Milly Chat in August 2026: The Quiz Becomes a Place to Buy, and Answers Show Their Sources

August was the month the shopping surfaces got finished rather than expanded. The quiz stopped being a recommendation screen and became a place to buy. Prices got one format across every surface a shopper sees. Review stars became something you can set up yourself. Underneath all of that, a long run of work went into the assistant knowing what it doesn't know — which products have no page to send a shopper to, which product details it was silently skipping, and which knowledge base entries it actually read before answering.

Each of these has an entry on the changelog.

The quiz becomes a place to buy

Shoppers can now add to cart straight from the quiz results, using the same variant picker the chat cards use, so a product with sizes or colours adds the one they actually chose rather than a default.

The quiz also moves faster. It advances by itself when an answer is picked — one tap per question instead of two. Multi-select questions and the last question still wait for you, on purpose. And the top recommendation now reads as the top one: a featured card, then a labelled break before the runners-up. If the top two genuinely tie, it stays quiet rather than crowning a coin flip.

One guardrail landed with it. A quiz that could send a shopper to an empty results screen can no longer be turned on. If an answer option has no products scored against it, publishing tells you which question and which answer still needs work. On stores with prices switched off, quiz results now respect that too.

Prices that look like your prices

Every price a shopper sees now uses one format: thousands grouped, cents shown only when there are cents. The same product used to render four different ways depending on where it appeared, which meant the sentence and the product card underneath it could disagree. They can't any more.

Prices also carry your store's own currency instead of an assumed dollar sign, so a non-USD catalogue is never quietly mispriced. There is nothing to configure — the default matches how storefronts already write prices — and a store with a different house rule can set one.

Review stars, now yours to set up

Review stars on product cards shipped in July, but turning them on meant asking us. Now it's in Configure: switch them on, pick where the ratings come from, and set the thresholds yourself.

The threshold worth knowing about is the minimum review count. A perfect 5.0 built on four ratings from a review app you stopped using a year ago is exactly the thing that should not go on a product card, and now it doesn't. Set a minimum rating too, and choose whether the card shows the average, a labelled count ("490 reviews"), or both. Products that fail a threshold show no rating row at all. The filtering happens on our side, so nothing that shouldn't be shown ever reaches the page.

Products your shoppers can't buy online

Some products exist in your catalogue but have no storefront page — discontinued lines, retired variants, items sold only through a rep. The assistant used to invent a product link for those and send shoppers to a page that didn't exist. It also described them as out of stock, which they usually aren't. They simply aren't sold online.

Both are fixed. Products with no page get no link, and they're no longer described as out of stock. You can tell the assistant how those items can be ordered, in your own words: a phone number, a support address, whatever your process is. If you'd rather it say nothing about ordering them, that's the default.

Every configured detail, on the product page too

July's update brought list, number and yes/no product fields to the assistant when it searches. They now reach it when a shopper is standing on the product page too, which is exactly when they ask "does this one have X?"

Three smaller things landed with it. A long field is no longer dropped whole; it's included up to its limit and marked as trimmed, so the assistant knows there's more rather than answering from a gap. Fields written in Shopify's rich-text editor are read as text instead of as markup. And references to other entries that can't be resolved are cleaned out instead of taking up a slot with an internal ID. On one live catalogue this took the number of products carrying a given attribute on the product page from 8 to 1,229.

See what the assistant read

Conversation replay now shows which knowledge base entries a reply drew on, in the order they came back. When an answer is wrong, this is the difference between fixing the source and writing a one-off correction. Long documents link to the entry you actually edit, not the fragment that matched, so a fix reaches every part of it.

One honest note on the wording: the panel says "retrieved", not "used". These are the entries that came back for that turn, and the assistant doesn't always lean on all of them.

The widget follows your storefront's privacy settings

On Shopify storefronts, the widget now reads the customer-privacy consent state your store exposes before storing anything on a shopper's device. A shopper who declines still gets the full experience — the chat, the quiz, add to cart — with nothing durable left behind on their browser. Your visitor-based rules and your revenue attribution keep working either way, and a shopper who accepts mid-visit keeps the same identity rather than being counted twice. Automatic for every store; nothing to turn on.

Also this month

A correction that has since been fixed at the source can be marked resolved, with a reason, instead of dismissed, so the Corrections page keeps an honest history. And on localized storefronts, the language hint is now applied only when the shopper's own message hasn't already settled the language.