How to track whether ChatGPT mentions your company, for free

By Roshan, founder of SearchMaxx - Sep 24, 2026 - updated Oct 18, 2026 - 8 min read

Write 20 to 30 questions your buyers ask, run each one several times in ChatGPT, Claude, Perplexity and Google, and log whether each answer names you, links to you, and where you appear. Report the share of answers that mention you, together with how many answers that share is based on. Repeat on the same day each month with the same questions, so the trend means something.

That is the whole method. The details below make the numbers worth trusting.

Why you have to ask more than once

AI answers change from one run to the next, even with the same question. In a SparkToro study, about 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI 2,961 times. ChatGPT and Google's AI gave the same list of brands less than once in 100 runs, and the same list in the same order about once in 1,000 (SparkToro).

Two lessons follow from that study:

• Track how often you appear, not your position. The author wrote that "visibility % across dozens to hundreds of prompts run multiple times is a reasonable metric", and that any tool giving a "ranking position in AI" is "full of baloney".
• One screenshot proves nothing. If a colleague sees you named once, or not named once, that is a single draw from a range of answers.

The study suggests "usually at least 60-100X" runs per prompt for a precise number. A free, manual method cannot do that. It can still show you whether you are named often, rarely or never, and whether that changes over months. Treat it as a rough gauge, and say so when you share it.

Step 1: Build your prompt set

Write 20 to 30 questions in the words a buyer would use. Spread them across these types, using Paidwick, a made-up invoicing app for freelancers, as the example:

Where to find real wording:

• Sales call notes and demo request forms. Copy the buyer's exact phrasing.
• Support tickets that start with "Can I..." or "How do I...".
• Your Google Search Console queries, especially longer ones phrased as questions.
• Forum threads where people ask for recommendations in your category.

Keep brand-check questions separate when you report. They tell you whether the engines describe you correctly. Only the other questions tell you whether buyers who have never heard of you will find you.

Once the list is set, freeze it. Changing questions between runs breaks the trend. If you must add one, add it as new and keep the old ones.

Step 2: Set up clean sessions

Your own account is the worst place to test. It knows who you are.

• ChatGPT memory. With "Reference chat history" on, ChatGPT draws on past conversations, and Temporary Chats do not use existing memories or create new ones (OpenAI Memory FAQ). Use a Temporary Chat or a fresh account with memory off. If you test from a work account where you have asked about your own product a hundred times, you will see yourself more than your buyers do.
• Other assistants. Check each one's settings for memory or personalization and turn it off, or use a fresh account.
• Google. Use a private window, signed out.
• Location. Results can depend on where you are. OpenAI's and Anthropic's web search APIs both take a user_location setting to localize results (OpenAI, Anthropic). Run from the same place every time, and write down where that is. If you sell in several countries, treat each country as its own test.

Record the model or mode you used, such as the model name ChatGPT shows, or whether search was on. When a model changes, your numbers may jump for reasons that have nothing to do with your site.

Step 3: Run each prompt several times

Start a new chat for every run. Paste the question exactly as written. Do not add follow-up questions.

A workable free setup:

• 25 prompts
• 3 runs per prompt per engine
• 4 engines: ChatGPT, Claude, Perplexity, Google AI Overviews or AI Mode

That is 300 answers. Split it across a few people, or across a few days in the same week. If you only have time for one engine, start with the one your buyers use most, and run 5 times per prompt instead of 3.

Step 4: Log every answer

Use one row per answer. Here is a column layout you can paste into a spreadsheet as the header row:

date,run_by,location,engine,model_or_mode,prompt_id,prompt_type,run,mentioned,linked,position,competitors_named,accurate,notes

Keep position even though it is noisy. It is useful as a rough check over many runs. Just do not report it as a ranking.

The accurate column can point to quick fixes. If engines keep saying you lack a feature you have, a clear page that states it plainly is your next task.

Step 5: Compute rates, with sample sizes

The main number is the mention rate: answers that name you divided by all answers. Compute it per engine, and per prompt type.

If your columns sit in a sheet called Log, with engine in column D and mentioned in column I:

=COUNTIFS(Log!D:D,"ChatGPT",Log!I:I,"Y") / COUNTIFS(Log!D:D,"ChatGPT")

Do the same for linked to get a citation rate. Always show the count next to the rate:

How much movement is noise? With 75 answers and a mention rate near 20%, a standard 95% margin of error is about plus or minus 9 points. With 150 answers it is about plus or minus 6. Runs of the same prompt are not fully independent, so the true uncertainty is larger than that. A move from 16% to 21% in one month is not news. A move from 5% to 30% over three months is.

How often to run it

• Monthly for the full set. Same week, same people, same location.
• After a big change, such as a new comparison page or a pricing update, wait a few weeks, then run only the prompts that change should affect.
• Quarterly, review the prompt list. Add new questions as new rows. Do not edit old ones.

Pitfalls

• Testing while logged in to your own account. Memory and history make you look more visible than you are.
• Rewording questions. Small wording changes give different answers. Paste the exact text.
• Mixing locations or people's accounts without recording them.
• Counting a warning as a win. "Avoid Paidwick, it has no mobile app" is a mention. Note what was said, not only that you were named.
• Confusing names. If your name is a common word, read each answer rather than searching the text.
• Reporting a rate without its sample size. "40% of answers" from 5 answers means very little.

Optional: automate it with the official APIs

If you can write a short script, the APIs let you run many more repetitions. Answers from an API are not the same as answers in the consumer apps, which have their own settings and features. Log API runs as their own engine value, such as "OpenAI API", and do not mix them with app runs.

• OpenAI. The Responses API has a hosted web_search tool. Cited sources come back as url_citation annotations on the message, and user_location sets an approximate location (docs).
• Anthropic. The Messages API has a web search tool, for example web_search_20250305. Citations are always on and return the cited URL and title. It costs $10 per 1,000 searches plus tokens (docs).
• Perplexity. Perplexity is moving from Sonar Chat Completions to its Agent API. Its docs give September 27, 2026 as the end of Sonar Chat Completions support. Agent API responses include a search_results item with the sources (docs).

An outline in Python for the first two:

from openai import OpenAI
import anthropic

openai_client = OpenAI()
claude_client = anthropic.Anthropic()

def ask_openai(prompt):
    r = openai_client.responses.create(
        model="MODEL_NAME",
        tools=[{"type": "web_search"}],
        input=prompt,
    )
    return r.output_text

def ask_claude(prompt):
    m = claude_client.messages.create(
        model="MODEL_NAME",
        max_tokens=2048,
        tools=[{"type": "web_search_20250305", "name": "web_search"}],
        messages=[{"role": "user", "content": prompt}],
    )
    return "".join(b.text for b in m.content if b.type == "text")

# For each prompt, call each function N times, check the text for
# your brand name, and append one row per answer to a CSV with the
# columns above.

Check your brand name with a case-insensitive match, then read a sample of the answers yourself. A script will miss misspellings and will count warnings as mentions.

Next step

If you are rarely named, check the basics first. Our free check shows whether the AI search crawlers can read your site. Then look at the questions where competitors appear and you do not. Those are the questions where a new page can help, and our guide to comparison pages AI assistants cite shows how to write it.

Want this done for your site?

Every day we make one change to your website that helps AI engines name you, and a person checks each one. Email me your website and the question your buyers ask AI.

Email Roshan to get started

Or check your site free: