AI Knew This Ingredient Supplier by Name. It Rarely Recommended It.

Over five weeks this summer, we repeatedly monitored how four AI systems answered buyer-style questions for a design partner: an established, mid-market B2B food-ingredient supplier. We asked the kinds of things a formulator or a procurement lead might type when they're sizing up who to work with — sweetener-alternative and clean-label formulation questions, and the taste, aftertaste, and ingredient-performance queries that decide a shortlist.

I expected to write about a number moving. That is not the story the data told. The story is a gap, and it barely moved at all.

---

Why I think this matters for ingredient suppliers

I want to be careful here, because it's easy to overclaim. My data measures what AI answers — not how many of your buyers have shifted their research into an AI assistant. What I can say is narrower and still useful: to the extent buyers ask these tools "who makes a clean-label sweetener system without a bitter aftertaste," the answer names a few companies. If you're not one of them, you weren't rejected — you were never in the consideration set.

For a category built on trust, specs, and long sales cycles, being absent from that first unaided answer is a real cost, and an easy one to miss.

How we measured it — plainly

Repeatedly over five weeks, per category, we sent a fixed set of buyer questions to four engines and logged every response and its cited sources. We separated two question types: branded ("tell me about [this supplier]") and unbranded ("recommend a supplier for [this problem]"). Branded questions measure Recognition — does AI know you when named. Unbranded ones measure Discovery — does AI bring you up when it isn't told to.

Before any claim, the data hygiene, because most "AI visibility" numbers skip it. We logged 114 focus-area runs across 20 dates. Only 67 were fully healthy; 43 came back partial and 4 failed. Mid-way we revised the question set, which changed the denominators — so early and later runs aren't directly comparable. I ran the numbers only on healthy runs, within a single category and a single question-set version, and I re-derived the appearances from the raw stored responses. Everything below survives that filter.

Recognition is AI answering questions about you. Discovery is AI spending one of its few recommendation slots on you.

Finding 1: Recognition was strong. Discovery was almost nothing.

Named, the supplier was recognized in nearly every check — on the order of 16 of 16, 12 of 12. The brand is known.

Unnamed, it nearly vanished. In 54 of the 67 healthy focus-area runs, it appeared zero times in unbranded questions. It did surface in four of the five categories — but only ever in tiny amounts: one, two, or at most three appearances out of 48 to 96. Not "declining." Just barely present in the answer AI gives a buyer who hasn't heard of you yet. That gap — recognized when named, absent when not — held across every category and every engine.

Finding 2: Every unaided appearance came from one engine.

This is the finding I checked hardest, because it's the kind of thing that's easy to assume and wrong to state loosely. I pulled all 22 runs where the supplier appeared unaided and re-derived them from the raw responses. Every one of those appearances — 36 in total — came from the single engine that retrieves live web sources at answer time. The other three engines returned it zero times unaided, in every run.

They answer from what they already "know," and in unbranded questions, they didn't know to bring this supplier up. If you treat "AI" as one thing, you miss this entirely: for this supplier, unaided discovery was only even possible on the retrieval engine.

Finding 3: A source hypothesis, stated as a hypothesis.

In 21 of those 22 runs, the answer that surfaced the supplier had the supplier's own website among its cited sources. And in the runs I reconciled, only the retrieval engine returned citations at all.

That is a correlation, not proof of cause — I won't pretend a single design-partner sample settles it. But it's a specific, testable idea I want to keep pressure-testing: unaided visibility tracked whether the retrieval engine had the supplier's own material to cite. Where its own content was in reach, it occasionally surfaced; where the cited sources were third-party research, regulatory pages, or a competitor's own domain, it stayed at zero.

What changed — and what I would not conclude

Almost nothing moved, and I'm not going to dress it up. Across comparable runs the aggregate held flat. The one thing that did move was cosmetic: the specific competitor ranked first flipped almost run to run — a small set of large, incumbent ingredient houses trading the top spot while the underlying picture stayed put.

What I would not conclude is that anything trended, or that the monitoring itself moved a number. We changed the question set mid-way, roughly 40% of runs were degraded, and the lone uptick I saw is both tiny and confounded by that change. Claiming a performance trend here would be exactly the overreach I'm arguing against.

What this means if you supply ingredients

Three things I'd stand behind:

Recognition is not Recommendation. Being known is not being recommended. This supplier's Recognition was strong and its Discovery was near-zero at the same moment.

Get into the citable set. In the runs I reconciled, unaided appearances came alongside the supplier's own content being cited — and only on the retrieval engine. If your material isn't where that engine looks for your category, you likely won't be recommended, however established you are. Worth testing on your own name.

Watch the aggregate over time, not one scan. A single scan can actively mislead you — that flipping first-place competitor is the proof. The pattern across weeks is the only honest read.

The close

This is why we run this as ongoing monitoring rather than a one-time audit. A single snapshot would have said "recognized, competitors present" and stopped. Five weeks of healthy runs showed something more useful and more uncomfortable: an established supplier that AI knew by name and almost never recommended, in the exact questions its buyers may be starting to ask. That gap is measurable — and you can only close what you can see.

Written from a design-partner engagement. Company, competitors, and cited sources are withheld; figures come only from healthy, comparable runs, re-derived from raw responses.