Shopify moved its AI agent knowledge base into the admin on October 26, 2026. Policies and contact details now publish to AI agents by default, but scanned facts wait for your approval, and cancellation terms need a fact you add yourself. A 15-minute self-test shows what agents will say.
Two of 88 Shopify stores gave an AI agent a usable answer about canceling an order. The fix for most stores is one sentence they never wrote down.
Picture a shopper on the Sunday before Black Friday. She asked an AI assistant for a gift, it found your product, and she placed the order. An hour later she changes her mind and asks the same assistant whether she can still cancel. The assistant checks what your store has published, finds nothing on cancellations, and gives her a vague answer. She doesn’t email you. She disputes the charge, or she just never buys from you again.
That is the gap this article is about. I first wrote about the Shopify knowledge base in February, when it lived in a separate app. As of October 2026 that app is on its way out, the knowledge base sits inside your admin, and some of it is already talking to AI agents whether you’ve looked at it or not. This is the updated version, with what changed, the one question most stores get wrong, and the self-test I’d run this week.
The Shopify knowledge base now lives in your admin under Sales channels > Agentic > Knowledge base, and it replaces the standalone Knowledge Base app, which Shopify’s October 26 announcement says will no longer be available after January 15, 2027. It holds the non product information an AI agent needs to answer a shopper’s question about your store.
Three details matter more than the rest. First, what publishes by default. According to Shopify’s Help Center page on the knowledge base, facts linked to your admin, such as store policies and contact details, are published as soon as Shopify creates them. Facts Shopify generates by scanning your storefront stay unpublished until you approve them. So your written policies are already speaking for you, and anything else is waiting on you.
Second, the structure. The knowledge base has ten categories: About the Store, Brand and Identity, Company policies, Contact, Gift services, Payment, Returns and warranty, Second hand, Fulfillment, and Additional facts. Additional facts can be a paragraph, a yes or no, or a list of tags, and you can attach a reference URL to each one. The whole knowledge base caps at 1,000 facts.
Third, the migration. Stores that had fewer than 1,000 facts in the old app had them moved over and published. Stores with more than 1,000 did not, and Shopify emailed them to re-add what they want before January 15, 2027. If you were a heavy user of the old app, check that email first. Search Engine Journal’s coverage of the move also flags that published facts are publicly available through the Storefront MCP, so the Shop app and third party AI tools can read them.
Two corrections to my February piece while we’re here. I wrote that the app cost $50 a month on standard plans and that you could run multiple knowledge bases. Neither holds now. There is one knowledge base, it sits in the admin, and Shopify’s help page lists no separate charge for it as of October 2026. If you’re on a custom storefront or the Shopify Agentic plan, you can only add additional facts, and a password protected store can’t be scanned at all.
Cancellation is the question most Shopify stores fail with AI agents, because none of the ten knowledge base categories is built for it and most merchants never write their cancellation terms down. Your return policy usually has a home. Your “can I cancel this after I order” answer usually lives in a support rep’s head or a macro in your help desk.
The best evidence I’ve seen is Bill Hunt’s test of 122 Shopify stores, published in Search Engine Journal on October 14, 2026. He queried each store’s policy tool the way an agent would, between August and September. Of the 88 stores whose policy tool he could reach, 70 gave a usable answer on the return window for an item. Only two gave a usable answer on whether an order could be canceled after it was placed, and by when. Five gave a usable answer on returning used or open box items. Two stores answered the cancellation question with their shipping FAQ.
That spread tells you where to spend your time. Return windows are mostly fine, because almost every store has a written return policy and Shopify publishes it. The failures sit in the questions a shopper asks right after buying or right before returning: can I cancel, can I return this if I opened it, can I swap it. Those are exactly the questions that turn into tickets, chargebacks, or silence.
It matters more than it looks. On episode 489, Nicole Jass, Senior Vice President of Enterprise Strategy at Signifyd, put the shopper side plainly: “I find myself shopping more and more at companies who give me a great return policy, right?” A shopper can only reward a good policy if the agent in front of her can actually say what it is. A generous policy the agent can’t find reads, to her, the same as no policy. You can hear Nicole’s full take on good and bad returns in episode 489.
The fastest way to know what AI agents say about your policies is to ask them yourself, the way a customer would, and write down the answers. Here is the test applied to one store so you can see the reasoning, not just the steps.
This is a composite, built from the kinds of stores I talk with every week, not a real merchant. Call it a women’s apparel brand doing about $800K a year on Shopify. Its return policy page says 30 day returns on unworn items. Final sale items are marked on the product page only. Cancellations are handled by support if the shopper emails within two hours, a rule the founder set and the team follows, but it is written nowhere.
Minutes one to five: open an AI assistant in a private window and ask three questions in plain customer language. “Can I return a dress from this store if I wore it once?” “I ordered an hour ago, can I cancel?” “Is the sale section returnable?” For this store, the first answer comes back roughly right, because the return policy publishes by default. The second comes back vague or empty, because the two hour rule doesn’t exist anywhere an agent can read. The third is a coin flip, because “final sale” lives on product pages, not in the policy.
Minutes six to twelve: open Sales channels > Agentic > Knowledge base and fix the two misses. Add an additional fact titled “Order cancellations” as a paragraph: orders can be canceled within two hours of purchase by contacting support, after which the return policy applies. Add a second fact titled “Final sale items” as a yes or no: sale items marked final sale cannot be returned. Then look at the scanned facts waiting for approval and reject anything that contradicts the written policy.
Minutes thirteen to fifteen: write the same two rules into the return policy page itself, in one plain sentence each near the top. Scans don’t always pick up storefront changes, and Shopify suggests adding important information directly, so the knowledge base fact and the policy page should say the same thing.
My call for this store: do not touch anything else this week. The knowledge base doesn’t need a brand voice essay or forty FAQs before November. It needs the two answers a buyer actually asks after checkout. Ask your three questions again in a few days and log whether the answers changed. That log is your baseline.
Put the facts a shopper needs to decide or to undo a purchase into the knowledge base, written in one plain sentence each, and leave out anything that only exists to sound good. The February version of this article told you to build three pillars: policies, FAQs and brand voice. Two of those still hold. The third needs a smaller role than I gave it.
Policies still come first, and the advice to make them scannable holds. If a support rep can’t find the answer in ten seconds, an agent won’t either. A plain summary at the top of each policy page with the legal detail underneath does most of the work. Our AI commerce scorecard tests whether an agent can quote your policies in two sentences, and it’s a good companion to this self-test.
Specific answers still beat generic ones. Pull the questions that show up weekly in your help desk or support inbox, and turn the top ten into additional facts. Cancellations, open box returns, exchanges, gift receipts and holiday return windows are where I’d start, because they map to the gaps in the SEJ test. Use the yes or no type for anything that is genuinely yes or no. It’s harder for an agent to misread.
Brand voice is the one I’d scale back. The admin has a Brand and Identity category, and it’s worth a few accurate lines about who you are and what you sell. It is not where your conversion lives. An agent quoting your brand personality correctly doesn’t help the shopper who can’t find out whether she can cancel.
On episode 491, I said the AI shopper arrives with “one or two kind of leftover questions. And, you know, the ones that the assistant couldn’t answer from your data.” The knowledge base is how you shrink that list before she ever lands on your site. You can hear the five purchase barriers those leftover questions map to in episode 491.
Stop treating an AI chatbot app as the fix for agent questions until your knowledge base answers the basics, because a chatbot can only repeat what your store has written down. The pattern goes like this: a merchant notices AI traffic, installs a chat app, and the chat app gives the same vague cancellation answer the agent gave, because it is reading the same missing information.
Stop maintaining two versions of the truth. If you have a long FAQ in the old app, a different FAQ page on your site, and a third set of answers in your help desk macros, an agent will find whichever one it finds. Pick the policy page and the knowledge base as the source, make them match, and retire the rest. The old app goes away on January 15 anyway.
Stop pasting legal text into facts. Each additional fact should answer one question in the words a customer uses. Long blocks of terms and conditions are exactly what the February article warned against, and it applies more now that the facts are public.
And stop waiting for a tool to tell you whether this is working. On the same episode I said that one line like free returns until January 15th, placed right next to the add to cart button, “beats a link to a policy page every single time.” The same logic applies to agents. The clearest, shortest statement of your policy wins, and you can check it yourself for free.
Under $500K a year, the 15-minute self-test is the whole job, and you can do it yourself tonight. You probably wrote your policies, you know the unwritten rules, and two or three additional facts will close most of the gap. Don’t build a knowledge base project. Fix cancellations and final sale, and move on.
Between $500K and $2M, this is where the unwritten rules pile up. A support lead made a call during last year’s holiday rush, it became policy, and nobody updated the page. Give one person the knowledge base the same way someone owns your return policy, have them run the self-test monthly, and add a line to your BFCM checklist to rerun it the week you change any holiday return window. If you want a wider read on how AI engines see your brand, the 30-minute AI search visibility check runs the same idea across ten buyer questions.
Past $2M, and especially if you sell in several markets, treat the knowledge base as part of your policy change process. Every time legal or operations changes a return window, a cancellation cutoff or a warranty term, the knowledge base update is part of the same ticket. If you had more than 1,000 facts in the old app, prioritize which ones to re-add by what your help desk data says customers ask, not by what you had before. For the bigger picture on where these agents send buyers, start with our agentic commerce guide for Shopify merchants.
A clean knowledge base improves the answers agents can give about your policies, but it doesn’t guarantee that any agent recommends you, and it doesn’t prove those answers will stay the same next month. Be honest with yourself about three limits before you put this on a slide for your team.
First, the plumbing is moving. SEJ notes that Shopify’s developer docs mark the Storefront MCP as deprecated but still working, and point developers to a newer tool Shopify calls a draft capability. The knowledge base itself is the stable part. How agents reach it may change again, which is another reason to keep the policy page and the knowledge base saying the same thing.
Second, you can’t see the questions. The old app had a query log of customer questions sent to AI shopping assistants. SEJ reports that neither the new knowledge base page nor the agentic storefronts reporting lists those questions. Your own self-test is the closest substitute right now.
Third, a scan is not a source of truth. Shopify says it never changes knowledge base content on your behalf, and it offers scanned findings as suggestions. That is good, and it also means a wrong suggestion approved in a hurry becomes a wrong answer published to every agent. Read before you approve, especially this month.
None of that changes the call. Ask your store the questions your customers ask, fix the two or three it gets wrong, and do it before the first big weekend of November.
The Shopify knowledge base is in your admin under Sales channels, then Agentic, then Knowledge base, as of October 26, 2026. It replaced the standalone Knowledge Base app, which Shopify says will no longer be available after January 15, 2027. The knowledge base holds the non product information AI agents use to answer shopper questions about your store, organized into ten categories including Company policies, Returns and warranty, Fulfillment and Additional facts. It caps at 1,000 facts. Stores with fewer than 1,000 facts in the old app had them moved over automatically. Stores with more were emailed by Shopify and need to re-add the facts they want to keep before the app retires.
The Shopify knowledge base publishes your store policies and contact details to AI agents automatically, because those facts are linked to your Shopify admin. Facts Shopify generates by scanning your storefront stay unpublished until you approve them, and Shopify says it never changes knowledge base content on your behalf. You can publish or unpublish everything at once, a whole category, or a single field. Published facts are publicly available through the Storefront MCP, so the Shop app and third party AI tools can read them. Because your written policies publish without any action from you, the first job is checking that those policies say what you actually do.
Add your order cancellation policy to the Shopify knowledge base as an additional fact, because none of the ten categories is built specifically for cancellations. Open Sales channels, then Agentic, then Knowledge base, add a fact with a clear title such as “Order cancellations,” and choose the paragraph type. Write one or two plain sentences: how long after purchase a shopper can cancel, how to request it, and what happens after the cutoff. Then put the same sentence near the top of your return or refund policy page, since scans don’t always pick up storefront changes. Ask an AI assistant your cancellation question a few days later to confirm the answer changed.
Test what AI agents say about your return policy by asking an AI assistant your own customers’ questions in a private window and writing down the answers. Use plain customer language, for example “Can I return this if I opened it?” or “I ordered an hour ago, can I cancel?” Ask three to five questions covering returns, cancellations, final sale items and exchanges. Compare each answer with what your policy actually says. Anything wrong or vague is a gap to fix with an additional fact in the knowledge base and one clear sentence on your policy page. Rerun the same questions monthly and after any policy change so you have a baseline to compare against.
You do not need the Shopify Knowledge Base app anymore, because the knowledge base now lives in your admin and the standalone app will no longer be available after January 15, 2027. If you used the app and had fewer than 1,000 facts, Shopify moved them into the admin knowledge base and published them, so review them there. If you had more than 1,000 facts, they were not migrated, and you need to re-add the ones you want to keep before the app retires. Use that as a chance to prune: keep the facts that answer questions customers actually ask, and drop duplicates or long legal text.
A complete knowledge base does not guarantee AI agents will recommend your products. It improves the accuracy of what agents can say about your policies, contact details and store information, which removes a reason for a shopper to hesitate or leave. Recommendations also depend on your product data, reviews, pricing and what other sites say about you. The knowledge base is the part you control directly and can check for free, which is why it is worth fixing first. Treat it as removing friction for shoppers an agent already sent your way, not as a ranking lever.
