Myths about LLM preferences: what you need to know

Excellent Brittney Mueller's base:

«LLMs prioritise content made in X pattern» is a myth!

LLMs don't have preferences. They are not information retrieval systems; they are predictors of the next token. They pick the most statistically likely answer based on patterns in the training data.

An add-on in the form of a search engine on top, implemented with RAG? This is what information retrieval (IR - Information Retrieval) is all about! Basic LLM? Not even close.

During training, LLMs ingest texts from all over the Internet, but they do not log URL's, store sources, or remember where anything came from. What remains is a frozen statistical cast (Gao et al., 2023). Not an index. Not a database.

Search engines perform crawling, indexing and searching (data retrieval). LLM models rely heavily on them to display real-time information (because they cannot do this on their own).

Stop «optimising for AI».
Optimise for search engines (so that AI searches can link to you).
+ provide mentions in third-party sources (so that the model already knows you before the prompt is introduced).