How to track brand mentions in ChatGPT, Gemini and Claude

AI answers aren’t deterministic. In a SparkToro and Gumshoe study, hundreds of volunteers ran 12 prompts through ChatGPT, Claude and Google’s AI, collecting 2,961 responses. ChatGPT and Google’s AI returned the same list of brands less than 1 time in 100. What held up was frequency: across 994 responses to headphone prompts, the leading brands appeared in 55–77% of them.

So tracking means sampling. Ask each question several times, count how often you’re named (your visibility), and treat position as secondary.

1. 01

   Build a question set

   10–30 questions across these shapes:

   • Category: “best meditation app”
   • Use case: “meditation app for anxiety before sleep”
   • Alternatives: “alternative to Headspace”
   • Comparison: “Calm vs Headspace”

   Keep branded questions (“is Still a good app?”) in a separate group. They tell you what the assistant says about you, but they inflate visibility if mixed in.

2. 02

   Remove personalization

   Memory and custom instructions change answers. In ChatGPT, use a temporary chat without personalization or turn memory off; in Gemini and Claude, use a clean session without memory. Note which you used so runs stay comparable.

3. 03

   Run each question three times per assistant

   Start a new chat for every run. Five runs per question if you’re about to make a decision on the result.

4. 04

   Log the same fields every time

   Date, assistant, question, run, named (yes/no), position, competitors named, cited URLs. See the sample below.

5. 05

   Compute the rate and repeat weekly

   Visibility = runs where you’re named ÷ total runs. Re-run the exact same set on the same weekday. Change questions only deliberately, and start a new baseline when you do.

By hand works for a baseline. For weekly tracking, script it. All three providers expose web search through their APIs:

• OpenAI Responses API with the web_search tool. You can pass an approximate user location including a two-letter country, and cited sources come back as url_citation annotations with URL and title.
• Gemini API with the google_search tool. The response’s groundingMetadata lists the web sources (groundingChunks) and which parts of the answer they support.
• Claude API (Messages) with the web_search server tool. It accepts a user_location with a two-letter country, and citations are always on: cited passages come back as web_search_result_location entries with URL and title.

curl https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "<any model with web search>",
    "tools": [{ "type": "web_search",
                "user_location": { "type": "approximate", "country": "US" } }],
    "input": "What is the best meditation app for sleep?"
  }'

The hard parts aren’t the calls:

• Matching names. String matching fails on generic names (“Notes”, “Focus”) and misses aliases. Extract the list of recommended products from each answer first, then match against your name and aliases.
• Volume. 20 questions × 3 assistants × 3 runs is 180 calls per check-in, each with search fees on top of tokens. Check current pricing for each provider before scheduling it daily.
• API vs app. API answers won’t match the consumer apps exactly (no memory, possibly different models). That’s fine for trend tracking as long as you’re consistent.

A category of AI visibility trackers now exists. They differ mostly in who they’re built for and how they charge. Plans change often, so check each vendor’s pricing page.

Disclosure: MentionReach is our product. It covers fewer engines than the enterprise tools; the trade is pay-per-check pricing and an app-first setup (paste an App Store, Google Play or web URL). See pricing.

• Low visibility, and the cited pages don’t mention you: work on getting onto those pages. See how to get cited.
• Named, but late in the list: look at how cited pages describe you versus the apps ranked above you.
• Strong on one assistant, absent on another: compare cited domains per assistant. They often differ.
• One competitor everywhere: read the pages that name them. That’s your outreach list. More in how to measure AI visibility.