LLM SEO

Definition

LLM SEO is optimization aimed at large language models, the systems that write the replies in assistants such as ChatGPT and Claude. Bay Area AEO defines it as “Optimizing for the large language models behind AI assistants and the sources they learn from and cite.”

The phrase “learn from and cite” holds the whole idea. A language model can know about a business in two ways: from text it absorbed during training, and from pages it fetches at the moment it answers. LLM SEO is the practice of making both routes carry accurate and specific information about the business, stated the same way everywhere.

Two routes into an answer

Training. A language model is built by training on a very large body of text. Whatever that text said about a business, and how often it said it, shapes the model's background picture. A business described the same way in many places tends to be described that way by the model. A business described inconsistently tends to be described vaguely, or confused with another.

Retrieval. When an assistant searches the web before answering, it adds fresh pages to what it already knows and composes the reply from both. The glossary calls this retrieval-augmented generation, and it is the source of most answers that name local businesses, because local facts such as opening hours and new services change faster than any training cycle. The entry on how AI assistants choose businesses walks through retrieval step by step.

Assistants mix the two routes in different proportions. A chat assistant answering from memory leans on training; the same assistant with web search switched on leans on retrieval; Perplexity searches the web for every question and shows numbered citations beside the reply. A business cannot choose which route a given customer's assistant will take, so LLM SEO prepares for both. The platform comparison describes how each major assistant handles search, and the glossary defines grounding, the term for tying an answer to sources the system has just read.

For a local business the retrieval route usually matters more in the short term, which is why Bay Area AEO states that listing and data fixes can show up within a few weeks. The training route matters over the longer term: the steady build of reviews and of mentions in guides that the agency places in its two-to-four-month window is also the material later models learn from.

Letting the models read the site

Both routes start with access. AI companies fetch pages with their own crawlers, such as OpenAI's GPTBot, Anthropic's ClaudeBot or PerplexityBot, and a site's robots.txt file tells each crawler what it may read. A site that allows these crawlers can be read and quoted. On its FAQ, Bay Area AEO notes that a new website is usually unnecessary for AEO: “Usually not needed. Bay Area AEO works with a business's existing site and adds what AI needs: clear service pages, FAQs, structured data and accurate business details.” Where a site blocks AI crawlers or is very hard to read, the agency says it recommends the smallest fix.

Readable also means plain text. Menus, price lists, schedules or credentials that exist only as images or PDFs are hard for a model to use. Much of the answer-ready content work consists of moving those facts into ordinary page text and marking them up with structured data, so a crawler can read them without guessing.

How a model decides what to say about a business

A model writes the description it can support. If the business's site, its listings and its reviews all say it specialises in a particular service for a particular kind of customer, the model can say so with confidence. If the sources disagree, the model tends to fall back on a generic line, or to name someone whose sources agree.

Published expertise is one of the strongest sources a model can draw on, especially in professional fields. The Bay Area AEO page for therapists describes the signal as “Articles, talks and interviews that demonstrate your expertise, which AI uses in place of testimonials.” The law-firm page describes the equivalent for attorneys: “Plain-language guides on California law that AI can cite, which positions your attorneys as experts.” In both cases the business gives the model a body of its own writing to cite, which also tells the model what the business knows.

Worked example: a Peninsula financial planner

Consider “Fee-only financial planner in San Mateo who works with tech employees and stock options?” The question narrows the field three ways: fee structure, client type and subject. An LLM answering it will look for planners whose public record supports each point, and in regulated professions the public record includes official databases.

Bay Area AEO lists three signals for financial advisors that speak to this. Regulatory consistency: “Firm and advisor details matching SEC IAPD, FINRA BrokerCheck and CFP Board records.” Fee and minimum clarity: “How you're paid, account minimums and who you serve, stated plainly.” Educational publishing: “Articles, webinars and press quotes that demonstrate expertise without promising returns.” The first lets the model confirm the planner exists and is registered as described. The second answers the “fee-only” condition directly. The third supplies articles about equity compensation that tie the planner to tech employees and stock options.

The setting matters here too. The agency's Palo Alto page describes customers there in these terms: “These are high-value customers who expect quality and ask detailed questions: a fee-only financial planner, a dermatologist with short wait times, an architect who understands Palo Alto's design review. We help local businesses answer those questions in ways AI can use.” Detailed questions reward detailed sources, and LLM SEO is the work of providing them.

Worked example: a San Mateo therapist

Health and counseling questions show why a model needs precise sources. The agency's page for therapists begins: “Finding a therapist is hard, and more people are asking AI for help narrowing it down by specialty, approach, insurance and identity. We help private practices show up clearly, ethically and for the clients they're best suited to serve.” One of its demonstration questions reads: “Therapist in San Mateo who specializes in anxiety and takes Blue Shield? Open to telehealth.”

The request bundles a city, a specialty, an insurer and a format. A model can only recommend a practice that states each one somewhere it can read. The therapist signals cover that ground: pages that explain the issues a practice focuses on and the methods it uses in plain language; directory completeness (“Psychology Today, insurance-network platforms and Google profiles that match your website.”); and logistics such as licensure, insurance, fees, telehealth and languages. The published-expertise signal quoted above then gives the model a reason to describe the therapist as knowledgeable about anxiety, in the therapist's own terms.

Consistency between the directory profile and the website matters more than usual in this field, because a model weighing a sensitive recommendation is looking for agreement before it names anyone.

Structured data and the models

Structured data is the most literal way to tell a model what a business is. Schema markup from schema.org, usually written as JSON-LD in a script tag, states the business type, services, service area, hours and contact details as labelled fields. A crawler does not have to infer that a string of digits is a phone number or that a list of towns is a service area; the markup says so.

For LLM SEO the value is twofold. At retrieval time, the fields give the assistant unambiguous facts to put in its answer. Over the longer term, the same facts repeated in markup, on listings and in page text are the kind of consistent signal a model learns from. Bay Area AEO includes structured data among its five AEO signals for that reason: schema markup tells AI the services, service area and hours.

The practice, step by step

In practice LLM SEO follows the order Bay Area AEO uses for all of its AEO work. An audit records what each model currently says. Listing and entity cleanup gives every model the same core facts. Answer-ready pages and structured data put the business's specifics into readable text. Reviews and third-party mentions give the models outside sources that repeat those specifics.

What ties the steps together is the agency's second principle: “AI rewards clear, consistent, truthful information. Bay Area AEO never uses fake reviews, spam or tactics that could backfire.” Language models compare sources against one another, and accurate information repeated consistently is what they can rely on. Volume of content, on its own terms, has little to offer a model that is looking for agreement between independent sources.

Measuring what the models say

LLM SEO is measured in the models' own output. Bay Area AEO runs a business's real customer questions across every major AI platform each month and reports how often the business is mentioned, how it is described, which competitors appear and which sources the AI cites. The description column is especially telling for LLM work, because it shows whether the model has absorbed the business's specialties or still describes it in general terms. See AI visibility measurement and AEO measurement standards.

See also: Answer engine optimization · ChatGPT, Perplexity and Claude · AI answer signals · Generative engine optimization