How to check if ChatGPT recommends your brand

Asking ChatGPT about your own market is now a normal part of competitive research. It is also easy to do badly. Most founders type one question, see their brand (or don't), take a screenshot and move on. That screenshot is a sample of one from a system that answers differently every time. This guide shows a method that gives you an answer you can act on, whether you do it by hand in a spreadsheet or use a tool.

The stakes are real. OpenAI CEO Sam Altman said in October 2025 that ChatGPT had reached 800 million weekly active users. Some of those people are asking for recommendations in your category.

Why one ChatGPT answer is not a result

AI assistants generate each answer fresh. Ask the same question twice and you can get different brands, in a different order. SparkToro and Gumshoe measured this directly: 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI a combined 2,961 times. They found there is less than a 1 in 100 chance that ChatGPT or Google's AI will give the same list of brands in any two of 100 responses, and roughly 1 in 1,000 before two lists appear in the same order.

The same study came to a useful conclusion: precise rankings are noise, but a brand's visibility percentage across dozens to hundreds of prompts, run multiple times, is a reasonable metric. That is the approach below. You stop asking "did ChatGPT name us?" and start asking "in what share of answers does ChatGPT name us?"

Step 1: Write the questions your buyers actually ask

Your prompt list decides everything that follows, so spend the most time here. Aim for 10 to 30 prompts to start. Write them the way a buyer would type them, not the way your marketing team would phrase them.

• Category prompts: "What is the best project management tool for small agencies?", "Top payroll software for US startups".
• Use-case and constraint prompts: "Invoicing tool for freelancers that supports multiple currencies", "CRM with a good free plan for a two-person team".
• Problem prompts: "How do I stop losing leads from my website form?" Buyers often describe the pain, not the category.
• Comparison prompts: "Alternatives to [market leader]", "[Competitor A] vs [Competitor B]". These are unbranded for you, and they show whether you are in the consideration set.

Keep branded prompts such as "Is [your brand] any good?" in a separate group. They are useful for checking how ChatGPT describes you, but they will mention you almost every time and should not count toward discovery.

Write several phrasings of each important question. In a follow-up analysis, SparkToro found that when hundreds of people wrote prompts for the same intent, the average semantic similarity was only 0.081: people almost never phrase the same need the same way. Testing one "perfect" prompt tells you little about the range of questions real buyers type.

Step 2: Set up a fair test

You want to see roughly what a stranger sees, and you want this month's results to be comparable with next month's. A few rules help:

• Start a new conversation for every prompt so earlier questions do not steer the answer.
• Be aware that your own account history and settings can shape answers. If you can, test from an account that has never talked about your company, and write down which settings were on.
• Note whether the answer used web search (it shows linked sources) or answered from the model alone. Treat the two as different tests.
• Run the full set on the same day, and keep the prompt wording fixed between checks. Changing prompts changes the baseline.

Step 3: Run each prompt more than once

Because answers vary, run each prompt at least three times, ideally five. Twenty prompts run five times gives you 100 answers, which is enough to see a pattern. If you also check Perplexity and Gemini, the count multiplies by the number of engines, which is the point where most teams stop doing this by hand.

Step 4: Record the same fields for every answer

The description column is easy to skip and often the most useful. Being mentioned as "an enterprise tool" when you sell to small teams, or with a price you changed a year ago, is a problem you can fix at the source.

Step 5: Turn the answers into numbers

Three numbers summarise the spreadsheet. Calculate them for unbranded prompts only.

1. Visibility: answers that mention you, divided by total answers, times 100. Mentioned in 23 of 100 answers means 23% visibility. Our guide to the AI visibility score explains how to read it.
2. Average position: the mean of your position, counting only answers that mention you. Treat it as a rough signal; the SparkToro study found ordering is the least stable part of an answer.
3. Share of voice: your mentions as a share of all brand mentions across the answers. If competitors were named 300 times in total and you 30 times, your share is 30 out of 330, about 9%. See how to track share of voice in AI answers.

The numbers above are a worked example, not data from a real brand. Your result depends entirely on your category and prompt list.

Step 6: Check that ChatGPT can read your site

If your visibility is low, rule out the simplest cause first. OpenAI's documentation says OAI-SearchBot is used to surface websites in ChatGPT's search features, and that sites opted out of it will not be shown in ChatGPT search answers, though they can still appear as navigational links. GPTBot is a separate crawler for model training, and blocking one does not block the other.

• Open yoursite.com/robots.txt and look for rules that block OAI-SearchBot, or block all bots.
• Check your CDN or firewall bot settings. Bot protection can block crawlers even when robots.txt allows them.
• After a change, wait. OpenAI says it can take about 24 hours from a robots.txt update for its systems to adjust.

Check the other AI engines the same way

Your buyers do not only use ChatGPT, and each engine pulls from different sources, so results often differ a lot between them. Use the same prompt list for each:

• Perplexity says it searches the internet in real time and that each answer includes numbered citations linking to the original sources, which makes the sources column easy to fill.
• Google's AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches across subtopics to build one response, so supporting links can come from pages that rank for related questions, not only the one you typed.
• Gemini apps can use grounding, which Google describes as providing content from the Google Search index to the model at prompt time. Note whether an answer shows sources, as you did for ChatGPT.

What to do with the result

Sort your prompts into three groups: ones where you are usually named, ones where you are sometimes named, and ones where you never appear. The last group is your work list. For each, look at the cited sources and the competitors named instead. Those sources are where you need to be present and accurately described. Our post on why ChatGPT doesn't mention your brand walks through the usual causes and fixes.

Then repeat the check on a schedule, weekly or monthly, with the same prompts. A move of a few points between two runs is normal variation. A trend over several runs is a real change.

When to stop doing it by hand

A manual check is worth doing once, because you learn how ChatGPT talks about your market. Repeating it is where it gets expensive: 20 prompts, five runs and three engines is 300 answers to read and tag every time. That is the job AI visibility trackers do. If you want to try it first, the free AI visibility check runs five buyer prompts for your brand and shows who is named, who is named instead and which sources are cited. If you are comparing paid options, see AI search visibility tools compared.

Frequently asked questions

Can I just ask ChatGPT "do you recommend my brand?"

You can, but it measures the wrong thing. A prompt with your name in it tells you how ChatGPT describes you when asked directly. It does not tell you whether ChatGPT brings you up when a buyer asks for options. Use unbranded prompts for that.

How many prompts are enough?

Start with 10 to 30 that cover your main use cases, each run several times. More prompts and more runs give a steadier number; the SparkToro study points to dozens to hundreds of prompts for a statistically meaningful visibility figure.

Why do I get different results from a colleague?

Answers are generated fresh each time, and account history, settings and whether web search ran can all differ between two people. That is why a fixed prompt list, repeated runs and a clean setup matter more than any single answer.