Shopify just gave every store the same AI file. Customizing it is the smallest thing you should do

In early May 2026, three new files appeared at the root of every Shopify storefront. No changelog entry, no admin banner, no email. Developers spotted them around May 7; Shopify staff confirmed on May 20 that /agents.md is now the primary storefront discovery file, with /llms.txt and /llms-full.txt serving the same content. On May 29, the customization mechanism followed: three optional theme templates that let a store override the defaults.

The AEO industry noticed immediately, and the advice arrived on schedule: the default files are generic, so customizing them is now urgent.

The defaults are generic—that part is true. Whether customizing them deserves the top of anyone’s list is a different question, and the evidence answers it clearly. This post covers what Shopify actually shipped, what AI systems actually read, and the order of operations that follows from vendor-confirmed facts rather than urgency.

What Shopify shipped

Every Shopify storefront now serves a default /agents.md. Pull one from any live store and it reads like an operator’s manual for shopping agents:

• The store name and description, pulled from storefront settings
• Supported Universal Commerce Protocol versions (2026-04-08 stable, 2026-01-23 fallback)
• The agent purchase flow: discover the store’s capabilities at /.well-known/ucp, search the catalog, build a cart, create a checkout, complete it
• Hard rules—agents must not complete payment without explicit buyer consent
• The store’s MCP endpoint at /api/ucp/mcp, rate-limited per IP

/llms.txt and /llms-full.txt point to the same content by default. Beyond the store name, description, currency, and contact details, the text is identical boilerplate across millions of storefronts.

Customization works through the theme editor: add templates/agents.md.liquid, templates/llms.txt.liquid, or templates/llms-full.txt.liquid, and each path serves your version. Miss one, and it falls back to your agents.md template, then to Shopify’s default.

One namespace note, because the industry conversation keeps blurring it: Shopify’s storefront /agents.md is unrelated to the AGENTS.md convention that lives in code repositories and instructs AI coding tools. Same filename, different standard, different stewardship, different audience. Shopify’s file is a commerce manifest. If an article treats them as one thing, that’s a signal about the article.

The evidence on llms.txt

The llms.txt convention was proposed by Jeremy Howard of Answer.AI in September 2024: a curated Markdown map of a site’s most important content, placed at the domain root for AI systems to read. Lumio covered what it does and doesn’t do in March. The evidence since then has hardened.

Google’s Gary Illyes has said Google doesn’t support llms.txt and isn’t planning to. No public statement from OpenAI, Perplexity, or Microsoft confirms using it for product surfacing. And in May 2026, an Ahrefs study found that 97% of llms.txt files received zero AI requests—and where requests did occur, SEO audit tools were the largest source. Software checking for the file outnumbers AI actually reading it.

The systems that demonstrably do read llms.txt are developer-documentation tools—docs platforms that generate it, coding assistants that consume it when pointed at a docs site. That’s a real use case. It is not shopping.

So a Shopify store that hand-tunes its /llms.txt today has a tidier file that, per the best available evidence, almost nothing reads. That’s worth 20 minutes. It is not worth the top of a priority list.

agents.md is different—but Shopify already did the work

The commerce file deserves more respect than the crawler file, because the layer it points to is real. UCP is an open standard co-developed by Shopify and Google, with Etsy, Target, Walmart, Wayfair, Stripe, Adyen, American Express, Mastercard, and Visa among its backers. Agents discover a store’s capabilities at /.well-known/ucp, then transact through the store’s MCP endpoint. Since Shopify’s Spring ‘26 developer edition, any developer can register an agent and call the endpoint self-serve. This is the plumbing AI shopping runs on, and it’s live on every Shopify store today.

Here’s the part the urgency-merchants skip: Shopify runs that entire layer for you. The default /agents.md correctly declares the protocol versions, the endpoints, and the consent rules. An AI agent transacting with a Shopify store doesn’t need the store to customize anything—the protocol works out of the box. Customizing the file adds your brand description, your category map, your policies in your words. Legitimate brand control, worth doing. But it changes what an agent can read about you, not whether an agent can transact with you or which products it picks.

The Instant Checkout lesson

If the last nine months of agentic commerce teach one thing, it’s which layers churn and which endure.

OpenAI launched Instant Checkout with Stripe on September 29, 2025—buy inside ChatGPT, powered by the new Agentic Commerce Protocol. By early March 2026, OpenAI had pulled the in-chat checkout flow, telling press that checkout was “moving to Apps,” which route buyers to the retailer’s own site to finish the purchase. Reporting cited very low in-chat completion; Walmart measured in-chat checkout converting roughly 3x worse than sending the shopper to its own checkout.

What survived the retreat: the data layer. OpenAI’s merchant product feed program is still how products get surfaced in ChatGPT—titles, descriptions, images, price, availability, refreshed as often as every 15 minutes. OpenAI’s product page states it verbatim: “Product results are organic and unsponsored, ranked purely on relevance to the user.” And its developer docs add that recommended feed attributes—“rich media, reviews, and performance signals—improve ranking, relevance, and user trust.”

The checkout mechanism lasted five months. The feed spec is the part every merchant still uses. Transaction layers are being redesigned in public; product data compounds through every redesign.

What actually gets products chosen

Across ChatGPT, Google, Perplexity, and Copilot, the vendor-confirmed picture is consistent:

1. Feeds are the entry ticket. OpenAI surfaces products from merchant feeds. Google’s AI shopping surfaces read the Shopping Graph, fed through Merchant Center. Microsoft says MMC feeds “help inform organic Copilot results”. Perplexity aggregates brand-supplied feeds and verified commerce sources, and states brands cannot pay for placement.
2. Structured data on the page is the feed’s twin. Google can crawl Schema.org product markup and generate the feed from it—price, currency, availability, and condition are required for automatic updates. Perplexity reads Schema.org data off the live page. The markup and the feed describe the same products; when they disagree, both lose.
3. Crawl access is the precondition. OAI-SearchBot, PerplexityBot, and Google’s crawlers have to be allowed in before any of the above works.
4. Discovery files are hygiene. No shopping provider confirms them as a ranking input. They’re the business card, not the inventory.

Notice what’s absent from the confirmed list: everything the loudest advice is selling this month.

The order of operations on Shopify

For a brand deciding what to do with an afternoon:

1. Fix the product data itself. Vague titles, missing attributes, absent GTINs, stale availability—this is what agents stumble on, in feeds and on pages alike. It’s also the slowest work, which is why it should start first.
2. Get the feeds flowing. Google Merchant Center at minimum; OpenAI, Perplexity, and Microsoft as your channels warrant. The same clean product data serves all four.
3. Check robots.txt. 2 minutes. Blocking OAI-SearchBot or PerplexityBot undoes everything upstream.
4. Then customize the discovery files. About 20 minutes in the theme editor: add the three Liquid templates, state your brand, your categories, and your policies in your own words, and keep Shopify’s agent-transaction instructions intact—that content is operational, not decorative.

Step 4 is real, and now that every store serves these files, a generic one is a small missed opportunity. But it’s step 4. Any advice that puts it at step 1 is selling you the file because the file is easy to sell.

The file at the root of the store was never the product. The catalog behind it is. Lumio scores that catalog the way AI agents read it—and fixes what they can’t.