A single AI assistant is not a single model. ChatGPT and Gemini each run a family of them, and the one that answers a given question depends on load, on how hard the question looks, and on what the product decides you need. That matters for a restaurant or hotel, because those models do not describe you identically. A business the strongest model names confidently can be absent from the faster one's answer to the same question.

Revised October 2026. An earlier version of this article framed this as a paid-versus-free subscriber gap and cited specific percentages and customer-spend multiples. We could not substantiate those figures and we cannot observe what any subscription tier is served, so they have been removed. What follows is the mechanism we can actually measure, and the limits of what our audit shows. Our methodology is published in full on the methodology page.

What we can and cannot observe

We query models through their APIs. That means we can measure, precisely and repeatably, how a specific named model answers a specific question with live web search enabled. It also means we cannot see inside the ChatGPT or Gemini apps. We do not know which model any individual user is served at any moment, and we do not claim to. Anyone telling you they measure "what Plus subscribers see" is inferring it, not observing it.

What we report is therefore model-level: four named models, each queried identically, each scored separately. You can see exactly which model named you and which did not.

Why a capability gap matters anyway

Routing to cheaper, faster models is not an edge case — it is how these products stay economically viable at the scale they operate. A faster model handles a larger share of real traffic than its benchmark scores would suggest. So if you appear only in the strongest model's answers, you are missing a meaningful part of the actual question volume, whatever the subscription mix happens to be.

The mechanism behind the difference is reasonably well understood. Models with a smaller inference budget lean harder on what is popular and heavily aggregated — review-site density, obvious name recognition. Models given more room to reason engage more with specific positioning, which is exactly where a distinctive independent business lives. A chef-led restaurant with a clear point of view and thin aggregator presence is the classic profile for a wide gap.

What this means for you

The useful question is not which model to optimise for. It is which one you are currently invisible to, because the remedies differ:

Named by the stronger model but not the faster one, and the gap usually lives in signal volume — review presence on the aggregators, recency and quantity of positive reviews, Google Business Profile completeness. These are the signals a model reaches for when it has less room to think.

Named by the faster model but not the stronger one, and the gap is usually in substance — what is actually written about you in press and on your own site that a reasoning model can use to justify a recommendation.

Missing from both, and this is not a routing problem at all. Fix the fundamentals first: Google Business Profile, name/address/phone consistency, structured data. Model-specific gaps are a refinement on top of basic visibility, not a substitute for it.

How we report it

Every audit breaks your citation rate out by model and by intent, so the comparison above is something you read off your own data rather than take on trust. Your audit names the models queried and the date they were queried, because model lineups change — the set we query today is not the set we queried six months ago, and scores are only comparable within the same set.

We also publish what we did not measure. That is the part most audits leave out.