Recommendation Intelligence Research™ · Study #45

Rank first in the cited source, and the model always agrees. Fourth place wins 1.2% of the time.

Inspired by a LinkedIn exchange with Olga Vainshtok: a company in her AEO audit sat 4th of 17 in a "brand alternatives" roundup, and none of the five engines she tested named it. We isolated the variable. Same 4 brands, same synthetic 17-slot source, only the target's rank moved. Selection rate: 100% at rank 1, 1.2% at rank 4, 0% at rank 15. Candidacy barely moves at all: 100%, 40%, 35.2%.

Atom Foundry · September 2026 · Inspired by a LinkedIn comment from Olga Vainshtok, part of the Recommendation Intelligence Research™ series

Cite this study: 10.5281/zenodo.23350942

100.0%Selection rate at rank 1

1.2%Selection rate at rank 4

0.0%Selection rate at rank 15

100.0ppRank 1 minus rank 15

Where this comes from

A LinkedIn comment about a company that never got named

Olga Vainshtok posted an AEO audit for a B2B SaaS company: five AI engines could describe the company accurately when asked directly, but none of them recommended it in a "brand alternatives" roundup on a relevant, reasonably authoritative site, updated 3 months prior. Her company sat 4th of 17 in that roundup. In reply to our comment connecting her finding to our own Possession vs. Deployment gap, she added her own observation: being mentioned in a source isn't enough, position inside that source seems to matter too.

That's a testable claim we hadn't isolated before. Ghost Citation (unpublished) measures a different layer: within one response, how often a cited source promotes into a named product mention. Position Stability measures the brand's position in the model's own answer across repeats, not the brand's position inside the source the model is reading. Neither tests whether rank inside a cited source, on its own, changes the odds a brand becomes a candidate or a winner. So we built the test Olga's example describes almost exactly.

The test

One factor, three positions

Same 4 brands as Winner vs Loser, byte-identical: Colored Organics, BodyArtForms, Barbaro Mojo, Hearthloom, each against its existing competitor across 20 real buyer intents. The system message describes a synthetic "brand alternatives" roundup source with 17 slots, matching Olga's real example exactly. The competitor sits fixed at slot 2 in every condition. The other 15 slots are generic anonymized placeholders ("Other Brand 1" through "Other Brand 15"), no real names. The target brand's own slot is the one thing that changes: rank 1 of 17, rank 4 of 17 (Olga's exact real case), or rank 15 of 17 (near the bottom, not the literal last slot). Forced two-way comparison, same judge prompt as the rest of this series, unchanged.

Model: gpt-4o, temperature 1.0

Brands: 4, reused from Winner vs Loser

Conditions: rank_1, rank_4, rank_15 (target's slot in the cited source)

Comparison: forced two-way vs. named competitor, LLM judge

Calls: 1,200 chat, 1,200 judge

Judge parse failures: 0 of 1,200

Spot-checked before publication. Sample rows read from judged.json across all three conditions: the model's stated reasoning matches the mechanism directly, responses cite the source ranking itself ("because it is ranked higher in the roundup") as the entire justification for the pick. Cluster-mean and pooled rates were independently recomputed from the raw judged rows and matched the reported results to one decimal place in every condition.

The result

Selection is a switch. Candidacy is a dial

Two different things happen to two different measurements. Winning a forced comparison is almost perfectly binary, tied to sitting first in the source. Simply getting mentioned barely cares about position at all.

Selection/winner rate by source rank

Cluster mean across 4 brands · n=400 per condition

Rank 1 of 17

100.0%

Rank 4 of 17

1.2%

Rank 15 of 17

0.0%

Named-candidacy rate by source rank

Cluster mean across 4 brands · n=400 per condition · is the brand mentioned at all?

Rank 1 of 17

100.0%

Rank 4 of 17

40.0%

Rank 15 of 17

35.2%

Candidacy drops 65 points from rank 1 to rank 15, but selection drops 100 points, and almost the entire selection drop already happened by rank 4. A brand sitting 4th in the source is still named in 40% of responses, roughly as often as one sitting 15th (35.2%), but it wins a head-to-head only 1.2% of the time, indistinguishable from last place (0%). Rank 1 vs. the average of rank 4 and rank 15: 99.4pp. Position-in-question control (target named first vs. second in the question sentence, a separate variable from source rank): 31.5% (n=559) vs. 35.7% (n=641), a small, clean gap next to the size of the main effect, confirming this is about position in the source, not word order in the question.

Per-brand breakdown

The rank-1 effect holds on every brand. Rank 4 is close to zero on three of four

All 4 brands hit 100% selection at rank 1, no exceptions. At rank 4, three of four brands never won a single one of their 100 comparisons; only Barbaro Mojo won any (5 of 100). At rank 15, all 4 brands land at exactly 0%.

Brand	Rank 1	Rank 4	Rank 15
Barbaro Mojo	100.0%	5.0%only nonzero at rank 4	0.0%
BodyArtForms	100.0%	0.0%	0.0%
Colored Organics	100.0%	0.0%	0.0%
Hearthloom	100.0%	0.0%	0.0%

Candidacy tells a different, brand-dependent story at rank 4: barbaro-mojo 64%, hearthloom 48%, bodyartforms 32%, colored-organics 16%. Selection doesn't inherit that spread, it collapses toward zero regardless of how often the brand still gets mentioned.

Why it matters

Being cited isn't the win condition. Being cited first is

Olga's real example, 4th of 17 and zero engines naming the company, isn't an edge case. It's close to the expected outcome under this mechanism: a rank-4 brand still shows up in the model's awareness about 40% of the time, but almost never gets picked when a comparison is forced. If a roundup, directory, or "best of" list is the source a model is reading, sitting anywhere other than first in that list is worth close to nothing for actually being recommended, even though it still helps the model recognize the brand exists. That's a sharper, more specific version of what Source Stability found: the cited source can change and the winner mostly doesn't, and what Signal Hierarchy found: some signals dominate a comparison outright. Source rank looks like one of the strongest single-position effects measured anywhere in this series so far.

Limitations

What this does not establish

4 brands

Same small cohort as the rest of this series, not a new or broader sample.

Synthetic source not live retrieval

The roundup source is described in the system prompt text, not fetched via real browsing from a live URL. This tests sensitivity to a described position inside a claimed source, not a direct replica of how a model parses rank during real web retrieval.

3 of 17 slots tested not the full curve

1, 4, and 15 were chosen to match Olga's real example, not systematically swept across all 17 positions. A trend across 3 points is suggestive, not proof of the curve's exact shape between them.

Competitor fixed always at slot 2

The competitor's own position never moved, which isolates the target's rank as the one variable but means we have no data on what happens when the competitor also sits at a worse position than the target.