The AI Recommendation Consensus Gap: High Visibility Can Depend on One Model

Share and consensus answer different questions

AI share of voice asks how much of a defined recommendation corpus names a firm. Model coverage asks how many assistants name it at least once. A firm can lead the first measure while remaining fragile on the second.

This is the consensus gap: the distance between aggregate visibility and cross-model recognition. It matters because a blended score can hide dependence. If one assistant supplies nearly all of a firm’s appearances, a provider update or retrieval change can erase much of the measured position without anything changing on the firm’s website.

The reverse matters too. A firm with modest aggregate share but four-model coverage may have shallow yet broad recognition. That is not automatically superior, but it is a different competitive asset.

Consensus differs dramatically by category

Viclaro counted the number of top-ten providers named by all four assistants after rebuilding every category from the same live extraction revision. IVF had nine, personal injury eight, cosmetic surgery seven, estate planning six, divorce four, immigration four, plaintiff employment four, and cosmetic dentistry two.

That range cannot be explained by concentration alone. Personal injury has the lowest top-five concentration but eight four-model firms. Plaintiff employment is the most concentrated category yet has four. Concentration and model breadth remain distinct even after extraction reconciliation.

These combinations produce distinct strategies. A high-consensus market rewards evidence that travels across models. A low-consensus market may offer faster entry, but gains should be diagnosed assistant by assistant.

A coverage count needs a strength check

Four-model coverage is not sufficient by itself. One appearance on a model should not carry the same interpretation as repeated appearances across a full prompt panel. The reconciled data expanded several model cells dramatically after citation-only entities were retained, which can raise nominal coverage without proving equal recommendation strength.

A useful consensus scorecard therefore reports aggregate share, per-model counts, number of models, and the response denominator for each model. It should also show confidence intervals where the sample supports them. Coverage is a breadth indicator, not a replacement for frequency.

This is the same reason Viclaro separates mentions from citations. Precise labels prevent an attractive number from acquiring a meaning the underlying event never had.

Set a breadth goal and a volume goal

For a firm present on one model, the next objective may be a second-model foothold rather than more share on the existing assistant. For a four-model firm with low frequency, the job may be increasing fit on decision and constraint prompts. A blended “improve visibility” goal cannot distinguish those paths.

Report progress as a pair: recommendation share and qualified model coverage. Define “qualified” with a minimum response denominator and repeat threshold before examining results. Then retain the raw counts so a nominal coverage change cannot disguise a one-off appearance.

The model-leader comparison shows how often the winning name changes by assistant. The NYC Recommendation Index places the consensus counts beside category concentration.

What the consensus gap does not prove

Cross-model presence does not prove causal authority, future stability, or commercial impact. The assistants may draw on overlapping sources, model outputs can change, and no defensible public weighting tells us exactly how much each assistant should contribute to a local-service market estimate.

Consensus is best treated as resilience inside a defined panel. It tells a firm whether recognition is narrow or distributed. Traffic, inquiries, and retained clients remain separate outcomes.