What is LLM visibility? And why your website is no longer the whole story

What changed while nobody was watching

Somewhere in the last two years, a chunk of your buyers stopped typing keywords into Google and started asking questions in full sentences: "Suggest a reliable CRM implementation partner for a 60-person trading company in Kolkata." The answer comes back as three named companies and a paragraph of reasoning. There is no page two. There are no ten blue links. Either the machine knows you, describes you correctly and considers you worth naming, or you were never in the running.

That property, whether the machines know you, describe you accurately and recommend you, is what I call LLM visibility. The industry also uses AEO and GEO for overlapping ideas, and the vocabulary will keep churning for a while yet. The thing itself is stable, and it is measurable.

The three parts of the definition

Each part fails independently, which is why the definition has three parts.

Known. The assistant has your company as a distinct entity: a thing with a name, a place, an industry. If it has never retrieved enough about you to form that record, you are not described wrongly, you are simply absent. Absence is the most common state for Indian SMEs today.

Described accurately. The assistant knows you exist but its description is stale or wrong. This is the failure mode I see most in companies that have evolved: the machine describes what you were five years ago, because the internet's record of you is five years old. My own company is a live example, and I will come to it.

Recommended. The highest bar. When a buyer asks for a suggestion in your category, your name appears. This depends less on your website than most founders assume, and more on whether independent sources agree you belong in the answer.

Where do assistants get their information about you?

Three places, and the mix matters.

First, training data: the snapshot of the web the model learned from. You cannot edit the past, and its influence fades with every retrieval-backed answer, so this is the part to worry about least.

Second, live retrieval. When the question needs current facts, assistants search the web and read pages on the spot. ChatGPT leans on Bing's index, Gemini and Google's AI features lean on Google's index, Perplexity fetches pages in real time, and Claude searches and reads live too. This is where your own site matters: if it is crawlable, quick, and written so a machine can lift a clean answer from it, you are supplying the words the assistant uses about you.

Third, and this is the one founders underestimate, everyone else's websites. Directories, industry listicles, review platforms, news mentions, LinkedIn, Reddit and Quora threads. Before an assistant recommends a company it looks for agreement across independent sources. Your website is one witness in your own trial; the machine wants corroboration.

How is this different from SEO?

SEO gets you retrieved. LLM visibility gets you quoted and recommended. The foundation is identical, and none of it is wasted: crawlability, speed, clear pages, good information architecture. What sits on top is new. Passages must be extractable, meaning a machine can lift 40 to 60 words that answer a question completely. Your entity must be unambiguous, meaning schema and consistent descriptions across every surface where you appear (the glossary entry on entity SEO is the primer). And your off-site record must agree with your on-site claims, because the assistant checks.

One more difference worth naming: in classic search you competed for position on a page. Here you compete for inclusion in an answer. Position ten exists; "fourth company mentioned by ChatGPT" mostly does not. The contest is harsher, which is bad news in crowded categories and surprisingly good news in specific ones, where being the best-documented answer is achievable for an SME.

What poor LLM visibility looks like in practice

A worked example I can vouch for, because it is mine. In August 2026 I ran a baseline check on my own company, EbizIndia, across the major assistants. The good news: they identify me correctly as a person, after deliberate entity work I have written up separately in the namesake disambiguation case study. The bad news: several assistants describe EbizIndia with positioning that is years out of date, as a "software development and digital marketing" company. Serviceable words in 2015. Not what we do now.

Why? Because the assistants were reading old third-party pages about EbizIndia, not our current site and not LinkedIn. The record of you that machines trust is distributed across the internet, and it decays unless you tend it. This publication is, openly, part of how I am correcting that record. Watching whether it works, and how fast, is part of what this site will report.

The five levers, in the order I would pull them

• Be retrievable. Nothing else matters if crawlers cannot reach you. That includes the AI crawlers, which some sites block by accident; the robots.txt guide covers who to allow.
• Make answers extractable. On every important page, answer the visitor's question in the first two sentences, then elaborate. Machines quote pages that make quoting easy.
• Fix your entity. One canonical description of your company, in your site's structured data, repeated consistently everywhere you control.
• Build off-site agreement. Directories, reviews, mentions: the independent sources assistants use to corroborate. Slowest lever, biggest effect on recommendations.
• Stay current. Dated, updated content wins retrieval, especially on Perplexity. A page last touched in 2021 tells the machine your company might be too.

Where to start on Monday

Not with any of the levers. Start by measuring where you stand: ask the four major assistants about your company and your category, and write down what they say. It takes an hour and it converts this whole subject from anxiety into a task list. The step-by-step method is in How to check what ChatGPT, Gemini and Perplexity say about your company.