LLM SEO: How Language Models Pick Which Businesses to Mention

Woman in profile with a circuit pattern over her head, on an orange background

When you ask ChatGPT, Claude, Gemini or Perplexity for "a good accountant for a small online shop," the assistant does not consult a ranking. It reads, in the span of a few seconds, a handful of pages and listings, and writes a paragraph naming two or three businesses. LLM SEO is the practice of being one of them. The term is new; the mechanics are surprisingly transparent, because every major assistant documents how it reads the web. This article walks through that machinery and turns it into a to-do list.

What an LLM Does When Asked for a Recommendation

A large language model (the LLM behind each assistant) has two sources of knowledge about your business:

  1. What it learned in training. A snapshot of the public web, frozen months before you asked. Useful for well-known brands, nearly useless for a small business with a two-year-old website.
  2. What it reads right now. When web search is on, the assistant runs searches, fetches a few pages and composes an answer from them, usually with citations. For a small business this is the source that matters, because it is current and it can be influenced this month.

Each company publishes the bots that do the second job, and each separates them from the training crawlers:

Assistant Search and user-fetch bots Training crawler Source
ChatGPT OAI-SearchBot ("used to surface websites in search results in ChatGPT's search features"), ChatGPT-User (visits a page when "users ask ChatGPT a question") GPTBot OpenAI
Claude Claude-SearchBot ("navigates the web to improve search result quality for users"), Claude-User (fetches content "in response to a user query") ClaudeBot Anthropic
Perplexity PerplexityBot ("designed to surface and link websites in search results on Perplexity. It is not used to crawl content for AI foundation models"), Perplexity-User none listed for that purpose Perplexity
Gemini and Google AI Overviews Googlebot, the normal Search index Google-Extended (training and grounding for Gemini apps; "does not impact a site's inclusion in Google Search") Google

Two conclusions fall out of the table. First, you can block training and still be recommended, because the search bots are separate. Second, if a "block all AI" setting in a plugin has closed the door on the search bots, no amount of content will get you mentioned. Check your robots.txt before anything else.

How the Model Decides Who to Name

Nobody outside the labs knows the exact weighting, and anyone who claims to is guessing. But the behavior is consistent enough to describe, and it matches the only peer-reviewed study on the subject, the GEO paper presented at KDD 2024, which found that adding citations, quotations and statistics to a page raised its visibility in generative answers "by up to 40%," while keyword stuffing did not help (arXiv 2311.09735).

In practice a model names a business when:

  • The fetched pages agree. Your site, your Google Business Profile, a directory and two reviews all say you are an accountant for online shops in a particular area. Agreement makes the model confident; contradiction makes it hedge or pick someone else.
  • There is a quotable sentence. Models write by paraphrasing what they read. "We prepare annual accounts for online shops, fixed fee quoted in advance, within ten working days" gives the model something to say. "Tailored financial solutions for ambitious businesses" gives it nothing.
  • Someone other than you says so. Reviews, a local news mention, a supplier's case study, a forum answer. Models weigh third-party confirmation the way Google weighs links.
  • The page is readable in one pass. Text in the HTML, not behind a login, a popup or a script that never finishes loading. The user-fetch bots are quick visitors.
  • The page is recent, or at least dated. A visible update date and current facts beat an undated page from 2021.

Why Classic SEO Still Carries Most of the Weight

The assistants search before they read, and most of them lean on conventional search indexes to decide which pages to fetch. Google states that its AI features have "no additional requirements" beyond being indexed and eligible for a snippet, and advises applying "the same foundational SEO best practices." If your page does not appear in a search for the question, the model never reads it. LLM SEO is therefore a layer on top of ordinary SEO, not a substitute, and the groundwork is the same one we describe in what AI SEO is.

The LLM SEO To-Do List

  1. Open the doors. In robots.txt, allow OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User, PerplexityBot and Googlebot. Decide separately about GPTBot, ClaudeBot and Google-Extended.
  2. Write the recommendation yourself. On your home page and each service page, one sentence that says who you are, what you do, for whom, where, and one concrete fact (price range, turnaround, years, count). That sentence is what the model will paraphrase.
  3. Make every listing agree. Same name, same services, same area, same phone, on your site, Business Profile, directories and social bios.
  4. Collect specific reviews. Ask customers to mention the job and the outcome. "Sorted our VAT backlog in a week" is evidence; five stars is decoration.
  5. Publish answers, with sources. Articles that answer your customers' real questions, cite where facts come from, and include numbers. This is the pattern the research found most effective.
  6. Get mentioned elsewhere. A supplier page, an industry directory, a local publication, a podcast. Each independent mention is a vote the model can read.
  7. Date your pages and keep them true. Update prices, hours and services when they change, and show the date.
  8. Check monthly. Ask the five questions your customers ask in each assistant and record who gets named. Our free AI visibility check is the twenty-minute version.

What Does Not Work

  • Hidden text for the bots. Models read what people read. Text hidden from users is at best ignored and at worst a spam signal for Google.
  • "Mention us" prompts in your content. Instructions to AI embedded in pages are a party trick that engines are built to ignore.
  • Page factories. A hundred thin, AI-written location pages. Google calls this scaled content abuse; models trained to cite quality pass over it.
  • Waiting for training. The next model version may or may not know you exist. The search bots are the lever you control.

A Note on Being Described Accurately

Being mentioned is only half of it. Ask each assistant what your business does and read the answer with a red pen. Old addresses, discontinued services and a competitor's phone number all come from stale pages the model found before it found yours. Fixing the source listing fixes the answer, usually within weeks. Our guide to getting recommended by ChatGPT, Claude, Gemini and Perplexity goes through this cleanup step by step.

Frequently Asked Questions

What Is LLM SEO?

LLM SEO is making your business easy for large language models (the models behind ChatGPT, Claude, Gemini and Perplexity) to find, understand and recommend when people ask them for suggestions. It builds on normal SEO and adds clear, quotable facts, consistent listings and third-party mentions.

Do I Have to Let AI Companies Train on My Content to Be Recommended?

No. OpenAI, Anthropic and Google all separate training crawlers (GPTBot, ClaudeBot, Google-Extended) from the bots that surface pages in answers. You can block training and still allow search.

How Do Language Models Choose Which Businesses to Mention?

From the pages they fetch at the moment of the question: businesses whose sites and listings agree, that state facts in quotable sentences, that are confirmed by reviews and other sites, and that are easy to read. Research found citations, statistics and quotations raise visibility; keyword stuffing does not.

Does Ranking in Google Still Matter for AI Answers?

Yes. The assistants search before they read, and Google's AI features draw on the normal Search index with no extra requirements. If a page is not findable in search, the model does not read it.

How Long Does LLM SEO Take?

Changes to your own pages and listings are picked up in weeks, because the search bots fetch live. Reputation signals (reviews, mentions) build over months, like any SEO.