Candidacy vs Selection showed that getting into a model's consideration set and winning inside it are two different games. This study manipulates what feeds that difference: the same real buyer question, the same brand, the same model, with only a single hidden line in the system message changing, something the shopper never sees and never said. Across 5 brands with weak real-world baselines and 5 different hidden-context conditions, a fabricated "you just saw an ad for this brand" claim moved winner rate further than any other condition tested, including a condition that combined the same claim with a real product fact.
Candidacy vs Selection separated "getting into consideration" from "winning" across 60,924 real stores, but it was observational, not a controlled manipulation. This study takes the same two-stage measurement and turns it into an experiment: 5 brands with weak real-world recommend rates, the same already-published real buyer question per brand, the same model, and the same deterministic brand-detection method throughout. The only thing that changes is one line in the system message, invisible to the user, describing something that supposedly happened before this conversation started.
Compared against the reused no-context control (22% winner rate, cohort-wide), all four brand-specific conditions produced a large, highly significant winner-rate lift (permutation test, 10,000 reshuffles, p<0.0001 for all four). But the sizes are not close to equal, and the order doesn't match what a straightforward more-information-is-better hypothesis would predict.
Candidacy tells a simple story: under any of the four brand-specific conditions, candidacy rate jumps to 97-100% for nearly every brand, regardless of how weak its real-world baseline is. Winner rate tells a messier one. Two brands convert candidacy into winning almost automatically, one brand is a genuine ceiling case with nothing left to move, and two brands get into consideration but keep losing to the same entrenched competitor, unless the context specifically frames prior ad exposure.
An exploratory, pre-registered check (n=5, not a confirmed test at this sample size) correlated each brand's real-world baseline recommend rate against how much hidden context lifted its candidacy and winner rate. Both correlations are negative and fairly strong: r=-0.63 for candidacy lift, r=-0.73 for winner lift. Brands that start further from the top have more room to move, and they move more, the same direction found in the multi-turn displacement study's brand-baseline correlation. The one near-ceiling brand in this cohort, already winning 100% of the time before any hidden context was added, shows zero measurable lift under any condition, a clean floor-and-ceiling illustration of the same pattern rather than a separate finding.
This sharpens the Candidacy vs Selection distinction into something closer to a mechanism. Nearly any hidden context, even a brand-blind instruction to be more decisive, is enough to move a brand into the consideration set. What actually wins inside that set depends on something else, and this study's most notable result is that the something else responds more to framing that implies the shopper already has some relationship with the brand than to framing that supplies more concrete, truthful information about it. A fabricated claim of prior ad exposure outperformed a real product fact. That is worth sitting with, both as a research finding and as a reason to be careful about what "personalization" signals actually do once they reach a model's context window.
Stating the limits up front. Brand detection here uses the same deterministic, case-insensitive substring method as the original report data, for direct comparability, and uses no LLM judge at all, which removes self-grading bias entirely but trades it for this simpler method's own blind spot: a closed, pre-registered competitor list. A manual spot-check of 20 full responses across every condition found no errors. A second, targeted check of all 244 records where a brand was detected as the winner found exactly one real misclassification: in a brand-blind condition, a model response explicitly named a different, unlisted tool as its "top pick" ahead of the brand being tested, but because that unlisted tool wasn't on the brand's known-competitor list, the detector still marked the tracked brand as the winner. Correcting this single record moves that condition's cohort-wide winner rate from 36% to 34.7%, the number shown above; the correction does not change any statistical conclusion.
Five brands, deliberately selected for weak real-world baselines, not a representative cross-brand sample. The H4 baseline-interaction correlation and the ordering of the five conditions are both exploratory at n=5, reported as such, not as confirmed findings. One category mix, consumer ecommerce, one model, one prompt per brand. None of the fabricated exposure claims in this study describe anything that actually happened to any real user of any real brand; brand identities are anonymized throughout this page for that reason, on top of the series' general anonymization-by-default policy.
Every record where a brand was detected as the winner was screened for the closed-competitor-list blind spot. One real error found, in the brand-blind condition where the model is freest to name anything. Corrected in the reported numbers above.
Every brand-specific hidden-context condition produced a statistically significant winner-rate lift over the no-context control, on a 10,000-reshuffle permutation test, cohort-wide.
If a single fabricated line in a system message can move winner rate this much, it's worth knowing what a model already assumes about your brand before anyone asks it a question. A free scan is the fastest way to check.
This study turns the observational Candidacy vs Selection split into a controlled manipulation, and sits alongside the series' other causal, context-manipulation studies.
Illustrative example · single-site signal for atomfoundry.dev.
View full signals →