Recommendation Intelligence Research™ · Study #42

A bad rating barely touches candidacy. It wrecks the final round.

A LinkedIn commenter argued that AI pre-filters bad actors before they ever get mentioned, so a mention itself is a reasonable proxy for success. We tested that directly on 4 brands reused from Winner vs Loser: a neutral system context against a disclosed-synthetic negative reputation, low rating, review count, one generic non-safety complaint. Two probes, an open buyer question and a forced two-way comparison against a named competitor. The result is the opposite of the pre-filtering theory. Candidacy rate went up under negative reputation, 39.0% to 54.5%. The forced-selection win rate collapsed, 83.5% to 24.3%.

Atom Foundry · September 2026 · Started from a comment thread under Jairo Guerrero's post on mentions as a KPI, part of the Recommendation Intelligence Research™ series

Cite this study: 10.5281/zenodo.23350436

+15.5ppCandidacy rose, not fell

-59.2ppSelection win rate collapsed

4Brands tested

2,400Chat + judge calls

Where this comes from

A specific claim from a comment thread, tested instead of argued

Jairo Guerrero, commenting on a post about mentions as a KPI, argued that AI filters out bad actors before a mention ever happens, so negative mentions are rare in the wild and a mention itself is a good enough proxy for success. Our own candidacy-vs-selection framework already complicates that story in general (599 stores ever recommended against 60,325 that never were, and the factors that predict getting in barely predict winning once you're in), but we had never tested negative reputation specifically as its own variable. Every earlier study in this cluster, Winner vs Loser, Authority Signal, Brand Familiarity, Signal Hierarchy, pit positive signals against each other or against a neutral rating. None of them asked whether a bad reputation blocks a brand from being mentioned at all, or only costs it the final pick.

The test

Two probes, same brands, same negative signal

4 brands were reused byte-identical from Winner vs Loser: Colored Organics, BodyArtForms, Barbaro Mojo, Hearthloom, each with its existing competitor and 20 real buyer intents. A new disclosed-synthetic negative reputation was added for each brand, a low star rating, a review count, and one short, generic, non-safety complaint (shipping delays, slow customer service, a confusing checkout), deliberately avoiding product-safety or quality claims given two of the four brands sit in trust-sensitive categories (baby clothes, body piercing jewelry).

Model: gpt-4o, temperature 1.0

Brands: 4, reused from Winner vs Loser

Conditions: neutral vs. negative reputation

Probe A, candidacy: open buyer question, brand_mentioned string match

Probe B, selection: forced two-way vs. named competitor, LLM judge

Total calls: 1,600 chat (800 per probe), 800 judge on Probe B

Pre-registered before the run. The design, the two probes, and which of them would support versus contradict the pre-filtering theory were written into STUDY-DESIGN.md before the run, so this page reports the pre-registered comparison, not one picked after seeing the data.

The result

Candidacy barely moves. Selection collapses

Across all 4 brands (cluster mean, neutral minus negative), candidacy rate did not fall under negative reputation, it rose, from 39.0% to 54.5% (for 3 of the 4 brands individually, see below). The forced-selection win rate, in a direct head-to-head against a named competitor, collapsed from 83.5% to 24.3%, a 59.2 point swing. Negative reputation does not appear to block a brand from being mentioned in an open buyer question. It decides the outcome once that brand is put in direct competition.

Candidacy vs. selection rate, neutral vs. negative reputation

Cluster mean across 4 brands · n=100 per brand per condition per probe

Candidacy · neutral

39.0%

Candidacy · negative

54.5%

Rose, not fell

Selection · neutral

83.5%

Selection · negative

24.3%

-59.2pp

Per-brand breakdown

The selection collapse holds for every brand. Candidacy is messier

Selection rate falls for all 4 brands without exception, sharply for 3 of them. Candidacy moves in the opposite direction for 3 of 4 brands, and falls for only one, BodyArtForms.

Brand	Candidacy, neutral → negative	Selection, neutral → negative
Barbaro Mojo	23.0% → 62.0%	97.0% → 30.0%
BodyArtForms	26.0% → 16.0%	100.0% → 56.0%
Colored Organics	80.0% → 86.0%	41.0% → 0.0%
Hearthloom	27.0% → 54.0%	96.0% → 11.0%

A possible priming effect, not yet confirmed. Candidacy rose for 3 of 4 brands under negative reputation, including Colored Organics, which was already near ceiling at 80% neutral. One reading: injecting a negative reputation still hands the model a longer, more specific block of text about the brand, a star rating, a review count, a complaint, and that extra salience may make the brand more likely to come up in an open answer even though the content is unflattering. This is an interpretation, not a demonstrated mechanism. A clean follow-up would compare against a neutral filler fact of the same length, to separate salience from valence.

Why it matters

Not pre-filtering. A finals problem

Jairo Guerrero's theory was that bad actors get filtered out before a mention ever happens, which would show up here as candidacy falling under negative reputation. It did not fall, on average it rose. What actually happens is that negative reputation is close to fatal once a brand is put in direct comparison with a named competitor, win rate falling from 83.5% to 24.3%. A brand carrying a bad reputation is not being quietly removed from consideration. It is still showing up in the room. It is just losing, decisively, every time it has to stand next to someone else.

Same shape as Candidacy vs Selection, different driver. That earlier study found one store factor separates who ever gets recommended from who never does, but barely predicts who wins once inside the recommended set, getting in and winning are governed by different things. This study adds a second, specific driver operating on the winning side of that same split: a bad reputation does not close the door, it loses the room once you're standing in it.

Limitations

What this does not establish

4 brands

Same small cohort as Winner vs Loser, not a new or broader sample. Extending to more brands is an open follow-up.

Synthetic reputation

The negative rating and complaint were written by one researcher, disclosed as synthetic, deliberately generic and non-safety to avoid realistic negative claims about a real company.

A vs B not paired

Probe A and Probe B use different prompts and mechanisms. The comparison between them is descriptive, not a formal statistical interaction in one model.

1 model, 1 round

gpt-4o only, single round. Self-grading risk on Probe B is mitigated by reusing the already-validated Winner vs Loser judge prompt unchanged.

Is your brand's reputation costing you the final round?

Free AI Commerce Score™ in 10 seconds.

See how your store is positioned before it ever gets compared against a named competitor.

Free · No signup · Results in 10 seconds