Large Language Model Optimization (LLMO)

How is LLMO different from SEO?

SEO earns a position in a list of ten links. LLMO earns a mention inside a written answer. That sounds like a small difference and it changes almost everything about the work, because the model doesn't rank your page; it reads it, decides whether it answers the question, and either repeats what it found or moves on.

The overlap between the two is real but smaller than most owners assume. Ahrefs (August 2025, 15,000 long-tail queries) found only 12% of the URLs AI engines cite rank in Google's top 10 for the same prompt: 28.6% on Perplexity, 8.6% on Copilot, 8.2% on Gemini, and about 8% on ChatGPT. A first-page Google rank buys you very little inside an AI answer. Ahrefs also tracked the share of Google AI Overview citations that come from top-10 pages falling from 76% to 38% between July 2025 and March 2026, which Ahrefs attributes to Google's query fan-out: the engine runs several related searches and stitches the answer from all of them.

So the SEO habits still matter (crawlable pages, clear titles, real content), but they are the entry ticket, not the prize. The prize goes to whoever gives the model a specific, quotable answer to the specific question.

How do large language models pick which businesses to name?

Each engine retrieves differently, and they disagree with each other far more than they agree. Profound's analysis of 100,000 prompts found ChatGPT and Perplexity overlap on only 11% of the domains they cite; 37.4% of citations were ChatGPT-exclusive and 51.6% Perplexity-exclusive. Even Google's own two surfaces diverge: Google AI Overviews and Google AI Mode cite the same URLs only 13.7% of the time (Profound). Optimizing for one model does not automatically help with another, which is why any serious LLMO effort tracks each engine separately.

What they have in common is a preference for sources they can trust and quote. Yext's analysis of 6.8 million citations found 86% came from brand-managed sources, where brand-managed bundles owned websites (about 44%) with managed listings like Yelp and Google Business Profile (about 42%). For a local business that means two things carry most of the weight: your own site, and the listings and review profiles the engines read alongside it.

It also means the aggregators are in the room. Yelp disclosed a data-licensing agreement with OpenAI in February 2026. Thumbtack was integrated into ChatGPT for home-services booking in October 2025. When the engine can answer from a licensed directory, a thin website loses to a complete Yelp profile.

What does an LLMO service actually do?

An LLMO service (you will also see it sold as a GEO agency, an AEO platform, or AI visibility software) does some combination of four jobs. Measurement: asking the engines the questions your customers ask, repeatedly, and recording who gets named, in what order, and with what language. Source analysis: working out which sites the engines cite when they answer, so you know whether the fix is your website, your Google profile, or a directory listing. Content: publishing the pages that answer the losing questions in plain, quotable language. And technical access: making sure the crawlers that feed the engines (GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended) can actually read your site.

Prices in this category run from about $29 a month for a starter tracker to enterprise contracts with no public price. Most tools were built for marketing teams: you write and maintain the prompt list, you read the dashboard, and you implement the findings on whatever platform hosts your site. Tibly's version is built for the owner: it generates the prompts from your industry and service-area cities, scans them daily across seven engines, and because Tibly also hosts your website, it drafts and publishes the page that closes a gap without a handoff. Plans start at $149 a month, and the free report at /ai-visibility shows where you stand before you pay anything.

How do you do LLMO yourself?

You can do the core of it without a tool, as long as you are honest about what you can and cannot measure by hand. The work, in the order it pays off:

• Answer the real questions in plain text. Prices, service area, hours, insurance accepted, licenses, response times. A model cannot quote a PDF menu or a “call for a quote” button. Put the answer in a sentence on a page.
• Make one page per question family. An emergency page, a pricing guide per big-ticket service, an insurance or fee page, a service-area page per city you actually serve. Generic pages lose to specific ones.
• Fix the sources the engines already trust. Complete your Google Business Profile, keep Yelp and the trade directories current, and respond to reviews; the engines read the responses as evidence of how you handle problems.
• Let the crawlers in. Check robots.txt for rules that block GPTBot, ClaudeBot, or PerplexityBot, and check whether your CDN's bot protection is turning them away at the edge. A site the crawler can't read cannot be cited.
• Skip the folklore. No AI system currently reads llms.txt (Google's John Mueller compared it to the old keywords meta tag), and there is no evidence that schema.org markup by itself earns citations. Schema may help indirectly by improving classic indexing, and that is the honest limit of the claim.
• Measure with repetition. One search proves nothing; answers change run to run. Ask the same question several times per engine, on a schedule, and track the share of answers that name you.

What does LLMO look like in practice?

One first-party example, attributed as Tibly scan data pulled 2026-09-23: East 2 West Collective, a boutique PR agency in Los Angeles and New York, launched a site generated on Tibly on 2026-08-05 and started tracking 15 buyer questions across three engines. On 2026-08-27 the agency was named in 0 of 30 answers. By 2026-09-23 it was named in 16 of 45, about 36%, and in position one in 8 of those. Over the same window, 34 pages answering the losing questions were published to the site and citations of e2wcollective.com in the engines' answers went from 0 in the week of August 24 to 121 in the week of September 14. The engine set changed once mid-series (Google AI Mode replaced Perplexity for two weeks), so the combined series is directional rather than a clean control.

That is the whole discipline in one case: find the questions you lose, publish the specific answer, and watch the share move. The label on the invoice (LLMO, GEO, AEO) matters far less than whether the loop runs.