# How to track AI visibility: a prompt-by-prompt method

> Track AI visibility step by step: build a prompt set, measure brand mentions in ChatGPT and 7 more AI engines, and add Search Console, Bing and GA4 data.

By Infinita team, AI search research · November 19, 2026 · https://goinfinita.com/blog/how-to-track-ai-visibility/

## Key takeaways

- Measure against a fixed prompt set, with branded and non-branded prompts, and keep it stable over time. No fixed set, no trend.
- AI answers change almost every time. In SparkToro’s research, the chance of getting the same list of brands twice was under 1 in 100. Measure in repeated samples, not screenshots.
- Control for personalization: in ChatGPT, test logged out or in a Temporary Chat set to “Unpersonalized.”
- Official data now covers part of the picture: Search Console’s generative AI report, Bing Webmaster Tools AI Performance, ChatGPT’s Business Hub and the AI Assistant channel in GA4. Everything else you measure by asking.
- Report a short, stable set of metrics: mention rate, citation rate, AI share of voice, sentiment and accuracy.

To track AI visibility, you run a fixed set of prompts through AI engines like ChatGPT, Gemini and Google AI Overviews, repeat them enough times to get a stable sample, and measure how often your brand is mentioned, cited and described correctly. Then you add the official data that now exists for some engines. This guide is the hands-on method: the prompt set, the testing conditions, the sample size, a worksheet, the formulas and a reporting rhythm. If you want the bigger picture first, start with [AI visibility: what it is and how to measure it across 8 AI engines](https://goinfinita.com/blog/ai-visibility/).

## What you’re actually measuring

A useful measurement answers three questions for every prompt, engine and month: does your brand **show up**, is your site **cited** as a source, and is what the AI **says about you** right? It does not answer “what’s our rank in ChatGPT?” As step 3 shows, list order changes in almost every answer. For definitions of terms like citation rate, GEO and AEO, see [GEO, AEO and AI visibility: a practical glossary for marketing teams](https://goinfinita.com/blog/ai-visibility-geo-aeo-glossary/).

## Step 1: Build your prompt set

Your prompt set is the list of questions you’ll ask in every cycle. If it doesn’t sound like your buyers, the numbers won’t mean much. Here’s a mix, using “Summit Example,” a **fictional** outdoor gear retailer based in Denver that ships nationwide:

| Prompt type | Example | What it tells you |
|---|---|---|
| Branded | “Is Summit Example legit for buying a tent online?” | What the AI knows about you, and whether it’s accurate |
| Non-branded category | “Best online stores for backpacking gear in the US” | Whether AI picks you when buyers don’t know you yet |
| Comparison | “Summit Example vs. [competitor] for ultralight tents” | How you’re positioned against rivals |
| Problem | “What sleeping bag do I need for camping in Colorado in October?” | Whether you show up before the buyer is shopping |
| Purchase | “Where can I buy a 2-person tent with easy returns?” | Whether you’re recommended at decision time |

Source prompts from real language:

- **Search Console queries.** If your property is eligible, the Branded and Non-branded query filter makes it easy to split them.
- **Sales and support.** The questions customers ask most in chat, email and calls.
- **Autocomplete, Reddit and forums.** How people phrase things before they know what to buy.

For why real prompts are longer and more specific than search queries, see [How your customers actually ask AI (and how to get recommended)](https://goinfinita.com/blog/how-customers-ask-ai/).

**How many prompts?** There’s no official number. We recommend starting with 20 to 50, weighted toward non-branded prompts, plus 3 to 5 competitors to track. Save the set with a date and version it: if you add prompts, add them as a new version instead of quietly editing the old one.

## Step 2: Pick your engines and lock the test conditions

Infinita tracks 8 engines: ChatGPT, Gemini, Google AI (AI Overviews and AI Mode), Perplexity, Claude, Grok, Microsoft Copilot and DeepSeek. If you’re starting by hand, pick the 3 or 4 your buyers use most and add the rest later.

Conditions matter more than most teams expect. Five people testing from five personal accounts aren’t measuring the same thing. Lock these rules:

1. **No personalization.** In ChatGPT, test logged out or open a Temporary Chat and choose “Unpersonalized” before your first message. OpenAI says Unpersonalized doesn’t use memory, custom instructions or plugins, and you can’t switch after the chat starts. A Personalized temporary chat can still use your memories.
2. **A fresh conversation for every prompt.** Earlier turns change later answers.
3. **Your buyers’ country and language.** Test from the market you sell in.
4. **The same mode every time.** If an engine can answer with or without web search, note which one you used and stick with it.
5. **Date and engine on every row.** Models change; without dates you can’t tell your progress from a model update.

## Step 3: Repeat every prompt and measure samples, not screenshots

SparkToro asked 600 volunteers to run 12 prompts through ChatGPT, Claude and Google’s AI, 2,961 times in total. The odds of getting the same list of brands twice were under 1 in 100, and the same list in the same order closer to 1 in 1,000. Their conclusion wasn’t “don’t measure.” It was that visibility percentage across many prompts, run multiple times, is a reasonable metric, and that “ranking position in AI” isn’t.

What that means in practice:

- **Run each prompt several times per engine, every cycle.** To fully map an AI’s recommendations, SparkToro suggests asking at least 60 to 100 times. For monthly tracking, we recommend starting with 3 to 5 runs per prompt per engine, and more for your most important prompts.
- **Calculate over the full sample.** 40 prompts × 3 runs × 3 engines = 360 answers. One or two answers moving isn’t a trend.
- **Compare periods with the same sample.** Same prompt set, same number of runs, same engines.

## Step 4: Log every answer in a worksheet

One row per answer. These columns cover what you need:

| Column | What to record |
|---|---|
| Date | Run date |
| Engine and mode | e.g., ChatGPT logged out, Perplexity, AI Overviews |
| Prompt ID and set version | So you always compare the same thing |
| Type | Branded / non-branded (plus category, comparison, problem, purchase) |
| Run | 1, 2, 3… |
| Mentioned? | Yes / no |
| Position | Where your brand appears in the list, if it does |
| Cited? | Yes / no, plus the cited URL |
| Sources cited | Domains the answer relies on |
| Competitors mentioned | From your list, plus any new ones |
| Sentiment | Positive / neutral / negative |
| Accurate? | Prices, locations, current products |

Copy this header row into a spreadsheet:

```
date,engine,mode,prompt_id,set_version,type,run,mentioned,position,cited,cited_url,sources,competitors,sentiment,accurate
```

## Step 5: Calculate your metrics

Calculate each metric per engine and per prompt type. Never blend branded and non-branded prompts into one number.

| Metric | Formula | Example (Summit Example, hypothetical) |
|---|---|---|
| Mention rate | Answers that mention you ÷ total answers | 24 of 120 non-branded ChatGPT answers = 20% |
| Citation rate | Answers that link to your site ÷ total answers | 6 of 120 = 5% |
| AI share of voice | Your mentions ÷ mentions of all tracked brands | 24 of 160 = 15% |
| Average position | Mean position across answers where you appear | 2.8 (across 24 answers) |
| Sentiment | Positive / neutral / negative ÷ your mentions | 15 positive, 8 neutral, 1 negative |
| Accuracy | Answers with correct facts ÷ answers that mention you | 20 of 24 = 83% |
| Sources cited | Most-cited domains in your category’s answers | Two gear review sites and a hiking forum |

*Illustrative example with a fictional brand and hypothetical numbers; not real data.*

How to read it: Summit Example shows up in 1 in 5 non-branded answers, but its own site is rarely the source, and 4 answers still list a discontinued tent. The sources column points to the next move: if ChatGPT leans on gear review sites and a hiking forum, that’s where the brand needs to be described well.

## How to track brand mentions in ChatGPT

Since November 2026, ChatGPT offers some first-party data. Combine three sources:

1. **Your prompt sampling**, using the conditions in step 2 and several runs per prompt.
2. **Business Hub.** Since November 10, 2026, in beta, businesses can claim their profile in ChatGPT and see aggregated insights: the number of mentions and website clicks they receive, plus discovery topics, the themes of conversations where their website is cited. OpenAI says it doesn’t share users’ conversations, and insights only appear above certain thresholds.
3. **GA4.** OpenAI has documented that ChatGPT adds `utm_source=chatgpt.com` to referral URLs, and GA4 groups visits from AI assistants into the AI Assistant channel.

Business Hub won’t show which competitors appear next to you or what ChatGPT says in a given answer; your sampling will. If ChatGPT rarely names you, work through [Why isn’t my brand showing up in ChatGPT? What to do](https://goinfinita.com/blog/why-my-brand-isnt-showing-up-in-chatgpt/), and to improve, see [How to rank in ChatGPT: a practical guide for brands](https://goinfinita.com/blog/how-to-rank-in-chatgpt/).

## Step 6: Add the official data for each engine

Official data doesn’t replace sampling, but it confirms it with numbers you can’t see from outside. As of November 2026:

| Source | Engine | What it shows | What it doesn’t |
|---|---|---|---|
| Search Console: Generative AI performance report (Search) | AI Overviews and AI Mode | Impressions of links to your site, by page, country, device and date. Rolled out to all sites worldwide as of August 31, 2026 | Queries, clicks, or what the answer said |
| Search Console: Branded / Non-branded query filter | Classic Google Search results (not AI) | Impressions, clicks, CTR and position, split by branded vs. non-branded queries. In the Search Analytics API since November 16, 2026 | AI features; only top-level properties with enough volume |
| Bing Webmaster Tools: AI Performance | Microsoft Copilot, AI summaries in Bing, select partners | Total citations, cited pages, grounding queries and trends. Topics Portfolio (beta, since October 13, 2026) groups queries into topics and lists opportunity queries where other sites are cited and yours isn’t | Full user prompts, or why a page was cited |
| ChatGPT Business Hub (beta) | ChatGPT | Aggregated mentions, website clicks and discovery topics | Individual answers, competitors |
| GA4: AI Assistant channel | Sources like ChatGPT, Gemini, DeepSeek, Copilot and Grok | Sessions from AI assistants (since May 13, 2026) | Visits from AI Overviews and AI Mode, which GA4 counts as Organic Search, and anyone who read the answer without clicking |

Setup notes:

- **Search Console:** if you don’t see the generative AI report, your site usually hasn’t had enough impressions in those features yet. Line up your top pages there with the prompts where Google’s AI cites you in your sampling. To improve, read [How to get cited in Google AI Overviews and AI Mode](https://goinfinita.com/blog/how-to-get-cited-in-google-ai-overviews/).
- **Bing:** grounding queries are grouped phrases, not users’ exact questions. Use them to find prompts missing from your set.
- **GA4:** if an AI source you care about (say, Perplexity) isn’t showing in AI Assistant, check the Session source / medium report. For Perplexity citations, see [How Perplexity picks its sources, and how to get it to cite you](https://goinfinita.com/blog/how-perplexity-picks-its-sources/).

For Gemini (the app), Perplexity, Claude, Grok and DeepSeek, we didn’t find an official visibility report for brands. There, your sampling is the only visibility data; GA4 only shows the clicks that come through.

## Step 7: Report on a fixed cadence

The best cadence is one your team can keep without changing the method. Our recommendation to start:

| How often | What to check | Why |
|---|---|---|
| Weekly | GA4 (AI Assistant) and Search Console | Spot drops or spikes in traffic and impressions |
| Monthly | Full run of your prompt set; the metrics in step 5 | Trend lines and competitor comparison |
| Quarterly | Review prompts, competitors and engines | Keep the set close to how buyers ask |
| After a change | Affected prompts (launch, price change, discontinued product) | Confirm AI picked up the new facts |

A monthly report fits on one page: metrics by engine, change vs. last month, prompts won and lost, top cited sources, and 2 or 3 actions for next month.

## Common measurement mistakes

- **One screenshot as proof.** One answer is an anecdote; a percentage across many answers is data.
- **Testing from your own account.** History and memories can change what you see.
- **Reporting a “rank” in ChatGPT.** Order shifts almost every time; use mention rate and average position.
- **Blending branded and non-branded prompts.** You nearly always show up when you’re named, which inflates the number.
- **Changing the prompt set every month.** No stable set, no trend.
- **Only watching traffic.** Many people decide from the answer and never click.
- **Ignoring accuracy.** An outdated price can cost more than not showing up.

## How Infinita helps

Measuring by hand works to start, but it eats hours as prompts, runs and engines add up. Infinita runs your prompts on a schedule in ChatGPT, Gemini, Google AI, Perplexity, Claude, Grok, Microsoft Copilot and DeepSeek, and shows which prompts your brand appears in and which it doesn’t, for both branded and non-branded prompts. Every answer is saved with its sources, so you can see which competitors show up when you don’t and which pages AI relies on for your category.

With [Ask Vera](https://goinfinita.com/ask-vera/), your AI marketing agent, you can ask “Which non-branded prompts did we lose this month?” and get a plan to win them back. To see where your brand stands in AI answers today, [book a demo](https://goinfinita.com/demo/).

## Sources

1. OpenAI, November 10, 2026: [Introducing Businesses in ChatGPT](https://openai.com/index/introducing-businesses-chatgpt/)
2. OpenAI Help Center, accessed November 19, 2026: [Temporary chat in ChatGPT](https://help.openai.com/en/articles/8914046-temporary-chat-faq)
3. OpenAI Help Center: [Publishers and Developers FAQ, archived copy from October 4, 2026](https://web.archive.org/web/20261004224023/https://help.openai.com/en/articles/12627856-publishers-and-developers-faq)
4. Google Search Central, June 3, 2026: [Introducing Search Generative AI performance reports in Search Console](https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports)
5. Search Console Help: [Generative AI performance report (Search)](https://support.google.com/webmasters/answer/16984139)
6. Google Search Central, November 20, 2025; updated March 11, 2026: [Introducing the branded queries filter in Search Console](https://developers.google.com/search/blog/2025/11/search-console-branded-filter)
7. Google Search Central, November 16, 2026: [New filtering and reporting capabilities in the Search Analytics API](https://developers.google.com/search/blog/2026/11/search-analytics-api-update)
8. Microsoft Bing, February 10, 2026: [Introducing AI Performance in Bing Webmaster Tools Public Preview](https://blogs.bing.com/webmaster/2026/2/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview/)
9. Microsoft Bing, October 13, 2026: [Introducing Topics Portfolio (Beta) in AI Performance](https://blogs.bing.com/webmaster/2026/10/introducing-topics-portfolio-beta-in-ai-performance/)
10. Bing Webmaster Tools Help: [AI Performance](https://www2.bing.com/webmasters/help/ai-performance-9f8e7d6c)
11. Google Analytics Help: [Default channel group](https://support.google.com/analytics/answer/9756891)
12. Google Analytics Help, May 13, 2026 release: [What's new in Google Analytics](https://support.google.com/analytics/answer/9164320)
13. SparkToro, January 28, 2026; updated June 28, 2026: [AIs are highly inconsistent when recommending brands or products](https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/)

## Questions

### How many times should I run each prompt?

There’s no official number. To fully map an AI’s recommendations, SparkToro suggests at least 60 to 100 runs. For monthly tracking, we recommend 3 to 5 runs per prompt per engine to start, with more runs on your most important prompts.

### Can I see brand mentions in ChatGPT officially?

Partly. Since November 10, 2026, ChatGPT’s Business Hub (beta) shows aggregated mentions and website clicks once you claim your business. It doesn’t show individual answers or competitors, so you still need your own prompt sampling.

### Can I test from my own ChatGPT account?

It’s better not to. History, memories and custom instructions can change the answers. Test logged out, or in a Temporary Chat set to “Unpersonalized” before your first message.

### Does GA4 show traffic from Google AI Overviews?

Not separately. GA4’s AI Assistant channel covers sources like ChatGPT, Gemini and Copilot, but visits from AI Overviews and AI Mode are counted as Organic Search. For Google’s AI features, use the Search Console generative AI report, which shows impressions.

### What is AI share of voice, and how is it measured?

It’s your share of brand mentions in AI answers compared with the competitors you track. Divide your mentions by the mentions of all tracked brands across the same answers: 24 of 160 mentions is 15%.

### Does Search Console’s branded query filter measure AI visibility?

No. It splits your classic Google Search performance into branded and non-branded queries. It’s a useful brand awareness signal, but it doesn’t show what happens inside AI Overviews or other assistants.

### What if my numbers swing a lot from month to month?

Check the method first: same prompt set, same runs, same engines, same conditions. If nothing changed, compare against official data and against what you shipped that month before drawing conclusions.
