The Shopify Knowledge Base app: the answers your store gives AI agents

What the Knowledge Base is

The Shopify Knowledge Base is a free, Shopify-built app that holds the question-and-answer pairs AI platforms read when they answer questions about your store. When a shopping agent is asked about your returns policy, your shipping coverage or what the brand is, it reads this feed instead of scraping your About page and guessing. Shopify describes it as a trusted data source for AI platforms, launched in May 2025 and installable from the App Store in a minute.

Two facts frame everything else about it. First, nothing in it displays on your storefront, because it is backend infrastructure read by machines, invisible to human visitors. Second, Shopify is explicit that curating it improves the accuracy of what AI platforms say about your store and does not change how often the store appears. It is an accuracy lever, and on that honest basis it is one of the highest-value hours a merchant can spend on the agentic channel.

How agents read it

Entries are stored as Shopify metaobjects, visible in the admin under Content, and distributed two ways. Neither route involves anyone visiting a page.

• The storefront MCP tool search_shop_policies_and_faqs, which any agent can call against the store's /api/mcp endpoint and receive the question-and-answer pairs as structured JSON. This is the same interface family covered in our piece on the UCP migration, and the policies tool is unauthenticated, so you can read your own feed the way an agent does.
• Shopify's syndication to connected AI surfaces through the agentic storefronts channel, where the same verified facts travel alongside structured product data to ChatGPT, Gemini, Copilot and the other connected platforms.

The part merchants consistently miss is that the feed works before anyone touches the app. Shopify auto-generates answers from store settings, policies and catalogue data, and serves them to agents today, on every store, errors included. Installing the app does not switch the feed on. It shows you what the feed has been saying.

What we found when we captured a live feed

As the baseline step of a recent engagement, we captured a live store's feed by querying search_shop_policies_and_faqs with 28 questions a shopper would ask: returns, exchanges, delivery costs and times, click and collect, gift services, appointments, care, sizing, warranty and payment. The store is a premium brand with a strong team, and nobody had ever opened the app.

Thirteen of the 28 questions returned nothing at all. No answer on refund timing, delivery cost, appointments, ring sizing, resizing, care or warranty, each an empty response an agent has to fill with a shrug or a guess.

The answers that did come back were worse than the silences. The returns answer stated a shorter window than the policy page it was meant to summarise. Click and collect came back as a raw system record, location ID and machine-formatted pickup time included, with nothing a buyer could act on. A gift service was confidently advertised that the store does not sell. Every one of those answers was live, being served to any agent that asked.

None of this is unusual, and Shopify's own documentation is upfront that auto-generated answers are a starting point to review. The finding that matters commercially is that the review step is not optional once agents are in the buying journey, because the difference between an unanswered question and a wrong answer is the difference between a lost sale and a service incident.

The question set to test your own store against

You do not need our tooling to repeat the baseline, because the question set is the method. Ask your store what a shopper would ask it, group by theme, and score each answer as accurate, wrong or missing. This is the set we run, and any store's version differs only in the category-specific block.

Run the set in the Knowledge Base app itself if the store is yours, or through the policies tool on the MCP endpoint to see the response as an agent receives it, which is covered step by step in how to test what a shopping agent sees. Twenty to thirty questions covers a standard premium store, and the scoring takes under an hour.

The three kinds of entry

The third type deserves more attention than it gets. The Top unanswered questions log is a list of things real buyers asked real agents about your store, with no answer available, which makes it a free source of genuine FAQ demand that no keyword tool can see. It complements Search Console the way returns data complements reviews: it shows you the demand you failed to meet.

The five categories to load

Auto-generation covers settings-derived facts at best, so the substance of the work is custom entries. We load every store against the same five categories, typically 15 to 25 entries in total.

1. Brand identity and positioning. What the brand is, who it is for, and what makes it the right recommendation, written as the direct answer to "What is this brand?". This is the single most important entry in the feed.
2. Return and shipping policies. The actual terms in plain English: return window, process, free thresholds, timeframes by region, international coverage. Stated to match the policy pages exactly.
3. Product category guidance. Which collections serve which use cases and occasions, so an agent asked for a gift, a first purchase or a specific situation is steered to the right part of the catalogue.
4. Bestselling and featured products. A direct answer naming the products you want surfaced and one line on why. This is the nearest thing the agentic layer has to a featured-products signal.
5. Brand differentiators. The answer to "why this brand over another", stated plainly, because comparison questions are asked constantly and an agent without your answer will improvise one or concede to a competitor with a clearer story.

A worked entry set for one category

This is the category-guidance block written out, for an illustrative premium footwear brand. The questions are phrased the way an agent receives them from a buyer, and each answer stands alone, names real parts of the catalogue and stays within the two-sentence guidance.

Note what the entries are doing. Every answer steers towards a named part of the range, states a fact an agent can safely repeat, and quietly answers the follow-up question inside the first. Multiply this pattern across the five categories and the feed becomes a sales assistant's knowledge, in a form a machine can relay without mangling it.

Writing answers an agent can relay

Shopify's guidance throughout the app is a brief one-to-two sentence answer, and the guidance is right, because whatever you write gets relayed, summarised or quoted by a consuming agent, and short declarative answers survive that intact.

Write the answer so it stands alone without the question, because that is how it arrives in an agent's context. An answer of "Yes, within 21 days" degrades badly when the question is lost, while "Unworn items can be returned within 21 days of delivery for a full refund" survives on its own. Lead with the fact, keep one condition if one matters, and put everything else on the policy page the entry points to.

And keep the feed consistent with everything else machines read. The returns window in the Knowledge Base, the policy page, the schema markup and the agents.md file must state the same number, because an agent that meets two different answers has learned to trust neither. That single-fact discipline is the working core of discovery-readiness.

Loading it: the exact flows

1. Create an entry: Apps, then Knowledge Base, then Add FAQ. Enter the question, a one-to-two sentence answer, and Save.
2. Fix an auto-generated answer: click the FAQ in Store FAQs, choose Override next to the information source, edit the answer and Save. Return to default reverts it later if the underlying settings change.
3. Answer real demand: in Top unanswered questions, hover over a question, choose Answer, tidy the question wording if needed, write the answer and Save.
4. Delete only manual entries. Auto-generated facts cannot be removed, only overridden, which is why wrong defaults have to be corrected rather than ignored.

Plan the loading around one constraint: there is no bulk import, no CSV path and no API route documented, so entries go in one at a time through the admin at roughly a minute each. A 20-entry set is an unhurried half hour, and the writing beforehand is where the real time goes.

Constraints and maintenance

Three constraints shape the ongoing work. The app is English-only as of mid-2026 according to merchant reviews, which multi-language stores need to plan around. Answers should stay within the brief-answer guidance because long entries get truncated or summarised downstream. And the feed inherits every policy change, so updating the Knowledge Base belongs on the same checklist as updating the policy page itself, on the same day.

The app's own metrics close the loop: it reports how often AI agents request your store's information and which questions they could not answer. Reviewing the unanswered-questions log monthly, and re-reading the auto-generated answers after any settings change, is the whole maintenance burden. It is small, and skipping it is how feeds drift back into misquoting policies.

One date for anyone testing the feed directly: Shopify is migrating the storefront MCP tools to UCP-conforming equivalents, and responses from the old /api/mcp endpoint already carry a deprecation notice pointing at 31 August 2026. The Knowledge Base itself is unaffected, and any scripts or tooling reading the feed through the old endpoint will need updating with the wider migration.

What the Knowledge Base will not do

It will not make your store appear in more AI answers, and Shopify says so directly. Appearance is governed by how the wider models see the brand, by the catalogue data in Shopify Catalog, and the authority signals the platforms weigh. What the Knowledge Base governs is whether the store is represented accurately once an agent is talking about it.

It is also not a customer-facing FAQ page, and merchants who install it expecting a storefront feature find nothing to see. If you want visible FAQs for humans, they live in your theme with FAQPage schema, and the Knowledge Base runs in parallel behind them, stating the same facts to a different reader.

Questions we hear about the Knowledge Base