Your product data is not ready for agents

The question most ecommerce businesses should be asking right now is not how to rank in ChatGPT. It is whether their product data is structured well enough to support an autonomous agent making a purchase decision on someone else’s behalf. Those are different problems, and the second one is considerably harder.

How search worked, and what it actually required

Traditional SEO was, at its core, a game of signals. Google built a system that used links, keywords, page structure, and engagement metrics as proxies for relevance and authority. The underlying content mattered, but the optimization layer mattered more. A product page with thin copy and good backlinks could outrank a detailed, accurate page with none.

This created an entire industry devoted to gaming the proxy rather than improving the underlying thing. The content itself was secondary. What counted was whether the content looked right to the crawler.

That incentive structure shaped a decade of ecommerce product data. Most online shops have product descriptions written by whoever was available at the time, inconsistent attribute naming across categories, empty specification fields, and images with no useful alt text. None of this prevented ranking in 2015. It rarely prevented ranking in 2020. The gap between what the search algorithm rewarded and what would actually be useful to a machine trying to understand the product was enormous, and it did not cost anyone much.

That gap is now starting to matter.

What GEO is actually asking for

Generative Engine Optimization is the practice of making your content useful to large language models rather than just to traditional search crawlers. When someone asks Perplexity or ChatGPT or Google’s AI Overview which running shoe to buy for flat feet, the system does not return ten blue links. It synthesizes an answer, usually with citations.

Getting cited in those answers requires something different from traditional SEO. LLMs do not respond well to keyword density. They respond to clarity, specificity, and authority. A product description that says “premium quality materials for superior performance” gives a language model almost nothing to work with. A description that specifies “full-grain leather upper, Vibram outsole rated to -40°C, 800g per shoe” gives it something it can use in a comparison.

This is the first place where product data quality starts to visibly affect commercial outcomes. Vague, generic product copy does not get cited. Specific, structured, complete product information does.

Schema markup also becomes more important here. Google’s rich results, Bing’s product answers, and the various shopping integrations in AI tools all draw on structured data to understand what a product is, what it costs, whether it is in stock, and what other people think of it. A Product schema with complete fields, accurate pricing, and linked reviews is the machine-readable version of your product catalogue, and it is increasingly the version that gets read first.

GEO has a brand dimension too. LLMs are more likely to cite products and brands they have encountered frequently in their training data and in the sources they can access at query time. Brand recognition, in the traditional sense, now has a direct effect on AI citation frequency. Content marketing and SEO produce compounding returns here that were harder to measure before.

Agentic shopping: a different kind of reader

Agentic shopping systems, whether built into assistants like Claude or Operator, or embedded in shopping platforms directly, do not browse the way humans do. They do not scan a page, notice the lifestyle photography, and decide the brand feels right. They parse structured information, compare attributes, check availability and pricing, and make a recommendation or complete a purchase based on criteria the user has specified.

When someone instructs an agent to find a waterproof hiking boot under 200 euros, size 43, available for delivery by Friday, the agent runs a structured query against whatever data it can access. That might be a shopping API, a crawled product feed, or a real-time scrape of product pages. The agent is not making qualitative judgments about the copy. It is checking fields.

If those fields are incomplete, inconsistent, or simply absent, the product does not make it into the results. Not because it was ranked lower. Because it could not be evaluated.

This is a qualitatively different failure mode from traditional search. A weak product page might rank on page three instead of page one. A product with missing or inconsistent data may not appear in agentic results at all, because the agent could not confirm it met the criteria. The product might be exactly what the user needs. The data said otherwise, or said nothing.

The state of ecommerce product data

Most ecommerce product data is in poor shape for this environment. This is not a criticism of the people who manage it. It is a structural consequence of how product catalogues are built and maintained.

Product data accumulates in layers. A brand launches with a core range and enters it carefully. Then the range expands, managed by different people under different deadlines. Products are imported from supplier feeds with whatever data the supplier included. Some fields get filled. Others do not. The attribute taxonomy made sense for the original catalogue and breaks down as the range grows. Sizes are recorded as “S/M/L” for some products and “36/38/40” for others in the same category. Weight appears on some products and not others. Materials are described inconsistently.

Nobody fixed this because nobody had to. The human browsing the site could look at the images, read the copy, and fill in the gaps. A search engine crawler could find the keywords it needed. The mess stayed invisible.

Agents cannot fill in the gaps. They work with what is there.

What data readiness looks like in practice

Getting product data into shape for agentic commerce is a PIM problem before it is a marketing problem. The work happens at the catalogue level.

The starting point is attribute completeness. Every product in a given category should have the same set of attributes populated. Not most products. All of them. An agent querying for waterproof hiking boots under 200 euros and finding that thirty percent of the products in that category have no waterproofing specification will exclude those products from evaluation. They cannot be confirmed as meeting the criteria.

Consistent taxonomy matters next. If waterproofing is recorded as “waterproof: yes/no” on one product, “water resistance: IPX4” on another, and “treatment: DWR coating” on a third, an agent comparing products across the catalogue cannot treat those as equivalent fields. A human editor would understand they are all describing the same property. A structured query does not.

Units and formats need standardization. Weight in grams on some products and kilograms on others. Dimensions as “30x20x10cm” in one place and “300mm x 200mm x 100mm” in another. These inconsistencies are invisible to human shoppers and a problem for automated comparison.

Rich product descriptions still matter, but the purpose has shifted. The description needs to serve two readers: the human who lands on the page, and the language model that might be asked to summarize or compare the product. That means writing in complete sentences with specific, verifiable claims. Specific claims are useful to both audiences. Vague claims are useful to neither.

Schema markup should be treated as a live, maintained asset, not a one-time technical implementation. The Product schema needs to reflect actual current pricing, actual stock status, and actual specifications. Outdated or incorrect structured data is worse than none, because it actively misleads automated systems.

Review data, handled correctly, is an underrated input. LLMs use review content to understand how products perform in real conditions, and agents use aggregate review scores as one signal in comparison decisions. Reviews with specific, substantive feedback about product attributes are more useful than five-star ratings with no content.

Finally, API access. Agents increasingly prefer to query product data through APIs rather than scraping pages. A product feed that updates in real time, with accurate inventory and pricing, is a significantly better signal than a page cached twelve hours ago.

The PIM as competitive infrastructure

None of this work is glamorous. It does not produce an immediate ranking improvement that appears in a dashboard. It is cataloguing work, taxonomy work, the unglamorous task of making sure every field is filled and consistent.

It is also, increasingly, the work that determines whether automated systems can sell your products at all.

A PIM that enforces completeness at the point of entry, standardizes attributes across categories, and maintains clean structured output is not just an internal efficiency tool. It is the infrastructure layer that makes a product catalogue legible to agents. Brands with well-maintained PIMs will have a structural advantage in agentic commerce. Brands with messy, inconsistently populated catalogues will find their products invisible to automated buyers, regardless of how good the products actually are.

The parallel to early SEO is imperfect but instructive. Brands that took technical SEO seriously in 2010 built a compounding advantage over brands that treated it as optional. The underlying mechanism is different now, but the dynamic is similar. The brands that get their data in order before agentic shopping becomes mainstream will be in a better position than those who try to catch up after the fact.

Data readiness is infrastructure. It should be treated that way.