How to Check Whether AI Recommends Your Company

“Does ChatGPT recommend my company when customers ask about services in my industry?” Answering that question takes a measurement, and a measurement is easy to inflate with a flawed method. The result depends on the conditions: the same questions asked in one conversation or from a logged-in account give a very different picture than the same questions asked by a stranger in a new chat.

Applies to ChatGPT Claude Gemini Copilot Perplexity Grok

Four results that are easy to confuse

“Visibility in AI” covers four different results, each worth something different to a company:

1. A bot fetching the page. It shows up in the server logs and only means the content can be read.
2. A brand mention. The assistant names the company, although it is not yet clear whether it describes it correctly.
3. A domain citation. The assistant lists the company’s website as a source for its answer.
4. A recommendation. The assistant names the company as the answer to a buying question.

Some bot visits only touch the robots.txt file and the sitemap without fetching any content. Not every visit signed with the name of an AI bot comes from that bot either. Any program can choose the name it presents to a server, so logs regularly show ordinary scans for password files signed with the name of an OpenAI bot.

Which questions to ask

Questions should read the way a customer types them before buying: a full sentence with a place, a budget or a problem. Search-style keywords such as “CRM small business” tell you little about an assistant. A good set combines four kinds of questions.

About two thirds of the questions should not contain the company’s name. The rest include it and show what the assistant says about the company when someone asks about it directly.

Measuring visibility step by step

The measurement means asking several AI assistants the same questions and recording whether and how they mention the company. An AI assistant is a program you talk to in a chat window: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Copilot (Microsoft), Perplexity or Grok (xAI). Each one answers with a different model and draws on different search engines and sources, so a company that appears in one assistant’s answers may not exist for another.

The times given for each step are an assumption for one company and three assistants.

1. A list of 30 questions (about 1 h). Following the table above: 20 questions without the company’s name and 10 with it. Keep the questions in a spreadsheet, one per row, so later runs ask them word for word the same way.
2. Choosing the assistants (a few minutes). At least three, ideally the ones the company’s customers use. For business audiences, ChatGPT, Gemini and Copilot cover three different providers.
3. The run (about 5 h). Ask every question in a new chat window, ideally in a temporary or private mode if the assistant has one, so the answer does not draw on earlier conversations. Ask every question twice, because the same assistant answers the same question a little differently each time. With three assistants that gives 180 answers.
4. A note for every answer (during the run). Record in the spreadsheet whether the company was named, in which position, whether the assistant searched the web (visible as source links under the answer), which sources it gave and which companies it named instead.
5. Summarising the result (about 1 h). Count how many answers mentioned the company, how many cited its website as a source and how many recommended it, separately for questions with and without the company’s name. The result for questions without the name is the baseline for later measurements.
6. A repeat after 6 to 8 weeks (as long as the first run). The same spreadsheet of questions, the same assistants and the same note-taking rules. Only the difference between two measurements shows whether changes to the website and the company’s profiles had an effect.

Errors that inflate the result

All questions in one conversation. When general questions follow a few questions about a specific company, the assistant remembers the context and readily recommends it. A customer opening a new chat does not have that context.

A logged-in account with history. ChatGPT, Claude and Gemini can use memory and the user’s earlier conversations. A business owner who often asks about their company will see it in answers more often than a stranger would.

An instruction customers never write. Adding “name specific companies” or “list five companies” pushes the assistant to name companies even where it would not do so on its own. When measuring through an API, the same applies to the system prompt.

A closed list of competitors. A list of companies to track drawn up before the measurement shows expectations about the market. Assistants often name companies outside such a list, so the list should grow with the answers.

Answers without search counted as answers with search. Some answers in web-enabled mode are produced without any search, from the model’s memory alone. Those answers need to be marked separately.

Errors that appear in the answers

Questions with the company’s name, 10 out of 30 in the set, reveal a different kind of problem. The most common ones look like this:

• the assistant quotes prices from an old, forgotten company profile on another website instead of the current price list on the company’s site,
• the assistant claims the company does not publish prices or run a blog, although both are on the website,
• the assistant gives an old address that survived in a business directory or an outdated map listing,
• the assistant confuses the company with another one with a similar name, often from a completely different industry,
• the assistant cites a review site or an aggregator instead of the company’s website.

What changes the result and what does not

The list of things a company can influence is short:

• Consistent data across the web: the same name, address, scope of services and prices on the website, in the Google Business Profile, in industry directories and on old profiles. Unused profiles are better deleted than left behind.
• Prices as text: an assistant will not quote a price placed in an image or a downloadable PDF.
• Content available without JavaScript: some bots do not run scripts. The test takes a minute: switch off JavaScript in the browser and check what is left on the page.
• Structured data that matches the content: schema.org markup describes the company, its services and prices directly, and Google notes that it should match the text visible on the page.¹

The llms.txt file is not on the list. Google states plainly that appearing in AI Overviews and AI Mode does not require new machine-readable files or special markup.¹ In May 2026 Ahrefs looked at traffic across 137,000 domains.²

• 28% of domains with a valid llms.txt file
• 97% of those files not fetched even once in May 2026
• ~1 in 5 fetches of the file made by AI bots, the rest by SEO tools and other bots

What to do with the result

The result points to the direction of further work:

The measurement and a plan of fixes can also be commissioned. We run this review as part of the AI Visibility Scan.

Frequently asked questions