When you ask an AI model an open-ended question without naming a specific entity, its response is not a random roll of the dice. The names that surface are selected by a complex interplay of pre-training statistics, context windows, and instruction tuning.
By analyzing thousands of standardized responses, we can isolate the primary factors that decide which public figures, brands, and organizations get mentioned, and why.
1. Pre-training frequency (The baseline representation)
The most fundamental driver of visibility is representation in the model's training corpus. If an entity appears millions of times in web scrapes, books, and news articles, its token associations are highly reinforced. When the model generates a response about a specific category (e.g., 'musicians' or 'tech companies'), the probability weights heavily favor these high-frequency tokens.
However, pre-training frequency is just the raw baseline. Highly visible real-world entities can sometimes be suppressed, and less famous ones elevated, by other forces.
2. Prompt formatting and context steering
The phrasing of a prompt significantly shifts token probabilities. Asking for 'influential contemporary figures' yields different results than asking for 'notable professionals'. In our methodology tests, we observe that models associate specific qualifiers (like regions, historical achievements, or niche sub-genres) with different subsets of entities. This context steering is what makes GEO a viable practice: entities must align their public profiles with the specific contextual signals that models are trained to group them with.
3. List effects and positional bias
Models demonstrate strong positional bias when instructed to output numbered lists. The first 3 to 5 slots are heavily dominated by the absolute highest-probability associations (the 'head' of the distribution). Slots 6 through 10 display much higher variance, showing that models explore the 'tail' as generation progresses.
Additionally, because of how causal language models work, the name generated in slot 1 influences the probability weights of names generated in slot 2. This creates clustered list responses (e.g., mentioning one Portuguese athlete often triggers mentioning another).
4. Instruction tuning and safety filters
Modern LLMs undergo extensive Reinforcement Learning from Human Feedback (RLHF) and safety filtering. These interventions deliberately alter token probabilities. For example, a model might be tuned to avoid over-representing a single demographic, or it might have filters preventing it from naming controversial figures in open-ended recommendations. These adjustments can silently drop an entity's visibility score to zero, regardless of their pre-training frequency.