget had no live region anywhere in it.

All thirteen are fixed and live for every store, with nothing to turn on. The full write-up — including the three layout bugs that were burying storefront footers and hiding theme add-to-cart bars — is in its own post.

Fewer answers it cannot back up

The largest share of July's work went here, and none of it is visible as a feature. The assistant is now much stricter about the line between what it retrieved and what it inferred.

A few concrete examples. It no longer promises a restock for a product that is not coming back. It only calls something popular when your own product data says so. It will not guess a URL on your site it was never given. It reconciles an "in stock" claim against the rows it actually retrieved that turn, rather than carrying a stale one forward. And when a shopper pushes back on a fit or compatibility answer, it re-checks rather than simply conceding.

The honest framing: most of these were cases where a plausible answer was being produced from an adjacent fact rather than the right one. Those are the failures that are hardest to notice, because they read perfectly well.

The product details shoppers keep asking about

A product attribute you configured only reached the assistant if its value happened to be plain text. Lists, numbers and yes/no flags were dropped silently — no warning, nothing in the dashboard. On one live catalogue, four of eight configured fields were reaching the assistant as nothing at all, including a spec list present on more than 800 products, in a store where roughly one shopper question in ten asked about exactly that.

Those value types now render. More usefully, the dashboard now shows you the literal line the assistant will receive for an attribute, sampled from your own catalogue — so instead of trusting that a field is configured correctly, you can read what it actually produces.

Review stars on product cards

Product cards in chat can now carry a star rating and review count, between the title and the price. Ratings come from either the standard Shopify review fields or from Yotpo, and are read when your catalogue syncs — so however slow a review provider is, it never lands in a shopper's wait.

It is off by default. Turn it on per store once your review data is in place.

Gorgias as a support desk

Gorgias now sits alongside Zendesk, Klaviyo and HubSpot as somewhere the assistant can hand a shopper off. Pick it in Lead Capture, connect it, and test the connection before you turn it on.

Tickets arrive already routed — your own ticket tags and custom fields, plus the page the shopper was on and what they were asking about — so your Views and Rules can filter on context instead of parsing the message body.

More of the widget sounds like you

Two smaller branding items. You can now replace the default "Thinking…" wait message with your own rotating phrases, edited directly in Configure, and the timing ramps — a short first beat, then it rotates — so a shopper waiting on a longer answer sees movement rather than one frozen word.

And the display name you set now appears in the banner format's header next to your logo, with image descriptions across all four formats using your store's name instead of a generic "Logo" or "Banner" — which helps shoppers using a screen reader and search engines reading the page.