LLM Visibility: What It Is and How to Measure It Properly

LLM visibility is how often large language models such as ChatGPT, Gemini, Claude and Perplexity name your business when people ask about your category. Most guides stop at the definition. This one shows how the models choose, the exact formulas, how many answers you need before a number means anything, and what changes the result.

Tomek Kuźniar · 8 min read · updated

LLM visibility is the share of answers from large language models that mention your business when people ask about your category. If ChatGPT, Gemini, Claude and Perplexity get 100 questions like "who is the best accountant in Bristol for freelancers?" and name you in 18 of them, your LLM visibility for that set of questions is 18 percent.

The term means the same thing as AI visibility or AI search visibility. People who work in SEO tend to say "LLM visibility", and business owners tend to say "does ChatGPT recommend me". This guide covers how the models decide who to name, the formulas worth using, how many answers you need before a number means anything, how the data is collected, and what moves the result.

What LLM visibility includes

A useful measurement has five parts. Each one answers a different question:

  • Mention rate: are you in the answer at all? This is the headline number.
  • Position: when the model lists several businesses, are you first or fifth?
  • Share of voice: how often you are named compared with each competitor on the same prompts.
  • Description: what the model says about you, and whether it is true.
  • Sources: which pages the model read and cited before answering.

Mention rate tells you whether there is a problem. Sources usually tell you how to fix it.

How LLMs decide which businesses to name

A model answers from two places: what it learned in training, and what it reads in a live web search. For questions about local services, prices or "best X near me", the assistants run a search and build the answer from the pages they find. That is why visibility changes week to week and why fixing the right pages works faster than waiting for a new model.

Each assistant searches differently:

AssistantWhere live answers come fromWhat that means for you
GeminiGoogle SearchYour Google Business Profile, reviews and pages that rank in Google weigh heavily.
ChatGPTOpenAI's search, built on third-party search providers and its own crawler (OAI-SearchBot)Bing indexing and being on widely cited lists and directories matter.
PerplexityIts own index and crawler (PerplexityBot), with citations shown for every claimClear, quotable pages and recent coverage get cited often.
ClaudeAnthropic's web search tool and its crawler (Claude-SearchBot)It tends to cite fewer, more authoritative pages.

Two consequences follow. First, a business can be visible in Gemini and missing in ChatGPT, so one model is not the whole picture. Second, the cited pages are often not your own site. For local businesses they are usually directories, review sites, maps and "best of" articles. If a page that lists your competitors is cited in half the answers and you are not on it, that page is your first task.

The formulas

Keep the definitions simple and fixed, so the numbers mean the same thing every month.

MetricFormulaExample
Mention rateanswers that name you ÷ all answers18 ÷ 100 = 18%
Average positionaverage place in lists, only for answers that name you2.6
Share of voiceyour mentions ÷ mentions of you and the tracked competitors18 ÷ (18 + 31 + 24) = 25%
Citation rateanswers that link to your site ÷ all answers6 ÷ 100 = 6%
Accuracyanswers with correct facts about you ÷ answers that name you15 ÷ 18 = 83%

Leave the score at that. Many tools wrap these numbers into a single "visibility score" with their own weighting. That is fine inside one tool, but scores from different tools cannot be compared.

How many answers you need

LLM answers are not deterministic. Ask the same question twice and you may get a different list of names. In a January 2026 study by SparkToro and Gumshoe, ChatGPT and Google's AI returned the same list of brands for repeated prompts less than 1 percent of the time. So one answer proves nothing, and a small sample swings a lot.

The margin of error for a mention rate depends on how many answers you have. For a true mention rate around 30 percent, at 95 percent confidence:

Answers collectedMargin of error
20±20 points
50±13 points
100±9 points
300±5 points

In practice: with 20 answers, a jump from 25 to 40 percent can be noise. With 300 answers a month, which is 10 prompts a day, a change of 6 to 8 points is real. When you compare two months, compare the same prompts, the same models and the same location.

How LLM visibility data is collected: API vs app

There are two ways to collect answers, and it is worth knowing which one a tool uses.

Official APIs with web search. The tool sends each prompt to the model's API with search turned on and a location set. It is stable, reproducible and allowed by the providers' terms. The answer matches what a signed-out user gets from the same search, but it does not include the personal memory or chat history a signed-in app user might have.

Scraping the chat apps. The tool drives a browser through the ChatGPT or Gemini interface. It can capture app-only features, but it breaks when the interface changes, it is hard to reproduce, and it often runs against the terms of service.

Neither method sees exactly what one particular customer sees, because nobody can. That is the reason to measure many answers over time rather than chase a single screenshot. AmIPicked uses the official APIs with web search and the location of your market.

How to track LLM visibility

You can do it by hand or with a tool. The manual method takes a spreadsheet and two to four hours a month. It is described step by step in how to check AI visibility. The short version:

  1. Write 10 to 20 prompts a real buyer would type, including your city or area if you serve one.
  2. Run each prompt three times in at least two models with web search on.
  3. Record mention, position, competitors named, the description and cited URLs for every answer.
  4. Calculate the five metrics above and save the cited URLs in a separate tab.
  5. Repeat with the same prompts every month.

An AI visibility tracker does the same work every day. AmIPicked asks each prompt once a day and rotates it through ChatGPT, Gemini, Claude and Perplexity, so each model sees every prompt every few days and the monthly sample stays large without paying for four models daily.

Can you see LLM visibility in Google Analytics?

Only partly. When someone clicks a link in an AI answer, the visit shows up as a referral from chatgpt.com, perplexity.ai, gemini.google.com, claude.ai or copilot.microsoft.com. ChatGPT also adds utm_source=chatgpt.com to many links. Create a channel group for these domains in GA4 so the visits are not lost in "referral".

But most AI answers are read without a click. A customer who hears your name from ChatGPT and then searches for you on Google shows up as branded search or direct traffic. Watch branded searches in Google Search Console alongside your mention rate. If both rise together, the AI answers are working, even when the AI referral traffic stays small.

How to improve LLM visibility

The steps that move mention rate for most small businesses, in order:

  1. Fix access. Check robots.txt does not block OAI-SearchBot, PerplexityBot, Claude-SearchBot or Googlebot. The AI crawler checker does this in seconds.
  2. Get onto the cited pages. Take the URLs the models cite most for your prompts and get listed, reviewed or mentioned there.
  3. Make the facts identical everywhere. Same name, address, services, prices and hours on your site, Google Business Profile and directories. Models repeat contradictions.
  4. Answer buyer questions on your site. Pages with direct answers, prices and an FAQ are easier to quote than general marketing copy.
  5. Add structured data. LocalBusiness or Organization schema with your services and area helps models connect your name to your category.
  6. Check the description. If a model says something wrong, find the page it came from and correct it there.

More detail on each step is in how to rank in ChatGPT.

A quick example

An accountant in Bristol tracks 10 prompts a day for a month: 300 answers across four models. She is named in 54 (an 18 percent mention rate, ±4 points), mostly in Gemini. In ChatGPT she appears in 6 percent of answers. The most cited page for her prompts is a local "best accountants" list that includes three competitors and not her. Claude describes her as "focused on limited companies", which she stopped doing two years ago.

Her plan for next month: ask to be added to the list, update the old directory profile Claude is quoting, and publish a page about accounting for freelancers with prices. The same 300 answers next month will show what worked. To see your own starting point, run the free AI visibility check.

Frequently asked questions

What is LLM visibility?

+

LLM visibility is the share of answers from large language models, such as ChatGPT, Gemini, Claude and Perplexity, that mention your business or brand when people ask questions about your category. It also covers where you appear in lists, how you are described and which sources the model cites. It means the same as AI visibility or AI search visibility.

How do you measure LLM visibility?

+

Write a fixed set of customer prompts, run each one several times in each model with web search on, and record whether you are mentioned, your position, the competitors named and the cited URLs. Mention rate is answers that name you divided by all answers. Track the same prompts over time so the numbers are comparable.

How many prompts do I need to track LLM visibility?

+

For a rough monthly reading, 10 to 20 prompts run three times in two models gives 60 to 120 answers, which puts the margin of error at about plus or minus 8 to 12 percentage points. With fewer than 30 answers a change of 15 points can be pure chance.

Is LLM visibility the same as SEO?

+

No, but they overlap. With web search on, models read pages from search indexes, so pages that rank and are easy to quote help. The difference is that the model writes one answer with a few names, often from third-party pages such as directories, reviews and forums, so being mentioned on the pages the model reads matters as much as your own rankings.

Do API answers match what users see in ChatGPT?

+

Mostly, but not exactly. The API with web search returns an answer like a signed-out user would get. The ChatGPT and Gemini apps can add memory, chat history and location details from the account. That is why trackers measure the share of answers across many runs rather than trusting a single response.

How can I improve my LLM visibility?

+

Get listed and reviewed on the sites the models cite for your category, keep your name, services, prices and area identical everywhere, publish pages that answer buyer questions directly, add structured data, and make sure robots.txt does not block OAI-SearchBot, PerplexityBot, ClaudeBot or Google's crawlers.

Read next