Your product feed is an AI agent’s reading list now

For twenty years, product feeds have been optimised for one reader: Google Merchant Center's crawler, feeding a human shopper who was going to look at pictures, compare prices, and click through. The feed was a means to an end. Get the required fields right, keep the prices in sync, don't let the image URLs 404, and you were mostly there.

That model is quietly ending. And most product feeds — including the ones being maintained right now by e-commerce teams across the UK — are nowhere near ready for what's replacing it.

The shift is this: AI agents are increasingly reading product feeds, search-result pages, and structured commerce data without a human on the other end of the query. OpenAI is building chatbot-native ads that launch business-specific agents instead of landing pages. Google's own product surfaces are being consumed by AI Overviews before a human ever sees them. And the structural quality of what's in your feed — the completeness, the freshness, the coherence — is starting to function as a ranking signal in ways the human-shopper model never punished you for.

The industry is not talking about this in feed-specific terms. It should be.

The feed used to be a container. Now it's a corpus.

When a human shopper landed on a Google Shopping result, they filled in the gaps. A slightly wrong product title? They read the image. A stale price? They clicked through and saw the real one. A missing size variant? They navigated the site. The feed was a routing layer, not the final answer.

The feed stopped being plumbing. It became content.

An AI agent doesn't do any of that. When ChatGPT or Gemini or a business-specific agent reads a product feed to answer *"which noise-cancelling headphones under £200 have the best mic for calls,"* the feed *is* the answer. There is no click. There is no image parsing to fix an ambiguous title. There is no human reading a product page to correct for a feed error.

The feed becomes the corpus. Whatever's in it is what the model reasons over.

That changes the economic value of every field you've been treating as optional. Attributes you've been leaving blank because Google Merchant Center didn't strictly require them are now the difference between being cited and being invisible. Product descriptions written for SEO in 2019 — keyword-stuffed, thin on actual specification — read to an agent as low-signal noise. Category taxonomies that were "close enough" for human browsing look, to a model, like structural incoherence.

The feed stopped being plumbing. It became content.

And almost nobody in the industry is auditing feeds through that lens.

The second-order problem: two readers, one file

Here's where it gets awkward. The systems that consume your product feed haven't split. It's still one file, one Merchant Center account, one set of structured data. But the readers have split — humans on one path, agents on another — and they don't want the same things.

A human shopper wants a compelling title. Emotional language. A hero image that pops. A price positioned against a strikethrough.

An agent wants exhaustive attributes. Unambiguous specifications. Clear categorical hierarchy. Machine-readable relationships between variants. The absence of marketing language that could bias its interpretation.

These aren't opposed exactly, but they pull in different directions. And right now, most e-commerce teams are writing feeds for the human reader and hoping the agent copes. That was fine when the agent was a minority reader. It's rapidly stopping being fine.

I've seen this pattern play out before in a different form. When mobile search overtook desktop, most sites had spent a decade optimising for desktop layouts. The mobile experience was a compressed version, worse in a hundred small ways nobody had noticed because desktop was the primary reader. Fixing it took years. Feed data is heading into the same situation, faster.

What agent-readable actually means

The GEO conversation, such as it is, has focused almost entirely on editorial content. Structured data on articles. Schema on FAQ pages. Getting cited by ChatGPT for a services query. All fine. All addressable.

Product feeds are barely mentioned in the same conversation, despite being one of the most machine-consumed data structures on the entire web. Which is strange, because the fundamentals of what makes a feed agent-readable are actually clearer than the equivalent for editorial content.

An agent-readable product feed has completeness in the attributes that matter — not just the required Merchant Center fields, but the optional ones a model would use to answer a specific question. Colour, material, dimensions, compatibility, certifications, use cases. It has titles that describe the product rather than sell it. It has descriptions that read as specifications, not brochure copy. It has variant relationships expressed cleanly so a model can understand that these seven SKUs are the same shoe in different sizes, not seven different products.

It has *freshness*. Prices that match the site. Stock levels that reflect reality. Descriptions updated when the product changes. Feeds that go stale in silent ways — where the price on the site is right but the feed still shows last month's — used to just cost you Shopping-ad quality score. Now they cost you agent citations, because the agent is going to answer *"is it in stock?"* from the feed, not the site.

And it has categorical coherence. A product classified as *"electronics > audio > headphones"* is legible. A product classified as *"gadgets > cool stuff > new arrivals"* is not, and never was, but only now is that going to actually cost you.

The audit nobody's running

Here's what I find striking. Every mid-sized e-commerce operation I've worked with has a technical SEO audit cadence. Some run monthly, some quarterly. Most cover crawl depth, canonical issues, Core Web Vitals, schema validation on category and product pages.

Almost none of them audit the feed itself as a first-class artefact. The feed is checked for Merchant Center errors — did any items get disapproved, are there any missing GTINs, are the prices matching. That's it. Nobody's asking whether the feed, read as a document by an AI agent, tells a coherent story about the catalogue.

That audit needs to happen. And the questions it needs to answer aren't the ones Merchant Center will surface for you.

Are your product titles machine-parseable, or are they written like Amazon listings from 2015 — keyword salads with the actual product name buried in the middle? Are your descriptions specifications or brochure copy? If a model tried to answer *"which of your products is compatible with X,"* could it, or would it have to guess from ambiguous text? Do your categorical relationships express real structure, or are they marketing groupings that made sense for merchandising three seasons ago?

Most feeds fail this audit badly. The teams running them don't know because nobody's told them what to look for.

The measurement problem, again

I keep coming back to measurement in these pieces because it's the actual bottleneck across most of AI search. The feed problem has its own version of it.

Even if you audit your feed and clean it up, you can't currently measure with any confidence whether an AI agent is citing your products more or less than a competitor's. There's no equivalent of Search Console for *"how often did ChatGPT recommend you in a shopping query."* You can spot-check with prompt testing — running the same query across models and seeing what comes back — but that's a manual, sampled, unreliable proxy. It's not a metric you can budget against.

This is the same measurement fog that hangs over the rest of AI-search performance work. The best you can do right now is fix the inputs on the assumption that better inputs produce better outputs, monitor the qualitative signal from spot-checks, and wait for the tooling to catch up.

For product feeds specifically, that means the ROI case for cleaning them up is going to feel weaker than the case for, say, running Performance Max campaigns. The traffic you would have earned from being cited in an AI shopping answer doesn't show up as a line item anywhere. It shows up as demand — someone arriving on your site already knowing what they want because a chatbot told them — but the attribution is invisible.

That's uncomfortable. It's also the situation, and pretending otherwise doesn't help.

What this argument doesn't cover

I'm not claiming feed quality is *the* deciding factor in AI shopping visibility. Brand strength matters more. Whether a model has been trained on content that mentions your products matters. Whether you have the reviews, the mentions, the earned coverage that establishes you as a real entity in the model's world all matter and probably outweigh anything you do to the feed itself.

I'm also not claiming this is urgent for everyone. If you're a service business, this doesn't apply to you at all. If you're an e-commerce operation running twenty SKUs, the audit is small and probably not the highest-priority thing on your list. The teams this actually matters for are the ones running catalogues of thousands to hundreds of thousands of products, where feed hygiene has been an operational afterthought and where AI-driven shopping is going to eat an increasing share of top-of-funnel demand.

And I'll concede: some of what I'm describing is speculative. The chatbot-native ad format from OpenAI is early. The extent to which agents read product feeds versus scraping product pages versus using entirely separate commerce APIs is going to shift. The exact mechanics of how a model weighs feed data versus other signals is opaque and going to stay that way.

What isn't speculative is the direction. Machine readers are becoming a larger share of the traffic that reads your commerce data. That share isn't going back down. The feeds that get read as coherent, complete, and current will be cited. The ones that don't, won't.

The practical version

If you're running e-commerce and taking this seriously, three things are worth doing in the next quarter.

First, run a feed audit that's actually about the feed as content, not about Merchant Center compliance. Sample your titles and descriptions. Read a slice of them the way an AI agent would — as your only source of information about the product. Are they doing that job? Or are they still selling to a human who was going to click through?

Second, get your attribute completeness up on the fields models actually reason over. Colour, material, dimensions, weight, compatibility, certifications, use cases. Fill the blanks. The marginal cost is low; the marginal value is uncertain but positive.

Third, tighten the freshness loop between your site and your feed. Every price mismatch, every stock discrepancy, every stale description is a place where an agent gets it wrong and someone else gets recommended instead.

None of this is glamorous. None of it will show up on a dashboard. It's the kind of unsexy structural work that separates the operations that quietly win the next few years from the ones that spend those years wondering why their traffic mix keeps shifting in ways they can't explain.

The product feed used to be plumbing. It stopped being plumbing about six months ago. Most teams haven't noticed yet. The ones that do — and act on it — get a compounding advantage that's going to look, from the outside, like they got lucky with AI search.

They didn't. They read the feed like an agent would, decided they didn't like what they saw, and fixed it.