LLM optimization (LLMO): the practitioner's guide

Large Language Model Optimization (LLMO) is the practice of influencing how an AI model represents your brand in its own knowledge — the description it produces when it is not reading a single page from the live web. Ask a model "who are the main vendors for X" with browsing disabled and whatever comes back is parametric knowledge: patterns absorbed from the public text the model was trained on.

That answer is not a ranking you can bid on and not an index you can submit to. It is a statistical summary of how the internet talks about your category. LLMO is the work of changing that conversation in enough places, consistently enough, that the summary changes with it.

Every AI answer your buyer sees is assembled from two sources, and they behave completely differently.

• Parametric knowledge lives in the model's weights. It is stable, fast, applied to every query, and only changes when the model is retrained or refreshed. This is LLMO territory.
• Retrieved knowledge is pulled at query time from search indexes, crawlers, or a vendor's own corpus. It changes within days of publishing and carries citations. This is GEO territory.

The practical consequence: retrieval decides who gets cited, but parametric knowledge decides who gets considered. If a model has never internalized that you belong in a category, retrieval often never goes looking for you — the query it generates does not include your name, and your carefully optimized comparison page is never fetched.

These are not competing strategies. They are four layers of the same funnel: SEO makes the page findable, AEO makes it extractable, GEO makes it retrieved and cited, and LLMO makes your brand something the model already knows to look for. Our AI visibility tracking guide covers the measurement layer that sits under all four.

Models learn from repetition across independent sources. That single fact rules out most of what marketing teams instinctively reach for — more owned content, more first-person claims — and points at four levers that actually compound.

1. Entity consistency. One brand name, one category descriptor, one boilerplate sentence — identical on your site, LinkedIn, Crunchbase, review profiles, press releases, conference bios, and partner pages. Models resolve entities by co-occurrence; three different self-descriptions produce a blurry entity that gets summarized as "a marketing tool."
2. Structured public data. Wikidata entries, well-maintained directory and marketplace listings, and machine-readable organization schema give models an unambiguous anchor for who you are and what category you sit in.
3. Third-party coverage in other people's words. Analyst notes, roundups, podcast transcripts, integration partner docs, and customer write-ups carry far more weight than your own copy because they are independent repetitions of the same claim.
4. Durable community text. Reddit threads, Stack Overflow answers, Hacker News discussions, and niche forums are heavily represented in training corpora and read as consensus. Being named in the "what are people actually using" thread is one of the highest-leverage LLMO assets available — and it is why we track Reddit gaps as a first-class signal.

1. Baseline the model's belief. Run your buyer's real prompts against ChatGPT, Claude, Gemini, and Perplexity with browsing off. Record who is named, in what order, and how you are described when you appear.
2. Audit the entity surface. Collect every public description of your company and flag mismatches in name, category, and positioning. Fix the mismatches before publishing anything new.
3. Fix the anchors. Wikidata, organization schema, directory listings, and your own about page — make them agree, word for word where possible.
4. Earn independent repetition. Target the specific sources that already get cited in your category's answers: the roundups, the review sites, the two or three threads that keep surfacing. Volume matters less than source diversity.
5. Re-baseline on a schedule. Repeat step one monthly. Movement in browsing-off answers is the only honest LLMO signal, and it is slow enough that a quarterly view is more informative than a weekly one.

The measurement trap in LLMO is treating a single chat session as evidence. Model outputs are sampled, so the same prompt can name you once and skip you the next time. You need repeated runs of a fixed prompt set to get a rate rather than an anecdote.

• Mention rate, browsing off. Share of runs where the model names you without retrieval. This is the closest proxy for parametric presence.
• Description accuracy. Does the model place you in the right category and attach the right capabilities? A wrong category is worse than absence.
• Position in the list. Being named fourth is materially different from being named first when the buyer only reads a shortlist.
• Retrieval delta. Mention rate with browsing on minus mention rate with browsing off. A large delta means your visibility is entirely rented from retrieval and disappears the moment a source drops.

Monroya runs fixed prompt sets across all four providers on a repeating schedule and splits results by buyer stage, which is what makes the browsing-on/browsing-off comparison usable rather than anecdotal. The methodology page documents the sampling.

• Publishing your way out of it. Owned content is discounted as self-description. Ten more blog posts rarely change what a model believes; one analyst roundup and three community threads might.
• Renaming and repositioning constantly. Every change of category language resets the co-occurrence signal that was starting to accumulate.
• Optimizing for one model. Providers weight sources differently. Work that only lands in one model's answers is usually a retrieval quirk, not an LLMO win.
• Expecting GEO timelines. Judging an LLMO program after three weeks guarantees you conclude it does not work.
• Prompt-stuffing tricks. Hidden text and instruction injection aimed at models are transient at best and reputationally expensive when found.