How do LLMs secretly change? Tracking invisible shifts in query fan-outs and source retrieval linkedin.com

LLMs are not static systems. They continuously evolve how they expand queries, retrieve information, and decide which sources to use when generating answers.

In SEO, we closely monitor Google Core Updates because we know that even small changes in ranking systems can have a significant impact on visibility. The same mindset needs to be applied to AI Search.

In this article, I analyze several patterns I observed in ChatGPT and explore how subtle changes in retrieval behavior can influence brand visibility and source selection.

The analysis covers:

How changes in Query Fan-outs can impact which domains are retrieved
Why Reddit suddenly gained visibility through a temporary shift in query patterns
How the increasing use of terms like “Official” changed source retrieval behavior
Why arXiv.org gained share as a reference source while Wikipedia declined
How ChatGPT 5.5 and ChatGPT 5.6 Luna show different Query Fan-out and retrieval strategies
Why iterative retrieval processes and precise search patterns (such as site operators and quoted terms) could make first-party content more important again

The key takeaway: AI Search visibility is not only influenced by the content we create. It is also shaped by how LLMs themselves evolve.

To understand changes in visibility, we need to look beyond the final answer and start analyzing what happens behind the scenes: query expansion, retrieval behavior, and source selection.