AI Can Find Your Products. It Can't Buy Them Yet.

Your best-selling product page has a price, a stock status, a brand name, a SKU, and nine photos, and an AI shopping agent will still leave it out of its answer completely, with nothing in your analytics to tell you it happened. Whether an agent can buy from you is decided by three boring fields that most stores have never filled in, and a July 2026 audit of 141 top-selling product pages from 29 major retailers found 70% missing all three. Not ranked lower. Missing. You need three fields, ten minutes with a free Google tool, and one sentence to send whoever built your site.

Passing every checklist you know about is not the same as being buyable

Reza Moaiandin, co-founder of SALT.agency, ran the audit that made this concrete. Writing in Search Engine Journal on July 21, 2026, he described pulling 207 top-organic-traffic product detail pages from 29 retailers using Ahrefs. Fifty-two of those URLs were unreadable to bots outright, returning 403, 404, or status-0 errors. Fourteen turned out not to be product pages at all. That left 141 pages, scored against a ten-point readiness rubric built from Google's Universal Commerce Protocol documentation.

The basics were in excellent shape. Ninety-nine percent of those pages carried price and availability. Ninety-nine percent carried an MPN or SKU. Ninety-six percent carried a brand. If your instinct is that the industry is broadly doing fine on product data, the audit agrees with you right up to the point where it stops mattering.

Then it falls off a cliff. Only 18% carried priceValidUntil, the expiry date on the price. Only 13% carried a shipping delivery time. Only 11% carried a machine-readable return window. Seventy percent of these pages had none of the three.

Inclusion here is binary. There is no page two of an agent's answer, no slow slide down the rankings you can watch in a dashboard and respond to next quarter. Picture a mid-sized outdoor gear store where one jacket carries a big slice of the year's revenue. In organic search, a bad month means that jacket drops from position three to position eight and sales dip. In agentic commerce, it is present or it is absent, and absent means the customer buys the competitor's jacket without ever learning yours existed. That is a different kind of risk, and it is why this is worth ten minutes even while the volume is still small.

The guides are written for engineers, so owners fix the wrong thing

Go looking for answers on this and you land in developer documentation. OpenAI's product feed spec asks for roughly 25 attributes per product and full catalog uploads every 15 minutes. Feed-template guides walk through all 25. Protocol write-ups are aimed at engineers. Every one of them hands you a 25-field spreadsheet and tells you to talk to a developer, which is the wrong answer for the person who actually runs the store. All of it is accurate, and none of it tells a business owner which fields decide the outcome, so the owner either forwards the whole thing to a developer with no priority order or closes the tab and decides this is next year's problem.

Both responses cost you the same thing. Meanwhile the actual gate is three pieces of information you already know off the top of your head: how long the price is good for, how fast it ships, and how many days a customer has to send it back.

There is a second trap, and it is more expensive because it feels like good work. Across the enterprise retailers in the SALT sample, category pages outpulled product pages for organic traffic by roughly ten to one. Barbour is the clean example. Any sane SEO team looks at a gap like that and pours its effort into category pages. Agents read structured product data at the product level, so every hour of that work is invisible to them.

Then there is the plainest failure of all. Thirty-two of the 207 URLs, about 15%, returned HTTP 403 Forbidden to bots. Not obscure stores. Product pages from Adidas UK, UGG, Converse, and Christian Louboutin. The same bot defenses that blocked the audit block Google's and OpenAI's shopping agents, which means somebody's security setting is quietly deciding the brand's commercial future. I have written before about the difference between a site that is safe from agents and one that is ready for them, and this is that distinction showing up on a balance sheet.

An agent does not want your copy. It wants your use-by dates.

Moaiandin put the mechanism in one line that owners should tape to a monitor: "agentic commerce isn't a content optimization problem so much as a data plumbing one." Both protocols, OpenAI's Agentic Commerce Protocol and Google's Universal Commerce Protocol, read your merchant feed, the product catalog file your store sends out to shopping platforms, and your on-page schema, the hidden block of product facts on your page written for machines instead of shoppers. Not your brand story. Not your product description. The plumbing.

Google's UCP documentation, which the audit scored against, names three fields that gate whether Gemini includes a product at all. Translated out of engineer: the price's use-by date, the delivery promise, and the returns window.

You know all three answers right now, without checking anything. That is the frustrating part. This is not a marketing capability you need to build, it is inventory hygiene written somewhere a machine can read it. And the penalty for skipping it is not subtle.

One more field belongs in this conversation. The three fields decide whether you are included. This one decides whether you win the comparison. Sixty-five percent of the audited pages had no GTIN, the barcode-style identifier that lets an agent recognize your product as the same item another retailer sells. Without it, the query "which retailer has the best price on this" ends with an answer that is not you, even when your price is the best one. That is a competitive advantage you are paying for and cannot spend.

Look at what the top scorers actually did.

None of the three switched platforms or rebuilt anything. They reconfigured what they already ran so it would expose the extra fields. Two of the three did it on Shopify, which is to say on software plenty of store owners already pay for.

Google's free Rich Results Test shows you in ten minutes what agents see

Do not start by hiring anyone. Start by finding out where you stand, because on a Shopify or Etsy store the answer is often that a setting and an app close the gap. This is the product-level slice of making your whole site readable to AI agents, and it is the slice with a direct line to revenue.

If the answer to step three comes back defensive or vague, that is information too. I have spent 17 years in search and worked with more than 300 businesses across the US, Canada, UK, Singapore, Australia, and New Zealand, plenty of them ecommerce. Ask for the screenshot of the test result anyway. Ninety seconds of proof beats a reassuring email every time.

Worth separating two fights here, because they get blurred constantly. Getting recommended by an AI is about the decision answers AI reads before it recommends you, the fit, tradeoffs, and comparisons a shopper actually weighs. Getting bought by an AI is about whether the transaction data exists in a format a machine can execute against. You can win the first fight completely and still lose the second one on a missing return window.

Count fields, not clicks, and ignore the referral dashboard this year

Coverage is the number to ask for: how many of your SKUs carry all three fields, and whether bots get a clean 200 response on your top product URLs. That is a number your developer can produce in an afternoon and re-produce monthly. No developer? Spot-check your top ten sellers in the Rich Results Test once a month. Same ten minutes.

Agentic checkout volume today is small, and the platforms themselves are still deciding what this is. OpenAI and Stripe announced the Agentic Commerce Protocol with Instant Checkout in September 2025, letting people buy single items inside ChatGPT. Google announced the Universal Commerce Protocol in January 2026. By March 2026, OpenAI had pulled back from Instant Checkout, saying the first version "did not offer the level of flexibility that we aspire to provide," and refocused its protocol on product discovery with retailers deciding how to handle checkout themselves. Google spent that same month adding shopping carts and catalog capabilities to UCP.

Read that timeline the way an owner should. The checkout mechanics are unsettled, so do not fund a custom checkout integration this quarter. The data requirements are not unsettled at all, and they are the same requirements Google Shopping, comparison engines, and your own returns policy page already imply. Three fields and an unblocked bot is a cheap insurance policy on plumbing you should own anyway. A custom agent commerce build is a bet, and it is early to place it.

Frequently Asked Questions

Do I need to be on Shopify for AI agents to buy from my store?

No. In the SALT.agency audit, the highest scorer was Uplift Desk at 9.0 out of 10 on a custom platform, and two Shopify stores, Carbon38 at 8.6 and Sigma Beauty at 8.0, rounded out the top three. None of them switched platforms or rebuilt their sites. They reconfigured what they already ran so it exposed the extra fields. Shopify and Etsy merchants do get a shortcut, since both are already eligible for ChatGPT's shopping surfaces without a custom integration, so check your settings and apps before you pay anyone to build something.

Is anyone actually buying through AI agents yet, or is this hype?

Volume is still small, and the platforms are visibly still working it out. OpenAI launched Instant Checkout inside ChatGPT with Stripe in September 2025, then pulled back from it in March 2026, saying the first version did not offer the flexibility they wanted, and refocused on product discovery. Google went the other way that same month, adding carts and catalog capabilities to its Universal Commerce Protocol. So do not fund a custom checkout build on this yet. Do fix the three fields, because they cost you an afternoon and they are the same product data your shopping feeds and returns policy already need.

What exactly do I ask my developer or agency about this?

Send them one sentence: do our product pages carry priceValidUntil, shippingDetails with a delivery time, and merchantReturnDays in the schema, and is our bot protection serving 403 errors to Google's and OpenAI's crawlers on product pages? Ask for a screenshot of Google's Rich Results Test on your best-selling product as proof. If the reply is that you already have structured data, that is not an answer, because 99% of the audited pages had price and availability and 70% still missed all three transaction fields. The screenshot settles it.

Most of the effort in ecommerce right now is going into being found. Being found was never the hard part for a store with a decent catalog and a few years of history. The retailer who wins the next few years is the one whose returns window is written somewhere a machine can read it, which is a genuinely stupid thing to be a competitive advantage, and it is one anyway.

Run the ten-minute check this week, fix what it turns up, and get back to running your store. If you want a second pair of eyes on what your product pages are actually handing to the agents, book a short AI visibility call and we will look at your best seller together and decide whether anything here is worth your money.