An FAQ that never guesses

Support tickets that were really the same question

Across the stores we support, a recurring pattern kept showing up: a handful of shipping and policy questions accounted for most of the support inbox, and the FAQ page that should have stopped them was either missing, out of date, or written once and never revisited. At the same time, Shopify started building a Knowledge Base that AI shopping assistants read directly, and a store with no accurate answers in it was invisible to that layer by default.

The obvious shortcut was to point a language model at the store and let it invent a plausible FAQ. We rejected that early. An FAQ is only useful if a merchant can trust it, and a generic AI-written page erodes that trust the first time it is wrong about a real policy.

Three problems that decided the architecture

None of these are visible in a feature list. All three are the reason the app is safe to publish straight to a storefront.

1. Refusing to guess

   The app reads a store's actual shipping zones, policies, local pickup settings, markets and existing FAQ pages, and writes an answer only when those facts support one. A question the data cannot answer is surfaced to the merchant to write, rather than filled with confident-sounding filler. That single rule shaped the whole content pipeline.

2. Tracing every answer back to its source

   Each generated answer carries a pointer to the store fact it came from, so a merchant reviewing it before publish is checking a specific shipping zone or policy line, not taking the app's word for it. That traceability also makes it possible to know exactly which answers are affected when a setting changes.

3. Staying fresh without re-reviewing everything

   Stores change constantly. Rather than re-generate the whole FAQ on a schedule, the app watches the specific facts behind each answer and flags only the ones whose source has actually changed, so a merchant's review queue stays small and relevant instead of growing without bound.

The same three problems turn up in client work

If your business needs AI-assisted content that has to stay accurate as the underlying facts change, the work is the shape above. Reading a system of record correctly, tracing generated content back to its source so it can be checked, and detecting staleness at the fact level rather than re-processing everything on a timer. We have those three in production, under load, with our own name on the support inbox.

What the app does, in short

• Reads shipping zones, policies, local pickup settings, markets and existing FAQ pages
• Writes answers only from those facts, and leaves unanswerable questions for the merchant
• Publishes a help centre page, product page FAQs, and the Shopify Knowledge Base
• Flags an answer out of date the moment the store fact behind it changes
• Translates answers into up to 8 storefront languages

Full detail, every screen and the plans are on the Miko AI Help Center & FAQ product page.

Two screens that show the principle

Interface work on a content tool is mostly about proof. Where did this answer come from, and what happens when it stops being true.