GigaCommerce

How to Train a Shopify Knowledge Base Brand Agents Will Quote

Train Shopify's Knowledge Base so Brand Agents quote your real return and shipping rules. Override generated FAQs, then get your free readiness score.

Sujan BhuiyanFounder, GigaCommerce12 min read
AGENTIC COMMERCEGigaCommerce · Insights

A Shopify Brand Agent quotes your return window, your ship-to countries, and your warranty from Shopify's Knowledge Base: a free admin app that shows generated FAQs from your settings and policies, logs what buyers ask, and lets you add custom FAQs. Training it means overriding the drafts that are wrong, writing the exceptions your support team already answers, and proving the quote against the policy URL before you leave an AI channel on.

This is the working method for operators in the $100K–$10M band — the policy half of a Brand Agent install. Catalog attributes and Copilot Checkout stay separate jobs. For a scored read of the whole stack before you edit copy, start with a free Agentic Commerce Readiness Score.

What the Knowledge Base actually feeds

Shopify's Winter '26 Edition (announced December 10, 2025) told merchants to track policies, FAQs, and brand voice in the Knowledge Base app so agents have answers in every conversation. The app listing, from Shopify itself, is narrower and more useful than the keynote line. It does three things: it shows Shopify-generated FAQs based on your store's settings and policies, it lets you monitor buyer inquiries, and it lets you create custom FAQs. It launched May 16, 2025, it is free, and it runs inside Shopify admin.

That is the surface your team can edit. Under it, storefront agents can call a public tool, search_shop_policies_and_faqs, on https://{your-store-domain}/api/mcp. The call takes a required query (the shopper's question) and an optional context (the product they are looking at, or their situation). The response is an answer plus a sources list: titles and URLs, typically a refund policy and an FAQ anchor. Training is finished when that response is right, not when the FAQ row looks tidy in admin.

Shopify Knowledge Base
The Shopify-built admin app where you review generated FAQs, read buyer inquiries, and publish custom FAQs that AI shopping agents use when they answer policy, service, and brand questions.

Shopify's June 18, 2026 guide, "Agentic Commerce on Shopify: How It Works," puts a number on the urgency. In Q1 2026, AI-driven traffic to Shopify stores grew 8 times year over year, and orders from AI-powered searches increased nearly 13 times. New buyers placed orders through AI channels at nearly twice the rate of other channels. Those shoppers still ask whether sale items are returnable, and the sentence they hear is this FAQ layer.

Year-over-year growth in AI-driven traffic to Shopify stores in Q1 2026

Shopify, "Agentic Commerce on Shopify: How It Works," updated June 18, 2026

Agentic Storefronts syndicates eligible products through Shopify Catalog to ChatGPT, Microsoft Copilot, and — for select shops — AI Mode in Google Search and the Gemini app. Shopify's June 2026 guide puts the controls under Settings → Sales channels → Agentic Storefronts. ChatGPT checkout, for merchants selling to US buyers, completes on your online store in an in-app browser. Eligible merchants can complete inside Microsoft Copilot via Copilot Checkout and the Universal Commerce Protocol. You remain the merchant of record. The Knowledge Base is what those agents read for "can I return this if I opened it."

A generated FAQ is not your policy

The app generates answers from settings and catalog data, then lets you customize them. Reviews on the app listing through July 2026 describe invented shipping coverage and invented URLs. Read every generated answer against the policy page before you treat the app as trained.

Fix the pages the tool will cite

The policy tool returns URLs. On a healthy store those URLs are your refund policy and a specific FAQ section. On an untrained store they are the theme-default page: a 30-day window you already changed to 14, or a "worldwide shipping" line from before you cut the country list to two.

Do this pass before you write a custom FAQ. A careful Knowledge Base row still loses if the tool keeps citing the stale policy page.

  1. 1

    Export the live policy set

    Open Settings → Policies and copy the current refund, shipping, and privacy text into a working doc, with today's date at the top. If a lawyer or ops lead owns the wording, this export is the only text you are allowed to paraphrase.

  2. 2

    List the exceptions in one table

    One row per exception: sale items, final sale, opened beauty, international, wholesale, gift cards, subscriptions. Columns: rule, window in days, who pays return shipping, refund destination, and the URL that states it. If the exception lives only in a support macro, it is invisible to the agent.

  3. 3

    Confirm ship-to countries in settings

    Open Settings → Shipping and delivery and write down the active zones. The FAQ is not allowed to name a country that is not in that list. This is the mismatch merchants keep finding: the generated answer says you ship worldwide because an old policy page said so.

  4. 4

    Open each cited URL in a private window

    Search your storefront for the refund policy and the FAQ page. If the policy tool's example shape is your target, you want a source like /policies/refund-policy and a source like /pages/faq#sale-items. A 404, a password page, or a duplicate FAQ in the theme footer is a failed source.

Worked example, labeled as a scenario: a skincare brand around $3M ships only to the United States and Canada, with 30 days on unopened product and 14 days on sealed sale items. The generated FAQ says worldwide shipping and one 30-day window. Support has answered the sale-item exception for two years inside a helpdesk macro. The training job is to make the policy page, the FAQ anchor, and the Knowledge Base row state the same 14-day rule, and to remove "worldwide" from every URL the tool can cite.

That alignment is the first hour. Failure modes of a bad agent install usually start on the retrievable page, while the real rule still lives in someone's head.

Override generated answers, including the ones you cannot delete

Start with the FAQs Shopify already generated. Those rows come from your settings, so agents reuse them first, and they stay wrong until someone reads them.

Merchant reports on the app through mid-2026 say generated questions are not freely deletable, and some stores see them stuck in English even when the shop's default language is not English. When you cannot delete a bad question, override the answer and put the correct statement on the policy URL the tool cites. An override that leaves the policy page wrong loses the next time an agent prefers the page.

Use a four-line card for every override:

  • Question — keep Shopify's wording if you cannot edit it. Add your own custom FAQ beside it in the shopper's language if the generated question is English-only and your buyers are not.
  • Answer — one or two sentences. Lead with the rule, then the condition. "Sale items can be returned within 14 days if the seal is intact. Refunds go to the original payment method."
  • Source URL — the exact policy or FAQ anchor a human can open. If you do not have a URL, you do not have an answer yet.
  • Last checked — a date. Policy copy rots. The card without a date becomes the next generated lie.

Add a voice line on the same card. Shopify's Winter '26 note groups brand voice with policies and FAQs in this app, and the useful form is a constraint list. A clinical skincare brand can allow "fragrance-free" and "patch-test first," and can allow "dermatologist recommended" only when that claim is on the PDP. A furniture brand can state lead time in weeks, and can say "in stock" only when inventory says so.

Match the sentence the policy page already uses

Agents and shoppers both compare the chat answer to the page. If the FAQ says "14 days" and the refund policy still says "30 days," you have trained a dispute. Change the page first, then the FAQ, in the same sitting.

Custom FAQs cover what generation missed. Pull them from buyer inquiries in the app, the last 90 days of tickets, and on-site searches for return, shipping, warranty, discount, and subscribe. Conversation design sorts product questions — comparisons, constraints, suitability. This list is store policy. Keep the two lists apart so a flow workshop does not crowd out the refund exception.

A practical first custom set for a mid-market store is ten questions, not fifty:

  1. 1What is the return window on full-price items?
  2. 2What is the return window on sale items?
  3. 3Are final-sale items returnable?
  4. 4Who pays return shipping?
  5. 5Where does the refund land?
  6. 6Which countries do you ship to, and which do you not?
  7. 7How long does delivery take to the primary zone?
  8. 8What does the warranty cover, and for how long?
  9. 9Can this discount combine with another offer?
  10. 10What is your policy on opened, used, or hygiene-sensitive goods?

Publish these ten once they match the pages. If question 9 has no written rule, write it with finance before the FAQ goes live. An agent will otherwise invent a stacking rule and state it as policy.

Test the quote before the channel stays on

Shopify's June 2026 guide says eligible products sit in Shopify Catalog by default, so Agentic Storefronts can already be syndicating while you read FAQs. Check the toggles first. Copilot, Google AI Mode, and Gemini let you turn direct checkout on or off per channel. ChatGPT checkout runs on your store in an in-app browser, with a separate control for whether order data is shared. An untested Knowledge Base is already in front of US shoppers when those toggles are on.

Call the policy tool the way an agent does. Record the shape Shopify documents: answer, plus sources[] with title and url.

  1. 1

    Ask the plain question

    Query: "What is your return policy for sale items?" Pass condition: the answer states your real window, your condition (seal, tags, unworn), and the refund destination. Fail condition: a single blended window, or a window copied from Shopify's documentation example rather than your policy.

  2. 2

    Ask it again with product context

    Pass the same question with context set to a sale SKU a shopper would actually be viewing, for example "Customer is looking at the marked-down 50 ml serum." The answer should not widen the window because the context is a specific product. If the sale rule depends on a tag, the answer should follow the tag.

  3. 3

    Ask the country question

    Query: "Do you ship to the UK?" If the UK is not a zone, the answer is no, plus the countries you do serve. An answer that says yes, or that hedges with "we ship worldwide," fails even if the rest of the FAQ looks polished.

  4. 4

    Check sources, not just the sentence

    Open every URL in sources. Each one must load, must be your domain, and must contain the fact the answer used. An invented URL is a fail even when the sentence happens to be right. Merchants have already reported that failure mode on the app.

  5. 5

    Record the miss and override the same day

    Write the question, the wrong answer, the URL you expected, and the FAQ row you changed. Re-run the same query. A pass you did not re-run is a note, not a pass.

The endpoint is your live domain, so run the ten questions before you announce anything or enable another channel. Shopify's tool docs use "What is your return policy for sale items?" as the sample query, and the sample payload happens to mention a 14-day sale window. That number is their example, not your rule. Use your window.

When the tool and the Knowledge Base disagree, follow the tool. That response is the contract agents call. Fix the surface it cited, then re-run the same query.

What a trained FAQ still does not fix

Three gaps stay open after every FAQ card is dated.

Catalog facts. "Is this fragrance-free?" depends on the SKU, so it belongs in a structured attribute. A brand-level FAQ is right for one product and wrong for the next. That pass is catalog enrichment.

Price and availability. Shopify Catalog is meant to reflect admin price and inventory across AI channels. A compare-at price, an expired discount, or a sold-out variant is an offer-sync fix on the product, not a new FAQ row.

Checkout rules. A FAQ that says codes do not stack does not configure checkout. Turn direct checkout on only after the policy test passes and someone has completed a test order to a country you actually ship to. The Copilot Checkout guide covers tax, rates, and payment edge cases.

Keep the ten-question log next to AI-channel order attribution in admin. Attribution shows the channel sold. The log shows what the tool answered on the day a shopper says the chat promised 30 days. Re-open generated FAQs the day a shipping zone or return window changes, and at least monthly if nothing obvious moved. Agentic Storefronts can answer off your domain even when the PDP has no chat widget.

Get a free Agentic Commerce Readiness Score

Ten minutes. See whether policy, catalog, and checkout are actually ready before you turn another AI channel on.

If the score shows the FAQ layer is the gap, the method above is the work. If it shows the catalog is the gap, fix attributes before you spend another hour on wording. The install that ties both together — Brand Agent, checkout, and the policy set — is Agentic Commerce Setup.

Frequently asked questions

What is Shopify's Knowledge Base in an agentic setup?

It is Shopify's free admin app for the FAQs AI shopping agents use. You review FAQs generated from your store settings and policies, you see what buyers ask, and you add custom FAQs. Storefront agents can retrieve those answers through the `search_shop_policies_and_faqs` tool on your store's `/api/mcp` endpoint, which returns an answer and the source URLs it used.

Should I trust the FAQs Shopify generates automatically?

Read them, then override anything that disagrees with the live policy. The app is designed around generated drafts plus your edits. Merchants using it through 2026 have reported answers that invented shipping coverage and URLs. A generated row is a starting point you either confirm against a policy page or replace.

Can I delete a generated question that is wrong or in the wrong language?

Do not plan on deletion. Merchants have reported that system questions cannot be deleted and can stay in English even when the store's default language is different. Override the answer, publish a custom FAQ in the right language, and make the policy URL itself state the correct rule so a citation of that page is still right.

Does a trained Knowledge Base mean my products will be recommended?

It means policy questions can be quoted from your pages. Product recommendation still depends on Shopify Catalog, attributes, price, and availability. Train the FAQ layer and the product data as two jobs.

How often should we re-test policy answers?

Re-run the ten-question set the same day you change a return window, a shipping zone, a warranty, or a discount-stacking rule, and at least once a month if nothing obvious changed. Record the answer and the source URL. A FAQ without a last-checked date drifts back into the generated draft you already rejected.
SB

Sujan Bhuiyan

Founder, GigaCommerce

Founder of GigaCommerce, part of Gigaverse. Works with mid-market Shopify and Amazon merchants on agentic commerce installs, AI-ready catalogs, and Commerce GEO.

The weekly brief

Get the weekly DTC + Agentic Commerce brief.

One email a week on what shipped in agentic commerce and the move to make. No fluff.