LLMO

LLMO (Large Language Model Optimization) lives in the same family as GEO and AEO but tilts toward a different question — not 'do we appear in AI search?' but 'what does the model actually know about our brand?'. The headline KPI is what ChatGPT, Claude or Gemini say when someone asks 'what is X?'. Surfaces come and go, but the brand facts encoded in a model's parametric memory tend to stick, which makes LLMO closer to long-term asset management than to a single campaign.

Mechanically there are two inputs to think about. Pretraining data — Wikipedia, press, reputable directories, well-structured brand sites — bakes facts into the model's weights. Inference-time retrieval (search tools, RAG) refreshes and corroborates those facts at query time. Good LLMO work shapes both, so the brand is described the same way whether the model is recalling from memory or fetching live.

Marketers should internalise that half of LLMO is decided off your own site. Wikipedia accuracy, consistent press descriptions, category mappings on review platforms, structured listings on industry databases — all of it feeds the model. Useful KPIs are brand-definition alignment, category-prompt presence, and factual-error rate. Villion runs identical brand prompts across major LLMs and scores definition alignment so drift becomes visible week to week.

Compared to GEO and AEO at surface level, the difference is one of focus. GEO and AEO ask whether you are cited in answers from ChatGPT, Perplexity or AI Overviews. LLMO asks whether the surrounding prose describes your brand correctly. A citation with the wrong category is a GEO win and an LLMO loss — and you need both lenses to make sense of it.

A frequent objection is 'we cannot touch training data, so why bother?' True in the literal sense, but the signals training pipelines absorb are almost all under marketing control. Another myth is that a single correction propagates fast. Pretraining cutoffs and live retrieval operate on different clocks, so updates show up unevenly — measure consistently rather than after a single fix.

Reasonable next steps: audit Wikipedia and top-tier press for description drift, standardise the definition sentences on your About, Brand and Press pages, harden the Organization schema that feeds the Knowledge Graph, and run a quarterly evaluation across four or five major LLMs with a fixed prompt set. LLMO is less a sprint than an ongoing yearbook of brand facts.