top of page

Agentic Commerce: What Happens When AI Agents Are the First to Read Your Store?

Writer: Jim Boudreau
Jim Boudreau
Sep 26
9 min read

They read the parts of your store you have never read and may even not know exist.

Not your homepage. Not the photography you invested in. An AI agent goes to the Page Title, the Meta Description, the Meta Keywords, the Brand field and the product feed — the machine-readable layer underneath your store — and it takes what it finds there as FACT.


This month we went through one particular eCommerce site's product catalog, field by field, 7,605 items. What we discovered in those fields is not what the owner believed was written in those fields. Not even close.


So, What Is Agentic Commerce?


An agent is a type of AI that does not just answer, it acts. A shopper tells ChatGPT to find a food-safe cutting board under $75 that ships this week, and the agent searches, compares, checks availability, and hands back a short list. The industry calls this agentic commerce, and almost everything written about it is written for enterprises, but the behavior is quickly translating into the consumer space.


What has happened in the last twelve months is worth knowing, because the shape of it is a necessary lesson for anyone that depends upon their website or online store for revenue. On September 29, 2025, OpenAI launched Instant Checkout — buy it inside ChatGPT — with Etsy sellers live, more than a million Shopify merchants slated to follow, and the Agentic Commerce Protocol open-sourced alongside Stripe.


On March 24, 2026, about six months later, OpenAI stepped back from it: "The initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide, so we're allowing merchants to use their own checkout experiences while we focus our efforts on product discovery."


Notice what got pulled and what did not. The payment rail lasted two quarters. The discovery layer — the agent reading product data and deciding what to put in front of a shopper — did not get pulled. It grew. Adobe measured traffic from AI sources to US retail sites up 393 percent year over year in the first quarter of 2026 alone.


So, while the protocols will keep changing, anyone telling you to bet a quarter on one of them is guessing. Being readable is the part that has to survive every version.


The Agent Reads the Fields You Never See


A human shopper lands on your homepage, notices the design, browses, forms an impression, and only then looks at a product. An agent does none of that. It queries feeds, parses structured data, reads titles and attributes, and assembles an answer. Your photography, your brand campaign, the header you spent eleven weeks getting right — none of it is in the room. It's a classic "black box".


A brightly lit handbag store where most of the merchandise appears translucent and ghostlike on the shelves — present and in stock, but invisible to anything trying to read it.

For a small eCommerce shop that is strangely good news. The beauty contest you were never going to win against a national brand matters less to a machine than it does to a person. What matters is whether your product data is complete, accurate, coherent and current...and that represents an opportunity that has seldom existed in the history of eCommerce.


It also means something less comfortable. A field that no human has looked at since 2014 is now a field that gets read on every single query. And the clock is ticking to fill in the blanks.


What We Found in One Catalog


As mentioned above, we audited a 7,605-product catalog this month — an established store, 16 years old, many hands on it over those years.


Start with the part that makes this so hard to see. By every measure a person would use, that catalog is finished. Every product has a name. Every product has a description. Page Titles are populated on all but one product in the catalog. Meta Descriptions, the same. If you opened it and scrolled for an hour you would say the store was in good shape, and you would not be wrong about what you were looking at.


Now the part nobody looks at.


Seventy-one percent of those products carry no UPC and no GTIN. Ten percent have no Brand Name at all. Set that against Google's own rule for product feeds — a product needs a GTIN, or failing that a brand together with a manufacturer part number — and 1,047 products, fourteen percent of the catalog, cannot be uniquely identified by a machine.


That is the gap that matters, because identity is the agent's first job. When a shopper asks for a specific thing, the agent is trying to match what you sell against everything else on the web that might be the same item. An identifier is how that match happens. Without one, your product is not a weaker candidate. It is an unrecognized object...therefore nonexistent.


Then there is depth. The median product description in this catalog runs 82 words. More than half of it — 56 percent — comes in under 100 words. Thirty-seven percent under 50. Sixteen percent under 25. These are not empty fields, which is precisely why nobody flags them. They are fields with something in them, and not enough of it for a machine to tell your product from a similar one three clicks away.


Five hundred and thirty-five products are both at once: no usable identifier and under a hundred words. For those, an agent has no way to know what the item is and not enough text to reason about it from.


There is one piece of outright damage worth reporting, because it shows how quietly this accumulates. Five hundred and forty-eight products carried a literal "\s" where an apostrophe belonged — "Moose\s", "Santa\s Workshop". Every one of them sat in Meta Description. Zero appeared in any other field. That is encoding damage, the kind an import or a platform migration leaves behind when it meets a curly quote, and a catalog collects it the way a basement collects boxes.


And here is what we expected to find and did not, which I think is the more useful finding. We went in braced for broken outbound links — an authority cited in a product description years ago, pointing at a page that has since moved or died. That is a real failure mode and it is there. It was 51 products out of 7,605. One percent.


Link rot is the tidy story. It was not THE story. The gaps were the story, and they were everywhere.


Why Nobody Caught It


Three reasons, and every one of them made sense at the time.


  1. A gap does not render. A missing GTIN does not show up anywhere on your product page. It does not turn red, it does not break the layout, it does not generate a support ticket. It is a column sitting off to the right of a spreadsheet nobody scrolls to. Some of these fields are not even reachable — on some eCommerce platforms, Meta Keywords cannot be edited at all any more. Nobody proofreads a field they cannot open, and nobody notices an absence that has no symptom.

  2. A mature catalog that has had a great many hands on it. Imports, migrations, bulk edits, spreadsheet surgery, a platform change, years of small corrections by people doing their best with the tools they had. Product lines came in from suppliers who sent what they sent. No single person put it there and no single person could have caught it.

  3. Until very recently, all of this cost you nothing. A human shopper never once asked for your GTIN — they looked at the picture and read two sentences and decided. Google said plainly in 2009 that it does not use the keywords meta tag in web ranking. Nothing depended on any of it. The rational thing to do was nothing — and for the better part of fifteen years, the rational thing to do was also the correct thing.

That is what changed. Not the fields. It's who reads them.

What Does Your Catalog Look Like?

On the surface it's impossible to tell. We see a lot of stores in aggregate, but to do this level of analysis is no small effort. We've looked at a number of catalogs beyond the one we've been referencing...the percentages may vary slightly, but there are always gaps. That is to be expected. To this point humans have been doing the writing, humans have been doing the reading. And the reality is that "busy life syndrome" makes it impractical to maintain this level of attention to detail for every product, catalog or brand.

If your store has gaps as well, whether more or less, you're not alone. In April 2026 Adobe ran a machine-readability check across US retail sites, scoring each type of page on how much of its content an AI model can actually read. FAQ pages came in at 80 percent. Returns and exchanges, 82. Contact Us, 81. Homepages, 75. Individual product pages came in at 66 percent — the lowest of any page type measured.

Retail's least machine-readable surface is the first place an agent goes and is the thing it evaluates with the most scrutiny. Adobe was measuring structure and clarity across whole pages rather than counting identifiers, so that is not our finding restated. But it says the same thing from a different direction: the part of the store built for machines is the part nobody has been tending.

How to Look for Yourself


You can measure your own catalog this week, for nothing, with a spreadsheet. Export everything — not just Description, but Page Title, Meta Description, Meta Keywords, Brand, UPC, GTIN and MPN. Then answer five questions.

  1. What percentage of your products have a GTIN or UPC? Not "do you use them." The percentage. This is one formula and it is the single most informative number in the file.

  2. How many products fail the identifier rule outright — no GTIN or UPC, and missing either brand or manufacturer part number? Those are the ones a machine cannot pin down at all.

  3. What is the word count of every description, sorted ascending? Look hard at the bottom quarter. Anything under fifty words is doing very little work for you now.

  4. How many descriptions are identical to another one in your own catalog, or to the supplier's original? When fifty stores carry the same paragraph, the engines pick one and filter the rest.

  5. Does Meta Description contain encoding damage? Search for literal backslashes, doubled spaces, and words fused together where a dash used to be. Damage of this kind concentrates in one column at a time, and the concentration is itself the tell.

One rule worth carrying into it: gaps travel in families. They arrive by supplier and by import batch, not one product at a time. When you find a product line with no identifiers, check every item that came in with it — that was the most productive hour we spent all month.

The Part Nobody Can Tell You Yet


The measurement layer for agentic commerce is genuinely unsolved. Nobody can tell you reliably how often an agent chose you, or what moved when you fixed something. Anyone selling agentic optimization as a monthly retainer is selling a promise they cannot yet measure, and the honest practitioners in this field say so out loud. Countless $$$ have been spent on the same promise "for $500/mo we can get you on page 1 of Google". Pure snake oil!


Which is the argument for starting where the ground is firm. You cannot verify your agent share. You can count your own GTINs today, and that number is true whatever the protocols do next.


Two more verifiable items belong in the same category, because most stores fail them quietly. Price parity between your pages and your feed: a mismatch is a shrug to a human being and a disqualification to a machine. And currency — Google Merchant Center's own documentation states that "all products expire from your Merchant Center account 30 days after the last refresh." A feed you built in 2023 and have not refreshed since is not a weak feed. It is an absent one.


None of this is a new discipline to go buy. It is the same foundation that earns visibility on every other surface, the practice we call Visibility Optimization. Agentic shopping is one more surface that foundation serves. The silver lining in this cloud – every bit of traditional SEO work that you've done over the years contributes to the same foundation AI Search depends on.


The Uncomfortable Part


The uncomfortable part of all this is not the technology. It is that for fifteen years everyone was rationally correct to leave these fields alone — and the ground moved underneath that decision without announcing itself. Google has never announced to the world that their algorithm has been changed, nor do they tell anyone why or how. Most people find out the hard way – when traffic fades.


In the years since 1998, when I began fighting these battles, the patterns have become familiar. There's always a new algorithm or tool or machine. What they all have in common is that they are all trying to solve a problem for a user. In eCommerce that is to find the product that best suits their need. No different than someone selling shoes face-to-face, the one that does the best job of determining which shoe looks and fits the best wins.


Sources

 
 
CataSEO Logo

CataSEO, a div. of

Studio 1119, Inc.

19 Crestwood Street

Worcester, MA  01605

info@cataseo.ai

(857) 777-1119

US Toll Free (844) 321-1119

  • Facebook
  • LinkedIn
bottom of page
Our website uses an intelligent AI assistant powered by Ultimo Bots to improve customer service.