Your Product Page Is Being Read by an Agent, Not a Customer: What Small Retailers Should Fix First

Note: The Salesforce agentic-search figure below was covered in small-business press on August 9, 2026, and the AI-referral-traffic figure on August 24, 2026 — both current as of this writing. The protocol landscape in this space has changed roughly quarterly for two years; verify anything protocol-specific before building against it.

You spent an evening writing the description for your best-selling product. It’s good. It has the story of where the thing is made, why the material matters, the small detail nobody else mentions.

Increasingly, nobody reads it.

Not because customers don’t care — because the customer’s first step has moved. Use of agentic search as the opening move in a shopping journey grew 200% year over year, according to Salesforce’s State of Commerce research. 67% of shoppers say they’ve used an AI tool — Gemini, Perplexity, ChatGPT — for a purchase in the last three months, rising to 80% among Gen Z. Retailers are reporting AI-source traffic up 1,200% while traditional search traffic fell about 10% year over year.

The uncomfortable implication for a small merchant: before a human ever sees your page, something else decided whether to show it to them. And that something does not care how the description reads. It cares whether your product has structured, machine-readable attributes — because a product without them doesn’t get recommended, no matter how well written it is.

This is not a “get ready for the future” article. It’s a list of unglamorous fixes, most of which take an afternoon.

What actually changed

The old shape of a purchase: search a keyword, scan ten blue links, click three, compare in tabs, buy.

The new shape, for a growing share of purchases: describe the problem to an assistant, get three to five candidates back with reasons, click one, buy.

That’s a different competition. In the old shape you competed for rank against pages targeting your keyword. In the new one you compete for inclusion in a shortlist assembled from structured data, and the assembling is done by something that reads specifications faster than prose and treats an unanswered attribute as a reason to pick someone else.

The scale is worth being sober about. Only 28% of commerce organizations use agentic AI today, though another 44% plan to within six months. J.P. Morgan estimates agentic commerce could reach up to 25% of U.S. online sales by 2030. So: not most of your revenue today, plausibly a quarter of it inside five years, and — critically — already the first touch for a large share of shoppers even when the purchase completes elsewhere.

That last part is the one small merchants keep missing. You don’t have to sell through an agent to lose the sale to one. If the agent doesn’t shortlist you, the customer never reaches your beautifully written page at all.

Your copy is now the second thing that matters

Here’s the reframe that makes the rest of this article obvious.

For twenty years the hierarchy was: copy first, data second. Write a page that persuades a human, then fill in the boring fields if you get around to it.

The hierarchy has inverted. Attributes decide whether you’re considered. Copy decides whether you convert once you are. Both still matter. But they’ve swapped order, and almost every small merchant’s effort is still allocated to the second one.

Product data quality has quietly become a revenue driver rather than a technical chore. Products with rich, structured attributes and accurate, current inventory data get recommended more frequently. Products without them are invisible to the layer that now runs first.

The three layers, in the order you should fix them

Layer 1: Availability truth

Start here, because it’s the one that gets you removed rather than merely ranked lower.

If an agent recommends your product and the customer arrives to find it out of stock, that’s a bad outcome for the agent’s operator, and systems that produce bad outcomes get demoted. Inventory accuracy stopped being an operations metric and became a marketing one.

For most small retailers this is the genuinely hard part, because stock lives in three places that disagree — the register, the storefront, and the marketplace listings. We’ve written both halves of this: what actually goes wrong across eBay, Amazon and Facebook and how to sync inventory across three platforms without a full ERP. If your three systems drift, no amount of schema markup saves you.

Practical minimum: your website’s stock status should be wrong for less than an hour, not less than a day.

Layer 2: Structured attributes

This is the schema.org work, and it is boring in exactly the way that means most people skip it.

For every product, the machine-readable fields that matter most:

Field	Why it matters	Common small-merchant failure
name	Matching	Marketing name with no product identity (“The Weekend One”)
brand	Filtering	Left blank on own-brand goods
gtin / mpn / sku	Cross-source identity	Missing entirely — the single biggest gap
price + priceCurrency	Comparison	Price only rendered in an image or JS
availability	Shortlisting	Static “in stock” that never changes
image	Presentation	One low-res photo, no alt text
Dimensions / weight	Fit and shipping questions	In the prose, not in a field
Material / colour / size	Attribute filtering	In the prose, not in a field
aggregateRating / review	Ranking within a shortlist	Reviews rendered by a widget the crawler can’t read

The pattern in that right-hand column is the whole lesson: small merchants put their best information in sentences. “Hand-stitched from 4mm full-grain leather, fits a 15-inch laptop” is a great sentence and a terrible data structure. It needs to be both.

If you’re on Shopify, Square, WooCommerce or Wix, most of this is a matter of filling in fields you’ve been skipping and confirming your theme actually emits the markup — many do it partially, some do it wrong, and almost none tell you. Validate a live product URL against Google’s Rich Results Test and Schema.org’s validator before you assume it’s fine.

And check that your product data is in the HTML, not injected by JavaScript after load. Some agent crawlers execute JS; not all do, and the ones that don’t will simply see a page with no price. That’s also, incidentally, a page-speed problem — the same one that makes your site feel slow to humans.

Layer 3: Feeds

If you sell through marketplaces or run shopping ads, you already have a product feed, and it’s probably an afterthought maintained by whoever set it up.

That feed is now a primary distribution asset. Keep it complete — GTINs populated, categories mapped to the real taxonomy rather than “Other”, availability updated on the same cadence as your site. A feed with 60% attribute coverage is a catalogue that’s 40% invisible.

The protocol soup, and why you should mostly ignore it

You will read a lot about agentic commerce protocols this year. Briefly, so you can stop worrying about it:

• ACP (Agentic Commerce Protocol) — co-developed by OpenAI, Stripe and Meta
• UCP (Universal Commerce Protocol) — co-developed by Google and Shopify

Both are open standards under permissive licences that any merchant can implement against any commerce backend. The landscape has been unstable: OpenAI launched Instant Checkout inside ChatGPT in September 2025 and retired it on March 5, 2026 after roughly 30 Shopify merchants had integrated; Google accelerated UCP; Amazon entered with its own approach.

Here’s the operator take: do not implement a commerce protocol yourself. Not this year, probably not next. Small merchants get access to these through their platform — Etsy sellers are already live, over a million Shopify merchants are in the pipeline, and PayPal’s ACP server is expected to bring tens of millions of small businesses along in 2026.

Your job is to be ready for the platform to flip the switch, and readiness is entirely Layers 1–3 above. Clean attributes and accurate stock make you eligible for every protocol that exists and every one that replaces them. A bespoke integration against a standard that got retired in March is the opposite of that.

How to tell if any of this is working

The measurement problem is real, because AI-sourced visits often arrive with thin or missing referrer information and land in your analytics as direct traffic.

What you can actually do:

• Segment by referrer where it exists. Assistant and search-AI referrers do show up; build a segment for them and watch it as its own channel rather than letting it dissolve into “Direct.”
• Watch the shape, not just the size. AI-referred sessions typically arrive further down the funnel — the shortlisting already happened — so judge them on conversion rate and pages-per-session against your site average, not raw volume.
• Track your own catalogue coverage as a number. What percentage of your products have a GTIN, a real category, populated dimensions, and current stock? That number is the actual KPI here, and unlike traffic you fully control it.
• Ask an assistant about your own category, monthly. Genuinely. “Best [thing you sell] for [use case], under $X.” See whether you appear, who does, and what the ones that do have in common. It’s crude and it’s the most useful fifteen minutes in this whole article.

If you don’t have analytics you trust, that’s the prerequisite — and it doesn’t have to be creepy or expensive.

What to ignore

• “AI SEO” services promising placement in assistant answers. There is no submission form and no placement to buy. What exists is structured data, accurate stock, and genuine reviews. Anyone selling more than that is selling the 2011 link-building pitch in a new jacket.
• Rewriting all your copy “for AI.” Your descriptions aren’t the constraint; your empty attribute fields are. Fix the fields first, then improve the copy for the humans who arrive after the shortlist.
• Building against a specific protocol. See above. Your platform will do this. If it won’t, that’s a signal about your platform, and we’ve compared the realistic options on cost.
• Panic. Agentic commerce is a growing minority of transactions with a credible path to a quarter of online sales by 2030. That’s significant and it is not an emergency. The work it requires is the same data hygiene that was already good practice.

The week-one checklist

Concretely, in priority order:

1. Pick your top 20 products by revenue. Not the whole catalogue. The tail can wait.
2. Fill in every attribute field for those 20: brand, GTIN/MPN, dimensions, weight, material, colour, size. Move facts out of the prose and into the fields, keeping them in both.
3. Validate the markup on three of those product URLs with Google’s Rich Results Test and Schema.org’s validator. Fix whatever the theme is emitting wrongly.
4. Confirm price and availability are in the server-rendered HTML, not injected client-side.
5. Audit stock accuracy for a week. Compare what the site says to what’s on the shelf, daily. If it drifts, that’s your real project — and a local agent watching for drift is a legitimate fix.
6. Check your product feed’s attribute coverage, if you run one.
7. Ask three assistants about your category and write down what you learn.

That’s an afternoon and a bit of follow-through, and it’s most of the available upside.

When this doesn’t apply to you

• You’re a service business, not a retailer. Trades, agencies, clinics, restaurants — there’s no product feed to fix. Your equivalent is local discovery data, and the work is your Google Business Profile and local presence, not schema on product pages.
• You sell one-of-a-kind items. Vintage, handmade, art. Fixed identifiers like GTIN don’t exist for you, and that’s fine — categories, materials, dimensions and condition carry the weight instead.
• Your revenue is entirely wholesale or repeat local trade. Then agentic discovery is a rounding error and this is a 2028 problem. Spend the afternoon somewhere else.
• You don’t sell online at all. Then the QR menu or catalogue is the higher-leverage version of this exact work — same principle, smaller surface.

Quick answers

Is agentic shopping actually big yet? As a share of completed transactions, it’s a growing minority — J.P. Morgan projects up to 25% of U.S. online sales by 2030. As a first step in the journey it’s already large: agentic search as the opening move grew 200% year over year and two-thirds of shoppers have used an AI tool for a purchase in the last three months.

What’s the single highest-impact fix? Populating product identifiers (GTIN/MPN/SKU) and real attribute fields on your top-selling products. Identity and attributes are what make a product eligible to be compared at all.

Do I need to implement ACP or UCP? No. Your commerce platform will handle protocol integration. Your job is clean product data, which makes you ready for whichever standard survives.

Will AI traffic show up in my analytics? Partially. Much of it lands as direct traffic with thin referrer data. Segment what you can, and judge the channel on conversion rate rather than volume.

Does my copy still matter? Yes — it just moved to second. Attributes get you shortlisted; copy converts the visit that follows.

The bottom line

The first reader of your product page is now a machine assembling a shortlist, and it makes that decision on fields most small merchants leave blank.

None of the work required is exotic. It’s product identifiers, real attribute fields, server-rendered prices, and stock that’s actually true. It is, almost exactly, the data hygiene that a good catalogue always needed — the difference is that the penalty for skipping it went from “slightly worse SEO” to “not in the consideration set.”

Do your top 20 products this week. Check whether stock is honest. Then go back to writing descriptions for the humans who show up afterwards, because they still buy the thing.

And the honest note to end on: the reason your stock is wrong on your website is almost never your website. It’s that your register, your storefront and your marketplace listings are three separate systems that only agree when a person makes them agree, on a Sunday, from memory. Structured data can’t fix that — that part we build.

Sources

All figures accessed August 29, 2026.

• Agentic search as the first step in the shopping journey, up 200% YoY — Salesforce
• Small-business coverage of the Salesforce data, and the 67%/80% shopper figures — Forbes / Gene Marks, August 9, 2026
• AI shopping traffic up 1,200% as discovery moves agent-led — Robotics & Automation News, August 24, 2026
• Agentic AI adoption rates among commerce organizations and the J.P. Morgan 2030 projection — commercetools
• ACP and UCP protocol landscape, Instant Checkout timeline, and small-merchant access — Opascope
• Product data structure and attribute requirements for agentic commerce — Bluestone PIM