Curated Collection and Curated Responses: Two Levers for Shaping AI Output on Shopify Chat

AI Instructions handle the global shape of every response. Corrections let you patch specific queries the AI gets wrong. Both are reactive and broad — Instructions apply to every turn whether they matter or not, and Corrections only kick in after you've watched the AI miss something.

There's a gap between them — the case where you know exactly what the AI should say or what catalog it should draw from, and you want that control without writing global rules or playing whack-a-mole with corrections. Curated Collection and Curated Responses fill that gap from two different angles. Curated Collection narrows the surface of the catalog itself; Curated Responses overrides the answer for a specific trigger. They're available on different tiers intentionally — scope on Essentials, precision on Core — and most merchants will discover the second after they've been using the first for a while.

Curated Collection: narrowing what the AI can recommend

By default the AI searches your entire product catalog. For most stores that's the right behavior — a customer asking for “something durable for hiking” should be matched against every product that fits, not a hand-picked subset. But there are cases where the full catalog is the wrong scope:

• Seasonal campaigns. Black Friday, a summer sale, a holiday gift drop — you want the AI to recommend from the campaign collection, not from products that are technically in stock but aren't part of the moment.
• Gift guides. A curated under-$50 set, a gifts-for-dad collection, a starter-bundle list. Hand-picked sets where the editorial choice is the value.
• Audience-specific stores. A wholesale-only subdomain pointing at a B2B-priced collection. A clearance landing page pointing at end-of-season SKUs.
• Brand lines or sub-catalogs. A multi-brand retailer running a chat scoped to one brand for a campaign-driven microsite, while the storefront-wide widget keeps the full catalog.

Curated Collection is a per-store setting on Configure → Response. You pick a Shopify collection — manual or smart — and from that point on every AI product search is pre-filtered to that collection's members. Hybrid search still does its job (semantic + keyword), but the candidate set is the collection, not the catalog.

The product list is resolved at save time and cached on the store, which keeps chat latency the same whether the collection has 12 products or 1,200. When the collection changes — products added, removed, smart-rule criteria adjusted — re-save the setting to pull the fresh list. (A webhook-driven resync is on the backlog so this becomes automatic; for now, save is the refresh.)

Curated Collection is available on Essentials and above. It's the kind of feature where the upgrade-from-Starter pitch writes itself: a Starter merchant running a seasonal campaign is suddenly getting product recommendations that ignore the campaign, and there's no way to fix that without scoping the widget. One tier up gives them the control.

Curated Responses: pre-authored answers for known asks

Some questions have one right answer and you want the AI to give it. “What's your return policy?” “Do you ship to Canada?” “What's the difference between your Pro and Standard lines?” — the merchant has the answer in their head, the policy doc, or the brand voice guidelines, and they want the AI to say that answer, not paraphrase from KB chunks and add product recommendations the customer didn't ask for.

Curated Responses lets you pre-author an answer and tie it to a trigger — a quick-question pill or a phrase pattern the AI watches for. When the trigger fires, the curated response takes over the turn. Two options shape the behavior:

• Search restriction. With this on, the AI stops at the curated response — no falling through to product search, no appended product cards. Useful when the answer is the whole turn (returns policy, shipping cutoff, support contact). Without it, the curated response shapes the answer but product search still runs underneath, so the response can lead with prose and then surface products.
• Guidance injection. The curated copy is injected into the system prompt for that turn — so the AI still generates the response, but it generates it with the curated content as must-include context. The result reads natural and on-brand rather than verbatim-canned, which matters when customers ask slight variations of the same question.

Curated Responses lives at its own sidebar entry — separate from Configure — because the workflow is different. AI Instructions are a small set of global rules you tune over months. Corrections are a backlog you work through after observing replay. Curated Responses is a library you build as you discover the questions that benefit from a precise answer — and pruning over time as your KB and brand voice converge.

It's a Core-tier feature. The pairing with the Quick Question editor is intentional: every QQ pill can either send a chat message (the default), trigger a quiz, or now route to a Curated Response. Wired together, a merchant can build an opinionated chat surface where the top entry points each lead to a controlled experience.

When to use which (and when to use AI Instructions instead)

Three controls overlap if you squint at them — Instructions, Corrections, and Curated Responses can all push the AI toward a desired output. The right one depends on how broad the rule is and whether you're steering content or steering catalog.

• AI Instructions — global rules. “Always mention free shipping over $75.” “Default to friendly, first-name tone.” Applied every turn. Use when the rule should fire regardless of the question.
• Corrections — pattern-specific reactive fixes. You saw the AI answer X wrong; you teach it the right answer. The match is fuzzy and you maintain a small set over time.
• Curated Responses — pre-authored answers for deliberate, known triggers. You're not reacting to a mistake; you're front-loading the precise answer for a question you know is coming. Tied to QQ pills or trigger phrases.
• Curated Collection — different axis entirely. Doesn't shape the AI's words; shapes which products the AI can pull from. Use when the surface of the catalog is the wrong surface for the moment.

A practical decision tree: is the rule about every turn? Instructions. Is it about a specific past mistake? Correction. Is it about a specific known trigger you want to answer precisely? Curated Response. Is it about narrowing what products are eligible in the first place? Curated Collection.

The Essentials → Core progression

These two ship at different tiers on purpose. Curated Collection is the more common need — most merchants will hit the seasonal-campaign or gift-guide moment within their first few months on Milly Chat, and the upgrade-to-Essentials path becomes obvious when they do. Curated Responses sits on Core because it's a higher-touch feature — you don't set it up once and forget it; you maintain a library, prune it, refine wording. It tends to be the third or fourth thing a merchant invests in after they've been running Milly Chat long enough to know exactly where their AI's tone needs to be locked down.

That sequencing matters for the broader product story: most merchants adopt Curated Collection first, get a measurable lift from focused recommendations during their first campaign, and only then start thinking about the precision-response use cases that justify the next tier. The two features pair as a natural progression — scope first, precision second.