On 8 September, the official @Shopify account posted three lines that most people scrolled past. "Muse from @Meta just dropped and Shopify merchants are already live. When someone uses their Muse for shopping, it searches Shopify Catalog to find the best match. Checkout happens directly in the app." (source)

Thirteen days later, on 21 September, Shopify CEO @tobi announced the deep partnership with Muse to enable agentic checkout with Shop Pay on every Shopify store, and the whole ecommerce internet noticed. (source) Within hours the advice started flowing: rewrite your landing pages for AEO, add FAQ schema, restructure your product pages so AI agents can read them.

Almost all of that advice skipped the important word in the first post. Catalog. Not storefront. Not landing page. The agent is not visiting your site the way a shopper does. If you run a Shopify store doing $20k to $500k a month and you have been told to spend this quarter rewriting pages for AI, this piece is about where the agent actually looks, why the AEO playbook written for blogs does not transfer, and what the two layers of a store need to do differently.

The sentence everyone skipped

The same day, Shopify President @harleyf spelled out the merchant side. "Shopify merchants are already discoverable. If your products are in Shopify Catalog and shipping to the US? You're in. Muse didn't exist yesterday. Today our merchants can already sell through it." (source)

Read those two posts together and the mechanism is clear. Muse does not crawl your storefront and interpret your hero section. It queries a structured feed that Shopify already maintains for every store: the Catalog. Your product title, description, variants, price, images, availability and shipping eligibility, as they exist in your admin. That feed is what the agent searches, ranks and hands to the shopper.

This matters because it changes what "optimising for AI" means for a store. For a blog, the AI has nothing but the page. For a Shopify product, the AI has a database record first and the page second, and Shopify's own posts say it is the record that gets searched.

Why the blog AEO playbook does not transfer

Most of what circulates as AEO advice was written for content sites, and it is worth understanding why it looks the way it does before applying it to a product.

The framework @chris_nectiv posted on 21 September is a good example. The perfectly optimised AEO page, in his version: an optimised title and H1, key takeaways up top, structured content formats, freshness updates, FAQs. (source) It drew over 400 engagements and it is sound advice for an article that needs to be retrieved and quoted by an answer engine.

But notice what it assumes: that the page is the unit the AI reads. For a product on Shopify, it is not.

Google made the broader point itself. Google published its official guide to optimising for generative AI features on 15 May 2026, and as @semrush summarised it, the message was blunt. AI search optimisation is SEO. Content that would not rank in normal search will not appear in AI answers either. And Google explicitly said to ignore llms.txt files, AI-only rewrites, content "chunking" and special schema built solely for generative AI visibility. (source)

@SkomorNick added the number that should reset expectations: in the C-SEO Bench study he cites, getting retrieved at all mattered roughly seven times more than any rewording of the paragraph. "A model cannot cite a page it never retrieved." (source) @jakezward listed what winning AI search really involves, and it runs from technical SEO and crawlability to review platforms and consistent facts, not "a handful of AEO hacks." (source)

Put those together and the translation for a store is simple. For a product, "getting retrieved" is not about the page. It is about whether your Catalog entry contains the facts the agent is matching against. Rewriting the landing page does nothing for a product the agent never pulled from the feed in the first place.

What the agent is actually comparing

Back on 27 August, before the Muse deal, @kurtinc described an AEO tactic his agency was testing, and it is the most specific post on this topic in the whole month. (source)

The setup: Shopify's team told his podcast that buyers now shop with long prompts like "hoodie for walking around SF on a windy summer day." The agent answering that prompt builds a comparison table. And, in his words, "It can only compare facts you actually wrote down."

Standard product descriptions are written for humans. They sell the feeling. They do not say the fabric weight, the wind resistance, the fit, the temperature range. So the agent has nothing to put in the table and your hoodie loses to one whose description happens to include those facts.

His tactic: replace the description Shopify sends to the Catalog with a product metafield written for a language model, full of the plain facts. He says the switch lives in Admin under Sales channels, then Agentic, then Shopify Catalog Mapping. We have not published our own test of that path, so check it in your own admin before planning around it.

The test he suggests is one anyone can run in five minutes. Paste your product URL into ChatGPT, Gemini or Claude and ask: "What data would you need to understand this product?" His warning: "It might surprise you how little it gets from your page."

That question is the whole of AEO for a store, compressed. The agent needs, as @MIKS_ae put it, to understand what you sell, who it is for, why it is better, what customers say about it, and whether the information can be trusted. (source) @WooCommerce asked the same thing in one line: "AI agents are reading structured product data to decide what to recommend and buy. Can they read yours?" (source)

Where the landing page comes back in

None of this makes your product page irrelevant. It changes its job.

@kurtinc shared the numbers this month from a store selling OEM replacement parts, the biggest share of ChatGPT-driven revenue he had seen this year. ChatGPT was 0.7% of traffic and 2.1% of revenue. Those sessions converted at 7.1% against 1.5% site-wide, 66% were first-time customers, nobody used a discount code, and 92% landed directly on a product page. His read: "Nobody is browsing from ChatGPT. They asked 'what part do I need,' got a link to the exact product, and bought it." (source)

That is the shape of an agent-referred visit. The agent has already done the comparison from your Catalog data. The shopper arrives on the product page with the decision mostly made and one job left: confirm that what the agent said is true. Price, the specific fact that won the comparison, the shipping promise, the return policy, the proof.

So the landing page is the second layer, and it has one rule the blog playbook never needed: it must agree with the Catalog. If the metafield told the agent "wind-resistant shell, 340 gsm fleece, true to size" and the page says "cosy vibes for autumn," the shopper who arrived for the facts does not find them, and the agent that sent them learns your data is not reliable. There is no public evidence yet of agents actively penalising mismatched stores, so treat that as a reasonable expectation rather than a measured effect. But the human bounce is real today.

This is where @chris_nectiv's framework does apply, once it is aimed at the right layer. Key takeaways at the top of a product page are the three facts the agent used. FAQs are the objections a shopper has after an agent recommended you. Freshness is your stock and price being current, because the agent will quote both.

The two-layer checklist for this week

Layer one is the Catalog, and it comes first because it is where retrieval happens.

  1. Run the five-minute test on your best-selling product. Paste the URL into an AI assistant and ask what data it would need to understand the product. Write down what it says is missing.
  2. Put the missing facts into the product record. Material, dimensions, weight, fit, compatibility, use case, who it is for and who it is not for. In the fields, not in an image of a size chart.
  3. Check whether your admin has the Catalog Mapping option @kurtinc describes. If it does, decide whether a fact-first metafield should be what the Catalog sends. If it does not, put the facts at the top of the standard description instead.
  4. Confirm shipping eligibility and stock are accurate. @harleyf's condition for being in Muse was "in Shopify Catalog and shipping to the US." An out-of-stock or unshippable product is not a candidate.

Layer two is the product page, and its job is to confirm.

  1. Mirror the Catalog facts in the first screen. The three facts that would win a comparison table go above the fold, in text.
  2. Proof in text, not only in widgets. Review count, rating, guarantee, certifications, written where a model and a shopper can both read them.
  3. An FAQ that answers the post-recommendation questions. Sizing, compatibility, returns, delivery time.
  4. One source of truth. Whoever edits the Catalog fields edits the page the same day.

@stacy_muur called AEO "still an experiment; no proven model that works 100%," and that is fair. (source) The one store with published numbers shows AI traffic under one percent. The reason to do the list above is not that agent traffic is big. It is that every step also makes the product clearer to a human, and the first step costs five minutes.

Run the test on one product today. If the assistant comes back asking for facts you know but never wrote down, you have found your AEO problem, and it was never the landing page.

Sources