Twenty years into watching catalogues move between systems, the pattern is boring and reliable: product data gets shaped by whatever is reading it. Pointed at a human with a filter sidebar, it grew sizes, colours, materials, and a two-line description nobody reads. Pointed at a search engine, it grew structured data and canonical URLs. Now a different reader has arrived, and the data was not built for it.
The reason this matters for a CFO and not just a merchandiser is where the invoice lands. When a retailer or a brand talks about getting ready for AI shopping, the budget line usually goes under SEO or discoverability: schema, feeds, maybe a vendor like Feedonomics or Botify on top. That is the smaller cheque. The larger one, the one almost nobody has written yet, is the product information management side: attribute models, inventory signals, and the governance that keeps the two in sync. It is a PIM bill dressed as an AI bill.
I asked where the honest view of this sits in print, and the clearest answer I found was from Forrester’s Emily Pfeiffer, who has been writing about agentic commerce for the last eighteen months with the useful habit of saying when something is not working. What follows is her published position on where product data has to change, with the operational levers she names, and then the one lever where almost nothing works yet. Then what any of this costs to ship inside a real WooCommerce store, because that is the part a decision maker actually has to budget.
Filter-Era Product Data Does Not Answer Agent Questions
Pfeiffer, Principal Analyst at Forrester, put it in one line in her 6 October 2026 piece with Tony Plec: “Your product data was built for filters, not questions” (Forrester blog, Oct 6 2026). That is the catch. The investment that got a brand through the last decade of retail, cleaning up SKUs, normalising colour names, filling in dimensions, was aimed at a shopper narrowing a shortlist. It was built to describe. An answer engine is doing something else: a shopper is asking what they can do with a product, which car seat fits their stroller, which backpack a nine-year-old can actually carry, whether the running shoes will hold up on wet pavement.
Pfeiffer’s framing in that piece is that the data has a missing dimension, not literally a third measurement, but a layer of context the filter-era schema never had to carry. “Conversational shoppers ask what they can do with a product,” she writes. The product data does not answer that unless somebody wrote the use cases in.
The scale is already there to justify the work. Forrester’s own research, cited in the same post, finds that among consumers who have used answer engines for retail, 89% have used them to research products or services (Forrester RES200830, Oct 6 2026). The buying may be slow, and we will get to that, but the research has moved. The front door is in a chat window and product data is being read aloud without anyone from the brand standing next to it.
“Agent-ready data gives the merchant more influence over how the product is represented on channels the merchant does not own.”
Emily Pfeiffer, Principal Analyst, Forrester, in Agentic Commerce Exposes The Missing Dimension In Product Data, 6 October 2026
Inventory Is the Line Nobody Priced In
The missed line in most 2026 agentic-commerce plans is not descriptive. It is inventory. Pfeiffer and Sucharita Kodali’s 7 March 2026 post on OpenAI’s Instant Checkout retreat is blunt: “Inventory management has been disastrously absent from the plan” (Forrester, Mar 7 2026). The product feeds that went into answer engines usually reflected only what was for sale on the merchant’s own website, which meant planned inventory, channel-specific holds, pre-orders, click-and-collect windows, and anything behind a loyalty tier were invisible to the agent surface.
That is a specific, boring, expensive problem. An AI agent that recommends a product it cannot actually get is doing the brand reputational damage on a channel the brand did not pick. The inRiver-reported account of Walmart’s early Instant Checkout surface, where Pfeiffer is paraphrased saying that the product data scraped from retailer websites was often inaccurate on stock, delivery timing and shipping cost, is the shape of the problem across the category: the feed is a snapshot of a website, not a view of the order management system. The order management system is where this gets fixed, and that is where the budget line really sits.
The Protocol Layer Is Real, and It Moves the Entry Ticket
Two protocols now define what being readable to an agent means in practice. OpenAI and Stripe published the Agentic Commerce Protocol on 29 September 2025, open source, with Instant Checkout launching for US users in ChatGPT the same day (Stripe newsroom, 29 September 2025). Google followed on 11 January 2026 with the Universal Commerce Protocol, co-developed with Shopify and endorsed by more than twenty companies including Walmart, Target, Wayfair, Etsy, Adyen, American Express, Best Buy, Mastercard, Visa, The Home Depot and Zalando (Retail TouchPoints, 11 January 2026).
The surface decisions follow. OpenAI’s product feed specification accepts updates every fifteen minutes, which is faster than most brands’ current PIM-to-feed sync. Each product carries two flags, enable_search and enable_checkout, so a merchant can be findable without being buyable in-channel, or vice versa (Forrester, 15 October 2025). Google’s protocol pushes direct checkout inside AI Mode in Search and the Gemini app, with Walmart stating it will put its full catalogue into Gemini.
The reason this is relevant to a buyer and not just a plumbing question is that the protocols raise the floor on what counts as a complete catalogue. A field that was optional in Google Shopping is required in ACP. A rule that was lenient for schema.org is strict here. Pfeiffer’s point in the October piece lands especially hard here: “Simply measuring how many attributes you maintain will lead to significant blind spots.” Completeness of attributes is not the same as agent-readiness. A brand can hit 100% attribute coverage on the human-era schema and still fail the agent-era read.
The Lever Almost Nobody Has Cracked: Buying in the Channel
This is the honest part of the case, and the reason not to pour the entire 2027 budget into one of these channels. By March 2026 OpenAI had pulled back Instant Checkout, moving transactional flows back to retailer apps instead of completing them inside ChatGPT. Pfeiffer and Kodali’s reporting from that date, citing Shopify directly, put the live-merchant count on ChatGPT’s checkout feature at roughly 30 and rising, against the millions of merchants the integration was originally pitched to reach (Forrester, 7 March 2026). Thirty is not a rounding error; it is the whole number.
Forrester’s March 2026 ConsumerVoices survey of US, UK and Canadian adults who regularly use answer engines put completing a purchase in the engine as the least-adopted use case among regular users. General questions and product research were the top two. The Forrester December 2025 Consumer Pulse found 23% of US Gen X online adults had used ChatGPT to search for products in the past month, 32% of Millennials, 35% of Gen Z. Research traffic is real. Purchase traffic inside the channel is not, yet.
Pfeiffer’s own summary of the state of the category in her 28 May 2026 post was two short lines: “Most ‘agentic’ experiences are still conversational” and “Hype is running ahead of behavior” (Forrester, 28 May 2026). Her operational recommendation to merchants, in the March post with Kodali, is the one worth copying verbatim: “No merchant should be 100% reliant on any one digital channel or partnership for their strategy.” Prepare the data. Do not bet the P&L on any one shopfront that currently reads it.
What This Costs to Ship on a WooCommerce Store
Here is the build section, because this is the half of the question Forrester does not answer and we do. Take a mid-market brand on WooCommerce or WordPress VIP with a catalogue of roughly 2,000 to 20,000 SKUs, a product team of one to three, a developer or an agency on retainer, and today a product data model that does what SEO wanted it to do. Getting that estate to agent-ready, in a version that survives OpenAI’s and Google’s specs as they drift, is five pieces of work.
One, a use-case field on every product. The shift Pfeiffer names, from describing to answering, has to land as a structured field, not as free text in the long description. A custom taxonomy or an ACF field group called intended_use, with controlled vocabulary per category, populated by merchandising, versioned in the database. Two to three sprints for a team with a PIM specialist on hand; longer if the controlled vocabulary has to be invented. The reason it is structured and not prose is that agents read the structured surface faster and the field can be audited for coverage. Free text in a 2,000-word description is not auditable.
Two, an inventory truth source that is not the website. This is the inventory line Pfeiffer and Kodali flag. The order management system, not WooCommerce’s product stock, has to be the source for anything an agent is going to read. Practically, that means a REST endpoint the feed generator reads from, with per-location availability, pre-order dates, backorder dates, and channel visibility flags. On a store without a real OMS, it means building one as a mu-plugin on top of WooCommerce custom tables, which is one to two months of real engineering, not a plugin install. On a store that already has NetSuite or SAP S/4HANA, it is an integration sprint and a contract with the feed vendor.
Three, a feed layer per protocol, not per channel. One Agentic Commerce Protocol feed. One Universal Commerce Protocol feed. A Google Merchant feed and a Meta feed kept because they still carry traffic. The feed is generated from the same canonical product data, not re-authored per surface. On WooCommerce this is a scheduled Action Scheduler job that writes to a static file, rebuilt on product save, with per-protocol transformers kept as small PHP classes. One sprint to set up, then an ongoing cost of maintaining the transformers as the specs drift. OpenAI’s spec currently refreshes every fifteen minutes, which the generator has to respect without bringing the database to its knees.
Four, the two flags, treated as governance, not toggles. OpenAI’s enable_search and enable_checkout are the hinges that decide whether an agent can find a product, buy a product, both or neither. A buyer should decide this per category and per channel. A $3,000 bespoke bike on a mid-market cycling store belongs in search surfaces and almost certainly not in a channel’s in-chat checkout. A $19 consumable belongs in both. Run this as a documented policy, enforced by the feed generator, not left to a tick-box in the editor.
Five, a measurement loop you will not love. Pfeiffer is honest that there is no clean referral path from most answer engines back to GA4 today; her October piece notes “there’s still no way for merchants to see the actual conversation.” The practical workaround is a combination of branded search lift, direct traffic lift to product URLs cited in agent answers, and a quarterly manual audit of what ChatGPT, Gemini and Perplexity return for your top fifty category queries, logged in a spreadsheet, dated, scored. Half a day per quarter per analyst. It is not satisfying measurement. It is better than nothing and it keeps the team honest about whether the work is landing.
Priced roughly, for the 2,000 to 20,000 SKU store described above: eight to twelve weeks of concentrated engineering and product work to stand the whole thing up, then a running cost of a sprint per quarter to keep up with spec drift and the measurement loop. That is not a cheap bill. It is also not a line most 2027 digital plans have on them, and the brands that have put it there quietly will be the ones Pfeiffer’s missing dimension stops being missing on.
What a Buyer Should Do With This
Prepare the catalogue for agents the way you would prepare it for a new wholesale partner: assume the reader is strict, the data has to be complete, and your inventory story has to be true. Do not pour money into being native in one in-chat checkout, because the live-merchant counts say the channel itself is not ready. Do pour money into the data that will travel into every subsequent channel, because that is where the compounding is. If you want to talk through what that looks like on your estate, with the WooCommerce or WordPress VIP implementation detail included, that is the conversation we have every week. We will be straight about which of the five items above we would start with on your specific catalogue, and which we would hold off on until the specs settle.
Last modified: November 16, 2026