We Said Product Feeds Alone Won't Win AI Search. Our Own Data Complicated That.

Six weeks ago we argued the research layer — the open-web content AI reads to decide what to recommend — mattered more than product feeds. Then we instrumented the orders. Most of the ones we can attribute came through the feed lane, not citations. Here's the honest revision.

In August we published a piece called The Research Layer: Why Product Feeds Alone Won't Win AI Search. The argument: an AI shopping answer is assembled in two stages — research (which products deserve the shortlist) and product lookup (what are they, what do they cost). Feeds serve the second. The first runs on the open web, is not commoditized, and is barely contested.

We still believe the structure. But we've since built the feed side ourselves, watched merchants use it, and — more importantly — connected orders back to where the shopper came from. Some of what we found argues against our own article, so we're publishing that rather than quietly leaving the original up.

What the attribution actually said

Across the orders we could trace to an AI source:

• The overwhelming majority of AI sessions land directly on product pages, not on articles.
• Of the orders we could attribute to a specific AI lane, most arrived carrying feed attribution, not article citations.
• A large body of published articles produced a comparatively small number of sessions.

If you'd shown us only that table, we'd have written the opposite article.

Why we didn't just retract it

Three reasons, and they're the interesting part.

Attribution is not symmetric. A shopping surface that passes a tracking parameter is countable. A citation inside a chat answer, where the shopper reads a recommendation and later types your brand into a browser, is not. The feed lane is instrumented and the research lane substantially isn't — so a raw comparison of "feed orders vs citation orders" is a comparison of measurement quality, not of value. We know some of our citation-driven traffic is invisible because we've watched specific articles get cited and seen no corresponding referrer at all.

The lag is real and the window was short. Publishing-to-citation runs on a scale of weeks, not days. An article published in month one can be earning recommendations in month three. Counting both lanes in a single window systematically undercounts the slower one.

Content landings grew sharply once volume arrived. In one recent week, sessions landing on published content went up roughly eightfold. Small absolute numbers, but the direction matters: the research layer isn't inert, it's slow — and slow looked like dead when we only had a few weeks of data.

The revised position

Here is what we'd write today, which is neither the original article nor its opposite.

The feed is not optional and we undersold it. It is the mechanism by which an AI shopping surface can actually transact with you. When a shopper is already on a buying surface, a clean feed is what turns the recommendation into an order. Our original piece treated it as a commodity checkbox. It isn't yet — see below — and even when it is, "commodity" doesn't mean "skippable".

The research layer is still where you get chosen, and it's still uncontested. Nothing about the feed data changes the two-stage structure. The feed makes you fetchable. It has no influence on whether the engine names you in the first place.

The two lanes serve different shoppers, and that's the actual insight. Someone on a shopping surface with a query is deep in the funnel, and the feed wins that. Someone asking an assistant "what should I look for in X" is upstream, and research-layer content wins that — then may arrive weeks later by a route nothing can trace.

Running only one of these means either being unrecommendable but purchasable, or recommended and then unbuyable.

The part nobody expected: the feed is hard to switch on

We assumed the product layer would commoditize fast — every platform ships feed syndication, every store gets it, advantage gone.

In practice the plumbing is a real barrier. Among merchants we've taken through OpenAI's product-feed setup, the store-side configuration gets completed by everyone who starts it. Then the process stalls, for nearly all of them, at the same place: the handover to the ads platform, where the account and SFTP credentials live. That step happens outside your store, requires access that isn't universally available yet, and has failure modes that report success.

So the "commodity checkbox" is arriving more slowly than we predicted. For now, having a working feed is itself a differentiator — not because the data is special, but because most stores haven't got one connected.

We'd guess that changes within a year. The research-layer advantage won't.

What to actually do

1. Get the feed working. It's harder than it sounds and it's the shortest path to countable revenue. Start by confirming you can create a feed on the ads platform at all, before configuring anything in your store.
2. Keep publishing anyway — and judge it on a horizon of months, not weeks. The lane that's hardest to measure is also the one competitors can't switch off.
3. Don't let attribution quality decide your strategy. The measurable lane will always look better than the unmeasurable one. That's a property of your instrumentation, not of your business.
4. Watch content landings, not just content published. Drafts have no URL and can't be cited. In our fleet the single biggest difference between stores with AI traffic and without remains whether the drafts went live.

The original article's closing line was: one of those is a checkbox, the other is a moat. We'd keep the moat and drop the checkbox. Both lanes are work, both are worth doing, and the one you can measure is not automatically the one that matters most.

Revises "The Research Layer: Why Product Feeds Alone Won't Win AI Search" (August 2026). Sources: FoundGPT citation corpus (207,297 citations, 17,695 domains, five engines, Apr–Aug 2026), first-party session pixel, and order attribution across our merchant base. Stores anonymized. Where numbers are given as proportions rather than counts, it's because the absolute figures are small enough that the ratio is the honest unit.