LLM Training Data vs. Citations: Ann Smarty Explains Visibility Foundation

In a recent LinkedIn post, Ann Smarty sheds light on a critical distinction in how Large Language Models (LLMs) generate information, emphasizing the foundational role of training data over real-time citations for visibility. Smarty uses a specific example involving a platform similar to her own, Smarty Marketing, to illustrate this point.

The discussion centers on an LLM’s response to a query about a particular platform. While the citations provided to the AI included pages from both the company in question and Ann Smarty’s own site, the AI’s answer predominantly focused on Smarty Marketing. This outcome, Smarty explains, is due to the AI’s underlying training data.

“👆BUT the answer is all about Smarty Marketing (even though not always correct because it does sync from that other website’s info) as it associates ‘Smarty’ with Smarty Marketing in the training data.”

The Primacy of Training Data in LLM Responses

Ann Smarty argues that the core of an LLM’s knowledge base, its training data, dictates its understanding and subsequent responses. Even when presented with current citations, the AI defaults to the associations and information embedded during its training phase. This leads to a scenario where the AI might generate information that is heavily influenced by, or even primarily about, entities that are strongly represented in its training data, regardless of the direct relevance of the provided citations.

As Ann Smarty notes, this has significant implications for how entities gain visibility through AI interactions:

“Been saying this for ages: What LLMs *know* about you (and how much) is the foundation of your answer visibility.”

This perspective challenges a common approach to AI visibility optimization, which often focuses heavily on ensuring a brand or website is cited in various sources. While citations are important for providing current context and specific data points, Smarty asserts they are secondary to the fundamental knowledge the LLM possesses.

Rethinking AI Visibility Optimization

Smarty suggests that professionals in the AI visibility optimization field need to broaden their strategies beyond merely focusing on citations. The underlying training data, and the strength of association an entity has within that data, forms the bedrock of its potential visibility in LLM-generated answers.

According to Ann Smarty, this means that efforts should also be directed towards influencing the AI’s foundational knowledge, a more complex but potentially more impactful endeavor. She elaborates on this by stating:

“If you are in the AI visibility optimization business and all you talk about is optimizing your site for citations, you are missing the foundation.”

In Ann Smarty’s view, understanding and leveraging the impact of training data is crucial for anyone seeking to ensure their brand or content is accurately and prominently represented by AI models. This nuanced understanding moves beyond surface-level citation strategies to address the deeper mechanisms driving AI comprehension and output.