Generative self-retrieval · Recall from model memory
Generative self-retrieval is when an LLM “retrieves” information from its own memory (its weights), with no external source, before writing its answer. It is what lets a model mention a brand even when no web search runs.
Standalone definition · 37 words
How generative self-retrieval works inside an LLM
During training, the model stores associations between words, entities and facts in its weights. Work such as Geva et al. (2021) shows feed-forward layers act like key-value memories. When answering, the model recalls those associations, like an internal search.
Why generative self-retrieval matters for GEO
Part of AI answers never touch the web. Dan Petrovic (DEJAN) stresses this point: if your brand is not well anchored in the model’s memory, it will not show up in those answers, even with strong SEO. RAG only helps when a search runs.
How to strengthen your presence in model memory
Get mentioned often, next to the right topics, on sources that end up in training data (press, Wikipedia, forums, Common Crawl). Consistency matters: same name, same description, same categories everywhere.
Frequently asked questions about Generative self-retrieval
Is generative self-retrieval the same as RAG?
No. RAG fetches external documents. Self-retrieval draws on the model’s internal memory.
Can you measure what a model knows about your brand?
Yes, by querying the API with web search off and repeating prompts. Some go further with logprobs.
Terms related to Generative self-retrieval
Further reading on Generative self-retrieval
Sources for the Generative self-retrieval definition
- Geva et al., “Transformer Feed-Forward Layers Are Key-Value Memories” (2021)
- Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” (2020)
- Dan Petrovic (DEJAN), work on model memory and brand visibility (2025)