LLM SEO and How Language Models Decide What to Mention
LLM SEO (also called LLM optimization or LLMO) is the practice of influencing what large language models such as ChatGPT, Claude and Gemini say about your brand. A model mentions brands from two sources: its parametric memory, meaning what it learned during training, and retrieval, meaning pages it finds in a live web search for that question. LLM SEO works very differently on each, so the first step is knowing which one produced the answer you are looking at.
This guide explains both memory systems, what each implies for a brand, and how to tell them apart when you measure. It stays on the model side of the picture. For the wider discipline and its other levers, see our pillar on generative engine optimization.
What LLM SEO means
Classic SEO tries to change where a page ranks. Doing SEO for LLMs means trying to change what a model says: whether it names your brand, how it describes you, and whether it links to your site. The term overlaps heavily with GEO and AEO, and our explainer on SEO vs AEO vs GEO and the other acronyms sorts out the labels. What sets the LLM SEO framing apart is its focus on the model itself, including knowledge that no search result can change on the day.
That focus matters because a language model does not always search. When it answers from memory, there is no ranking to win. The only way in is to have been part of what it learned.
The two memory systems behind every answer
The distinction comes from machine learning research. The 2020 paper that introduced retrieval-augmented generation, by Patrick Lewis and colleagues, described pre-trained models as storing factual knowledge in their parameters, a “parametric memory”, and paired them with a searchable index as a “non-parametric memory” (Lewis et al., 2020). The paper also named the limits of parametric memory: it is hard to update and hard to trace back to a source. Today’s AI assistants use the same two-part design at web scale.
| Attribute | Parametric memory (training data) | Retrieval (live search) |
|---|---|---|
| Where knowledge lives | Inside the model’s weights | In a search index queried at answer time |
| When it was formed | At training, up to the model’s cutoff date | Seconds before the answer |
| Visible sources | None; the model cannot cite its training data | Usually shown as citations or a sources panel |
| How fast it changes | Only with new model versions or fine-tunes | As fast as pages are crawled and indexed |
| What drives inclusion | How often and how consistently your brand appears across the web | Whether your pages are retrieved and judged useful for the query |
| Typical failure | Unknown brand, outdated facts, wrong category | Page blocked, not indexed, or outranked by third-party pages |
| Crawler that feeds it | Training bots such as GPTBot and ClaudeBot | Search bots such as OAI-SearchBot, Claude-SearchBot and PerplexityBot |
When a model answers from memory and when it searches
Most assistants decide per question whether to search. Google’s Gemini documentation describes the flow: the model judges whether a search would improve the answer, generates queries if so, and returns an answer with citations. Dan Petrovic of DEJAN reported that Gemini’s grounding scores each query from 0 to 1 and searches above a default threshold of 0.3, with recent-event and accuracy-sensitive questions scoring higher.
In practice, the pattern looks like this:
- Usually from memory: definitions, general how-it-works questions, well-known category questions such as “what is a CRM”.
- Usually searched: prices, “best X in 2026”, local options, recent news, product comparisons, anything with a date.
- Always searched: engines built around retrieval, such as Perplexity and Google’s AI Overviews and AI Mode, which draw on a search index for nearly every answer.
The same brand can therefore look current on one engine and out of date on another. Duane Forrester calls an engine’s default lean its “memory posture” and argues it has to be measured per engine rather than assumed.
What parametric memory means for your brand
A model learns about a brand by seeing it many times, in many contexts, across its training data. Research on long-tail knowledge shows why that frequency matters. Nikhil Kandpal and colleagues found that a model’s ability to answer a factual question tracks how many pre-training documents mention the entities involved, and that facts with little support are learned poorly (Kandpal et al., 2023). A small brand mentioned on a handful of pages sits in that long tail.
The implications are practical:
- Volume and spread of mentions matter. Being discussed across many independent sites gives a model more chances to learn the association between your brand and your category. This matches Ahrefs’ finding that branded web mentions correlate with AI Overview visibility far more strongly than backlinks do.
- Consistency matters. If your brand is described as a “project tool” in some places and a “CRM” in others, the model learns a blurred picture. Use the same category, audience and core facts everywhere.
- Changes are slow. Petrovic estimates that influencing training data takes about three to six months to show in fine-tunes and sub-releases, and about a year for major model releases. He also notes that a model’s built-in view of a brand, its “primary bias”, is hard to shift.
- Blocking training bots is a trade-off. Disallowing GPTBot or ClaudeBot keeps your own pages out of future OpenAI or Anthropic training. It does not stop other sites from describing you, and it does not affect live search inclusion in ChatGPT or Claude, which use different bots. Google-Extended is a special case: it also governs grounding in the Gemini app, but not AI Overviews or AI Mode. Our list of AI crawlers shows which token controls what.
What retrieval means for your brand
Retrieval is closer to classic SEO, but with different rules for winning. The model searches, reads passages from the top results and decides what to use.
- Access first. Allow each engine’s search crawler and keep key content in the raw HTML, since many AI fetchers do not run JavaScript.
- Be in the index the engine uses. Google’s AI features use Google’s index. ChatGPT citations have tracked Bing results more closely than Google’s in third-party tests, and Perplexity runs its own index.
- Win the sub-queries. Engines rewrite a prompt into several searches. Pages that answer specific follow-up questions get more chances to be retrieved.
- Write liftable passages. A model reads chunks, not whole pages. A direct answer with a specific fact near the top of a section is easier to use than a conclusion buried at the end.
- Third-party pages count as retrieval too. If the top results for “best X” are review sites and listicles, being described well on those pages gets you mentioned even when your own site is not cited.
Unlike training data, retrieval fixes can show up within days or weeks of a page being recrawled, which makes retrieval the faster lever for most teams.
How to tell which memory produced an answer
When you test your brand in an assistant, check these signals before deciding what to fix:
| Signal in the answer | Likely source | What to fix |
|---|---|---|
| Citations or a sources panel are shown | Retrieval | Access, indexing, passage quality, third-party pages ranking for the prompt |
| No sources, and facts match an old version of your site | Parametric memory | Consistent, widespread current facts; wait for model updates |
| Brand missing without search, present with search | Weak parametric memory, decent retrieval | Build off-site mentions to strengthen the brand-category link |
| Brand present without search, missing with search | Strong memory, weak retrieval | Check bot access and whether your pages rank for the sub-queries |
| Wrong facts even with citations | Retrieval of inaccurate third-party pages | Correct the facts on the cited sources |
Petrovic recommends comparing grounded and ungrounded answers separately for the same reason: a single blended score hides which layer moved.
How to measure LLM visibility
LLM visibility means how often and how favorably language models mention your brand for the prompts that matter to you. Because answers vary from run to run, measure rates across repeated runs, not single screenshots. A workable routine:
- Write a fixed set of buyer prompts, mixing questions likely answered from memory with questions likely to trigger search.
- Run each prompt several times per engine and record mentions, citations, position and sentiment.
- Tag each answer as cited (retrieval) or uncited (likely memory) and report the two groups separately.
- Re-run weekly, and note model releases so you can link jumps to updates rather than to your own changes.
Our guide to AI visibility metrics defines each number in detail. Asheard runs this kind of tracking across ten engines, including ChatGPT, Claude, Gemini, Perplexity, DeepSeek and Mistral, and records mentions, citations, position and sentiment per prompt. See the AI visibility tracking page for how it works.
Frequently asked questions
What is LLM SEO?
LLM SEO is the practice of influencing what large language models such as ChatGPT, Claude and Gemini say about your brand. It covers two layers: the knowledge a model learned during training, and the pages it retrieves in a live web search. Training data responds to widespread, consistent brand mentions over months, while retrieval responds to accessible, well-structured pages within days or weeks.
Is LLM optimization different from SEO?
Partly. The retrieval side of LLM optimization looks a lot like SEO, because models search indexes such as Google, Bing or their own before answering. The training-data side has no SEO equivalent: there is no ranking to win, only how often and how consistently your brand appears across the web when the next model is trained.
Can I get my brand into ChatGPT’s training data?
Not directly. There is no submission process. A brand enters training data by being written about on many public pages that training crawlers collect, such as reviews, news coverage, forums and directories. Allowing training bots like GPTBot to read your own site adds your pages too. Changes typically show only when a new model version is released.
Should I block AI training crawlers?
It depends on your goals. Blocking training crawlers such as GPTBot or ClaudeBot keeps your pages out of future training sets, but other sites can still describe you. It does not remove you from live AI search, which uses separate bots. Many brands that want visibility allow search bots and decide on training bots as a separate policy question.
