What changed in the Agentic Commerce Protocol update?

The ACP update pushed agents to be stricter about trust and disambiguation, which means they rely more on structured fields than on vibes. When an agent has to guess what a product is (or which variant the shopper meant), it down-ranks it. That is the whole story.

If you sell online, you have already lived through this pattern. The scoring systems get tighter, the tolerance for messy data drops, and the stores with clean catalogs get a quiet advantage. Same movie. Different screen.

Here is what I see changing in practice when teams talk about "ACP tweaks":

• Identity fields got heavier. Brand + GTIN (or a stable MPN) acts like a trust anchor. Agents want to know you are talking about a real thing.
• Category precision matters more. Vague product types make matching harder, especially for long-tail questions.
• Variant clarity matters more. If size/color/pack count is ambiguous, the agent can not confidently recommend it.
• On-page structured data is checked more often. Agents cross-check feed claims against what your product page says.

None of this is random. It lines up with the direction of structured product data specs and schema standards. If you want a neutral reference point, start with schema.org Product and work backward into your feed fields. Citation: https://schema.org/Product

Which product feed signals matter more to agents now?

Most stores fail this. Agents are putting more weight on fields that reduce uncertainty: identity, category, attributes, and constraints like price and availability. Your description still matters, but it can not rescue a product that is missing the basics.

Think about how an agent answers a buyer question. It has to do three jobs fast: match, filter, and trust. The signals below map to those jobs.

1) Identity signals (match + trust)

• brand (consistent spelling)
• gtin / upc / ean where it exists
• mpn for custom products you manufacture
• sku as your internal handle (less important to the agent, still useful)

Google is blunt about product identifiers in Merchant Center. If you sell branded goods and leave GTIN empty, you are making trust harder. Citation: Google Merchant Center product data specification

2) Classification signals (match)

• google_product_category (or Shopify Product category mapped correctly)
• product_type (your hierarchy, consistent across the catalog)
• condition (new, used, refurbished)

Classification fields are how an agent stops recommending the wrong thing. "Trail running shoe" and "road running shoe" are both shoes. They are also totally different products.

3) Attribute signals (filter)

• size, color, material, pattern
• gender, age_group where relevant
• multipack / bundle / unit counts

These are the fields that map cleanly to real shopper language. When someone asks "black linen shirt, relaxed fit, size medium," you either have those fields or you do not. And if you do not, the agent is guessing. Not great.

4) Constraint signals (filter + trust)

• price and sale_price
• availability (in stock, out of stock, preorder)
• shipping weight and sometimes dimensions

Agents get punished for recommending items a shopper can not buy. Price/availability mismatches are one of the fastest ways to fall out of recommendations.

How do you know if your Shopify feed is agent-readable?

Here is the simple test: can a machine confidently tell what your product is, which variant it is, and whether it can ship today? If any of those answers are fuzzy, you will see it as low impressions, weird matches, or sudden drops after a feed refresh.

I look at three places first. Do this before you touch copy.

1. Google Merchant Center diagnostics. Even if you are not living inside Google Shopping, the diagnostic feedback is a great proxy for what agents can and can not parse.
2. Your product page structured data. Many Shopify themes output Product JSON-LD by default, but it can be incomplete or broken by apps. Shopify has a solid primer on ecommerce schema that is worth reading. Citation: https://www.shopify.com/blog/ecommerce-schema
3. A quick Rich Results Test. Paste a product URL into Google's Rich Results Test and see what it detects. You are looking for Product + offers (price, availability).

First-person aside: I have seen stores chase "AI shopping" for months while their Product schema was missing price. That is like putting up a billboard with no phone number.

How can you fix your Shopify product feed in one afternoon?

Fixes beat theory. If you do the 10 items below, you will clean up the signals agents need to match and trust your catalog. Start with your top sellers and your highest-margin products. Then work outward.

1) Rewrite titles so they carry the variant meaning

Your title is a match engine, not a slogan. A good pattern looks like: Brand + Product + Differentiator. Example: "Acme Trail Runner 2 - Men's - Wide - Size 11 - Black".

• Put the product type in the title if the name does not contain it.
• Put pack count in the title if you sell multi-packs.
• Do not cram every keyword. You want clarity, not a word salad.

2) Set Shopify Product category and keep it tight

If you are on Shopify, use the built-in Product category field when it applies. It gives you a cleaner mapping into downstream channels, and it forces discipline.

3) Build a consistent product_type hierarchy

Product type is your internal taxonomy. Agents use it as a hint for grouping and relevance. Write it like a path: "Footwear > Running Shoes > Trail".

Rule: one meaning per path. If the same product appears as "Shoes > Trail" and "Footwear > Trail Running", you are teaching machines that you do not know what you sell.

4) Fix brand strings and stop the drift

Pick one brand string and enforce it. "Darn Tough" is not "DarnTough" is not "Darn Tough®". Humans do not care. Machines do.

5) Add GTINs for branded products (and stable MPNs for your own)

If a product has a UPC/EAN, put it in. If you manufacture your own products, use an MPN that does not change every time you tweak a title.

One practical move: export your catalog, sort by vendor/brand, and find the rows with blank identifiers. Those blanks are where trust falls apart.

6) Expand thin descriptions into buyer-answer descriptions

Descriptions should answer the question the buyer meant, not the question they typed. Add who it is for, real materials, sizing notes, compatibility, care, and what comes in the box.

If your description is 2 lines and a bullet list, agents have very little to work with. You can still write it clean and short. Just make it complete.

7) Move important attributes into structured fields

Put size, color, material, and other attributes into variants and metafields. Do not bury them in the last sentence of a description.

• Size and color should be variant options.
• Material can be a metafield if it is not a variant.
• Bundle contents should be explicit ("Includes 2 filters and 1 brush").

8) Clean up images and write real alt text

Agents do read images and surrounding metadata. Give them clean inputs: a strong primary image, and enough supporting angles to remove doubt.

Alt text should be descriptive, like: "Black linen button-down shirt, relaxed fit, front view". Not "IMG_3829". Zero. Nothing. Blank.

9) Fix price and availability mismatches

Price and availability need to match between your product page and your feed. If an agent sees "in stock" in one place and "out of stock" in another, it will treat the listing as risky.

• Watch for sale pricing that updates on-site but lags in the feed.
• Watch for preorders that look like in-stock items.
• Watch for currency issues if you run multi-currency.

10) Verify Product structured data on-page (do not assume your theme is fine)

Your feed is one input. Your product page is another. Agents cross-check. If your schema.org Product JSON-LD is broken, you are asking machines to trust incomplete facts.

Quick check: open a product page, view source, and search for "application/ld+json". You should see Product markup with an Offer that includes price and availability.

One use of "instead": do not add more apps when the data is the problem. Fix the catalog fields first.

What should you watch after you fix the feed?

After you make changes, watch for two things: fewer feed errors and better match quality. Agents can take days to re-crawl and re-score, but diagnostics tend to react faster.

• Merchant Center warnings and item disapprovals: you want these trending down, not up.
• Query-to-product match quality: fewer weird matches, fewer "close enough" recommendations.
• Variant selection accuracy: agents picking the right size/color more often.

If you want a fast sanity check, pick 10 products that should match obvious questions and see if your own site data actually contains the words a buyer would say. Not fluff words. Product words.

FAQ: Agentic Commerce Protocol updates and Shopify feeds

What is ACP, really?

ACP is shorthand for how shopping agents evaluate product relevance and trust. You can call it a protocol, a scoring system, or a ruleset. The practical point stays the same: clean identity + clean attributes + consistent constraints win.

Do I need GTINs for every product?

If you sell branded products that have GTINs, yes, you should add them. If you manufacture your own, use stable MPNs and keep brand consistent.

Which three fixes should I do first?

Start with titles, product categories/types, and identifiers (GTIN/MPN). Those changes reduce ambiguity, and ambiguity is what agents punish.

How do I know if my theme schema is broken?

Run a product URL through Google's Rich Results Test and inspect the detected Product markup. If has are missing price or availability, fix that. Agents will notice.