What it means
LLMs (large language models) are AI systems trained on large amounts of text to predict and generate language. Their knowledge is frozen at training time. To answer current or specific questions, an LLM retrieves outside sources at query time and grounds its answer in what it pulls back.
Why it matters
Because an LLM's parameters are fixed after training, you can't optimize the model itself. The leverage is what an LLM retrieves and cites at query time, not what it already knows. Many LLMs use search engines as retrieval sources, so ranking in the top 10 raises your probability of entering the candidate pool and getting cited.
Say a mid-market B2B SaaS company spends a quarter trying to "get into ChatGPT" by seeding brand mentions across the web. Nothing moves, because the model was trained months earlier. When the same team instead earns top-10 rankings for 30 long-tail question variations of its core topic, its citation rate climbs from near zero to showing up in 4 of 10 test prompts.
How to use this knowledge
Stop trying to influence the model's weights and optimize the retrieval layer instead. Rank for many fan-out query variations, beyond the head terms, so more of your pages qualify for the candidate pool. Use internal links to signal entity relationships and topical authority. Track citations by model, since grounding varies from one model to the next and not every model prioritizes Google's top results.
Growth Memo guidance
Many LLMs use search engines as retrieval sources. Higher organic rankings increase the probability of entering the LLM's candidate pool and receiving citations. — State of AI Search Optimization 2026
Search engines and LLMs are mapping relationships between entities and judging your brand's authority accordingly. — Internal Linking Grows Up: Evolving from Link Juice to Entity Maps
To earn lasting visibility, and not short-term visibility bought by hacky LLM visibility tricks, your brand needs to signal to search engines and LLMs that it's an authority in topics related to your offerings. — Operationalizing your topic-first SEO strategy
Query fan-out — how a model splits one prompt into many sub-queries, each a fresh retrieval chance.
Candidate pool — the set of pages a model can cite, which you enter by ranking in the top results.
Retrieval-augmented generation (RAG) — the mechanism that lets an LLM ground answers in sources fetched at query time.
Topical authority — the signal LLMs use to judge whether your brand is worth citing.
Entity maps — the relationships an LLM forms from your internal links.