Ricardo Batista
September 19, 2026
9 min read
Software buyers increasingly ask ChatGPT, Perplexity or Google’s AI Mode for a recommendation before they open a search box. The answer names two or three products, and those are the ones a buyer looks at next. If your brand is not in that answer, you are not in the comparison.
The hard part is that none of these engines tells you how often it recommends you. There is no Search Console for ChatGPT. So teams do the obvious thing: they type their category into ChatGPT, see whether their name comes up, and treat the result as a finding. That single check is close to worthless. Below is why, followed by a setup that produces numbers you can act on.
Don’t miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Why checking ChatGPT once tells you almost nothing
AI answers change from one run to the next. Ask the same engine the same question twice and you will often get a different set of recommended brands and a different set of cited
We measured how large that effect is. In paired runs, the same model answering the same prompt minutes apart repeated only about 40% of its cited domains. More than half of what it cited the first time was gone the second time, with nothing changed on anyone’s website.
That changes what a single answer can tell you. If ChatGPT names you once, you have learned that you are in the pool of brands it can draw from. How often it draws you is a separate question. If it leaves you out once, you have learned even less. A real answer to “how visible are we” is a rate, measured over many prompts and many runs, and that is what the rest of this guide builds.
What AI brand tracking actually measures
Four numbers do most of the work. Keep them separate, because they move independently.
- Mention rate. The share of answers that name your brand at all. This is the headline number most tools report.
- Citation rate. The share of answers that link to your own website as a source. An engine can recommend you without citing you, and it can cite your blog while recommending a competitor.
- Position. Where your name appears when it does appear. Being the first of three recommendations is worth more than being the sixth item in a list.
- Share of voice. Your mentions divided by the mentions of every brand in your competitor set. This is the number that tells you whether you are gaining ground, because it controls for answers simply getting longer or shorter.
A fifth input is for diagnosis: which third-party pages the engines cite when they answer your category’s questions. Those pages are where the engines get their opinions, and Step 6 covers what to do with them.
Prerequisites
Before you start, have these ready:
- A list of three to eight competitors. Share of voice needs a denominator.
- A spreadsheet or database that stores every answer in full, because you will want to re-read old answers when a number moves.
- A cadence. This guide uses daily collection and weekly reporting.
- A budget ceiling for either a dashboard subscription or API credits, depending on the method you pick in Step 3.
Step 1: Write your prompts from real buyer questions
Your prompt set is the sample every number is measured on, so it deserves more care than any other step.
Write prompts the way a buyer would ask them. “What is the best tool to monitor our brand in AI answers” is a real question. “AI-powered brand intelligence platform” is not something anyone types.
- Questions prospects ask on sales calls.
- The “People also ask” questions on your main search terms.
- Threads in the communities your buyers read, where people ask each other for recommendations.
Aim for coverage of distinct intents. Ten phrasings of the same question are one intent sampled ten times. Most teams land somewhere between 50 and a few hundred prompts, and the dashboard market prices around the same range: Peec’s plans cap tracked prompts at 50, 150 and 350.
Tag each prompt with the intent it represents (comparison, “best tool for”, pricing, alternatives) so you can later see which kind of question you win and which you lose.
Step 2: Choose the engines
Track the engines your buyers actually use, and track them separately. Each engine runs its own retrieval, and in our own monitoring a brand that is well represented in one engine can be nearly absent from another. A single blended “AI visibility” score hides exactly the gaps you need to find.
For most B2B and software categories, the set that matters is:
- ChatGPT.
- Perplexity, which shows its sources alongside every answer.
- Google AI Overviews and AI Mode, which sit on top of Google Search.
- Gemini and Microsoft Copilot, if your buyers live in those ecosystems.
Check the answers with web search switched on. In our monitoring, almost every buying question triggers a live search, and that search decides which
Step 3: Pick a collection method
There are three ways to collect the answers. They trade cost against control.
| Method | Good for | Watch out for |
|---|---|---|
| Manual checks in the chat window | A first look at a handful of prompts | Personalization, no history, and far too few answers to measure a rate |
| A visibility dashboard (Profound, Peec, Otterly, Semrush’s AI toolkit, Ahrefs Brand Radar) | Marketing teams who want charts and alerts without engineering work | Prompt caps per plan, fixed engine lists, and limited export of the raw answers |
| An API that returns the raw answers | Teams who want every answer in their own database and their own metrics | You write the scoring logic yourself |
Manual checks are fine for curiosity and useless for measurement. Dashboards are the fastest way to a chart. The API route suits teams who already run a data warehouse and want the underlying answers themselves.
On the API side, cloro returns the answer from ChatGPT, Perplexity, Copilot, Gemini, Google AI Overviews and AI Mode as structured JSON, including the list ofth, and the Lite plan is $30 a month for 37,500 credits, which is enough to prototype a tracker before committing to anything
Step 4: Run every prompt more than once
Because answers vary from run to run, one run per prompt per week is too little data. Collect every prompt daily, on every engine you track, and keep all of it.
The numbers add up quickly. Fifty prompts across six engines sampled once a day is 2,100 answers a week. In our analysis, around 300 answers per period is enough to resolve a visibility rate to roughly plus or minus five percentage points. Below that, the week-to-week movement you see is mostly noise.
Two rules keep the data clean:
- Hold the prompt wording fixed. Changing a prompt changes the sample. If you need a new phrasing, add it as a new prompt rather than editing the old one.
- Hold the location fixed. Answers differ by country, so run each prompt set from the market you care about.
Step 5: Record mentions, citations and position
For each answer, store the raw text and the list of cited:
- prompt, engine, date and country
- whether each brand in your competitor set is named
- the order in which the brands appear
- whether your own domain is among the cited sources
- the full list of cited URLs
Match brand names carefully. Count “Acme” and “Acme AI” as the same brand, and make sure a short brand name does not match inside other words. In our experience, matching is where early tracking setups go wrong.
Step 6: Read the
This step turns the numbers into a list of things to change.
When an engine answers a question about your category, it cites pages: review sites, comparison articles, forum threads, vendor blogs. Aggregate thoseages that shape what the engines say about your market
Then sort that list into three groups:
- Pages that already name you. Keep the information on them accurate.
- Pages that name your competitors but not you. These are your best opportunities. Getting added to a page that engines already cite changes the answers more directly than publishing a new page they have never seen.
- Your own pages. Check which of them get cited, and for which prompts.
Step 7: Report weekly, not daily
A daily chart of mention rate mostly shows the run-to-run randomness described at the start, and it will send your team chasing noise. Weekly totals smooth that out.
A weekly report needs only a few lines:
- mention rate and share of voice per engine, against last week and against a four-week average
- the prompts where you gained or lost the most
- new third-party pages that started appearing in citations
When a change matters, check it against the sample size before acting on it. A move from 20% to 30% sounds large, but on a few dozen answers it is well within normal variation.
What the output actually looks like
For teams taking the API route, here is the shape of one call and its response. The request sends a single prompt to ChatGPT with web search on:
curl -X POST "https://api.cloro.dev/v1/monitor/chatgpt"
-H "Authorization: Bearer YOUR_API_KEY"
-H "Content-Type: application/json"
-d '{"prompt": "What is the best tool to track brand mentions in AI answers?", "country": "US"}'{
"success": true,
"result": {
"text": "Here are some of the most popular options...",
"sources": [
{ "position": 1, "url": "https://example.com/best-ai-visibility-tools", "label": "The best AI visibility tools" },
{ "position": 2, "url": "https://example.org/ai-brand-monitoring-guide", "label": "AI brand monitoring guide" }
]
}
}result.text is what you score for mentions and position. result.ude.searchQueries also returns the search queries ChatGPT ran behind the scenes to build the answer, which shows you the exact searches your content needs to rank for
Common mistakes that make AI brand tracking misleading
- Reading one answer as a result. One answer is one draw from a pool that changes every run.
- Blending engines into one score. It hides the engine where you are weakest.
- Editing prompts mid-stream. Every edit breaks the comparison with past weeks.
- Tracking mentions and ignoring citations. Mentions show where you stand, and the cited sources show what put you there.
- Tracking only your own brand. Without competitors there is no share of voice, and no way to tell a market-wide shift from a change that affects only you.
- Checking from a logged-in personal account. Personal history and memory shape the answer. Use clean, repeatable sessions.
How do I track brand mentions in ChatGPT?
ChatGPT itself has no analytics view for brands, so you build the measurement yourself:
- Write 50 or more prompts from real buyer questions.
- Pick the engines your buyers use and track each one separately.
- Collect the answers through a dashboard or an API.
- Run every prompt daily, with fixed wording and a fixed country.
- Record mentions, citations and position for your brand and each competitor.
- Aggregate the sources the engines cite to find the pages that shape the answers.
- Report weekly, and check each change against the number of answers behind it.
How many prompts do I need?
Enough distinct questions to cover the ways buyers ask about your category, usually between 50 and a few hundred. Coverage of different intents matters more than the raw count.
How often should I check?
Collect answers every day and look at the numbers once a week. Daily readings swing too much to act on, and weekly totals give you enough answers to see a real change.
Why does ChatGPT give a different answer every time?
Each answer is generated fresh, and with web search on, the engine retrieves a new set ofains repeated when the same prompt was asked minutes apart
What should I do if competitors are cited and I am not?
Look at which pages the engines cite for those questions. Getting accurate information about your product onto the pages they already trust is usually faster than waiting for your own content to be picked up.
Related reading
- GPT Store vs Claude Skills vs Gemini Gems: 159K Gap [2026]
- Grok Imagine API Setup: 13 Steps, $0.032/Image [2026]
- Muse Spark Tops GPT-6 Astra, Claude Fable: 75% SWE [2026]
