You Don't Have One Feed Problem. You Have Three — And a Black Box on Top of All of Them.

Somewhere in the last few months, a small industry sprang up to sell you the same six-word promise: clean your data, get found.

You now know that's only ever a partial answer, at best — because there isn't one house, there are five, and they don't agree on the rules. This chapter zooms into the two that matter most for pure discoverability — OpenAI and Google — and asks a harder question: even inside the houses built on open standards, how much can you actually verify?

Feed audits. "GEO" agencies. Dashboards with a single glowing score out of 100. All of it points at one villain — your messy product feed — and one hero — whoever's selling the fix.

It's a good story. It's also not true, and the part that's missing is the part that actually matters.

There is no "the feed"

Ask three different platforms how they decide whether your product exists, and you'll get three completely different answers, because there are three completely separate systems, built by three separate companies, that don't talk to each other and were never designed to.

The first is the one you can see: your own page. Whatever schema.org markup sits in your HTML. For Google, this is real but secondary — a verification layer checked against something else, not the primary source. For ChatGPT, it's closer to irrelevant: OpenAI has been explicit that product data is not crawled at all. Your beautifully marked-up product page can be flawless and ChatGPT will never once look at it.

The second is a file you'll never see anyone else's version of. OpenAI's actual mechanism is a private, structured feed — JSONL, CSV, or Parquet — pushed directly to a secure endpoint, refreshed as often as every fifteen minutes. Direct submission is invite-only for enterprise retailers; most brands reach it secondhand, syndicated through a commerce platform. Google runs a parallel private channel of its own, Merchant Center, feeding what it calls the Shopping Graph — an index of tens of billions of listings that AI Mode and Gemini actually read from. Neither of these files is public. Neither is crawlable. You cannot audit a competitor's. You can barely audit your own without asking the platform directly.

The third didn't exist a year ago, and it isn't a file at all — it's infrastructure you have to host. Google and Shopify's Universal Commerce Protocol asks a merchant to run live, callable API endpoints — cart creation, checkout, identity linking, fulfillment — discoverable through a manifest published at a predictable address. This isn't data cleanup. It's a service you now have to build, deploy, and keep running, on top of everything else.

Three planes. Three separate integration points. Fix one and you've fixed a third of the problem, at best.

And most of what's being sold to you right now only touches the first — the one plane that matters least to the platform growing fastest.

And even if you get all three perfectly right, there's a locked room on top

Here's the part almost nobody selling "AI visibility" says out loud: after your data clears every documented requirement on all three planes, a separate, undisclosed process decides which of the qualifying products actually gets said out loud to the shopper. OpenAI does not publish its full ranking function. Google has never published one for the Shopping Graph either. Both companies say, on the record, that placement isn't for sale and results reflect data quality — which may well be true, and is also, functionally, unfalsifiable from the outside. You cannot inspect a scale you cannot see.

The practitioners who track this obsessively, full-time, are reduced to reading tea leaves. One industry research group observed a sharp, measurable shift in which merchants ChatGPT named in its answers, coinciding with a model update — a real, documented pattern, sitting right next to an honest admission that OpenAI never confirmed the change or explained it. That's the most rigorous version of "we think this is what happened" currently available in the entire industry. Nobody has more than that. Not the loud agencies. Not us. Not you.

This is the discomfort worth sitting with for a second: the metric everyone is chasing — did I rank — is being measured by people staring at smoke, because that's genuinely all there is to look at.

Which makes the confident pitches the actual red flag

Anyone promising you a guaranteed ranking, a formula, a "we cracked the algorithm" deck — is either wrong, or lying, or selling you plane one and calling it the whole house. This isn't a cynical take; it's the plain, published position of the platforms themselves. And it explains something that should otherwise seem strange: adoption is real and accelerating — 68% of retailers plan to adopt agentic AI within the next year — while, at the exact same time, Gartner projects 40% of agentic AI projects will be cancelled by 2027, and MIT has found 95% of generative AI pilots fail to show measurable impact. Those numbers aren't in tension. They're the same story: a lot of money chasing a result nobody can honestly guarantee, sold by people who didn't say so.

So what's actually worth doing

Not nothing. Just something more modest and, oddly, more durable than a ranking promise.

You can audit against every plane's documented, public specification — because the requirements themselves genuinely are published, even though the outcome isn't. That's real, checkable work: does your public markup validate, does your submitted feed carry every required field, do you have a live UCP profile if that surface matters to your category. None of that requires seeing inside anyone's black box.

And you can watch the black box's output, even though you can't read its insides — by asking the same real customer questions of ChatGPT, Gemini, and Copilot before you fix anything, and again after. Not a promise. An observation, repeatable, on a schedule, because the ground keeps moving under everyone at the same time.

THE HONEST VERSION

That's a smaller claim than "we'll get you to the top." It also happens to be the only claim in this entire industry that's actually true.

REFERENCES