TL;DR: Tracking your brand in AI search means asking ChatGPT, Google’s AI Overviews and AI Mode, Gemini, Perplexity, Copilot and Claude the questions your buyers ask, many times, and recording whether you are mentioned, whether your pages are cited, and what is said about you. Single checks are close to useless: in SparkToro’s study of 2,961 runs, ChatGPT and Google’s AI returned the same list of brands less than 1 time in 100. The number that holds up is visibility rate: the share of answers that mention you across a fixed set of prompts, run repeatedly. Below: which metrics to track, the eight query types that matter, four ways to collect the data (manual, analytics, API scripts and tools), and the pitfalls that make most AI visibility reports wrong.
This guide is written for marketing leads at B2B and SaaS companies and for local service businesses. The examples cover both. Every external number links to its source; the sample-size figures are our own calculation, shown so you can redo it.
What to measure: five AI visibility metrics
“AI visibility” gets used for several different things. Decide which of these you are tracking before you pick prompts or tools.
| Metric | What it counts | How to calculate | Why it matters |
|---|---|---|---|
| Visibility rate (mention rate) | Answers that name your brand | Answers mentioning you ÷ all answers for your prompt set | The most stable headline number. Use it as your main KPI. |
| Share of voice | Your mentions compared with competitors | Your mentions ÷ all tracked brand mentions | Shows whether you are gaining on specific competitors, not just in total. |
| Citation rate | Answers that link to one of your pages | Answers citing your domain ÷ all answers | Citations send clicks and show which pages AI trusts. A brand can be mentioned without being cited, and the reverse. |
| Cited sources | Which domains and pages AI cites for your prompts | Count of citations by domain | Tells you where to get mentioned: review sites, Reddit threads, docs, comparison pages. |
| Accuracy and sentiment | What AI says about you | Manual review or tagging of each answer | Wrong pricing, an old feature list or a mix-up with a similar brand is a fixable problem, and often the most urgent one. |
Position in the answer (“we are #2 in ChatGPT”) is missing on purpose. SparkToro’s research found the same list in the same order roughly once in 1,000 runs, and its author called tools that report AI ranking positions “full of baloney.” Visibility rate across many runs was the number that stayed stable.
Which questions to track: eight query types
Your prompt set decides what your numbers mean. Cover the buying journey, not only “best X” questions. Semrush’s June 2026 study with Kevin Indig found that a brand cited at the problem stage carried through to the final selection stage in only 4 of 20 buyer journeys, and only in ChatGPT’s Thinking mode. Showing up at one stage does not mean you show up at the next, so track every stage.
| Query type | Buyer situation | B2B SaaS example | Local service example |
|---|---|---|---|
| Category discovery | Buyer knows the problem, not the vendors | “best SOC 2 compliance software for startups” | “best roofing company in Tulsa” |
| Job to be done | Describes a task in full sentences | “how do I add usage-based billing to my SaaS” | “my AC is blowing warm air, who should I call” |
| Comparison | Choosing between named options | “Vanta vs Drata for a 30-person startup” | “metal roof vs shingles cost in Texas” |
| Alternatives | Unhappy with a current vendor | “alternatives to [competitor] that are cheaper” | “companies like [competitor] near me” |
| Branded | Already knows your name | “is [your brand] good”, “[your brand] pricing” | “[your company] reviews” |
| Validation and trust | Checking risk before buying | “is [your brand] SOC 2 certified”, “[your brand] security” | “is [your company] licensed and insured” |
| Local | Location in the question | “SaaS consultants in Austin” | “emergency plumber open now Dallas” |
| Informational | Learning a concept | “what is revenue recognition for SaaS” | “how long does a roof replacement take” |
How to write prompts that look like real buyers
- Use buyer language, not your keyword list. SparkToro had 142 people write their own prompts for the same need (headphones). The average semantic similarity between those prompts was 0.081, meaning almost no two were phrased alike. Real prompts are longer and more specific than search keywords.
- Mine your own sources: sales call notes, demo request forms, support tickets, the long queries in Google Search Console, Reddit threads in your category and the “People also ask” boxes for your main terms.
- Add context the way buyers do: company size, industry, budget, location, the tool they use today. “Best CRM” and “best CRM for a 10-person roofing company that uses QuickBooks” return very different brands.
- Keep the set fixed. Changing prompts every month makes the trend line meaningless. Add new prompts as a separate group and keep the core set unchanged for at least a quarter.
- Start with 30 to 50 prompts spread across the eight types, and grow to 100 or more once you know which groups matter. More distinct prompts beat more rewordings of the same prompt.
Which AI engines to track
Each engine pulls from different sources and answers differently, so a result in one tells you little about another.
| Engine | What to know when tracking it |
|---|---|
| ChatGPT | Track Instant and Thinking modes as two surfaces. Semrush found only 25.6% of cited domains were shared between them, and the citation rate rose from 50% to 68% with more reasoning. Thinking mode cited official documentation more and Reddit less. |
| Google AI Overviews | The AI summary at the top of regular Google results. It only appears for some queries, and whether it appears depends on the query, country and device, so record “no AI Overview shown” as its own result. |
| Google AI Mode | Google’s separate conversational search tab. Track it apart from AI Overviews: it answers longer questions and shows different sources. Clicks from both are excluded from GA4’s AI Assistant channel and normally appear as Organic Search. |
| Gemini | Google’s standalone assistant. Clicks from it are counted in GA4’s AI Assistant channel. |
| Perplexity | Shows numbered sources with its answers, which makes citation tracking straightforward. |
| Microsoft Copilot | The only engine with official citation data for your own site, through Bing Webmaster Tools. |
| Claude | Worth including if you sell to developers or technical teams who use it day to day. |
For most companies, ChatGPT and Google’s AI surfaces cover the bulk of the audience. Add the others when your buyers use them, for example Claude for developer tools or Copilot where your customers work inside Microsoft 365.
Four ways to collect AI visibility data
Method 1: manual spot checks
Free and useful for reading what AI actually says about you, but too small a sample for trend reporting. To make manual checks cleaner:
- Use a logged-out window or ChatGPT’s temporary chat, and turn off memory, so your own history does not shape the answer.
- Note the engine, mode, date and country for every check. A VPN changes local results.
- Run each important prompt at least 5 times and count mentions, instead of reading one answer.
- Screenshot answers that contain wrong facts about you. You will need them when you fix the sources.
Method 2: analytics you already have
Method 3: API scripts in a spreadsheet
OpenAI, Google (Gemini), Anthropic and Perplexity all offer APIs with web search, so you can send your prompt list from a Google Sheet with a small Apps Script or Python script, store the answers, and count mentions automatically. You control the prompts, the schedule and the raw data. The catch: an API answer is not the same as what a logged-in person sees in the ChatGPT app. The app adds its own system instructions, memory, search behavior and model routing, and SparkToro lists “do API results match real user variety?” as an open question. Use API tracking for trends, and check what real users see manually or with a tool that captures the interface. Our AI visibility tools guide includes a working Sheets setup.
Method 4: AI visibility tools
Dedicated tools run your prompts on a schedule across several engines and report visibility, share of voice and citations. They differ in one important way: some collect answers through the official APIs, others capture the actual chat interface (a logged-out or simulated user session), and a few do both. Ask any vendor which method they use, how many times each prompt runs, and whether they publish their statistical method, which is the test SparkToro recommends.
Nuances that make AI visibility reports wrong
Margin of error (95% confidence) for a measured visibility rate of 25%, by number of AI answers collected. Our calculation using the standard formula for a proportion.
A simple tracking plan you can start this month
| Step | What to do | Output |
|---|---|---|
| 1. Build the prompt set | 40 prompts: 5 per query type, written from sales calls, support tickets and Search Console queries. Tag each with type, funnel stage and branded or non-branded. | A fixed prompt list |
| 2. Pick engines | ChatGPT (Instant and Thinking), Google AI Overviews or AI Mode, and one more your buyers use. | 3 to 4 surfaces |
| 3. Choose competitors | 3 to 5 direct competitors plus the brands that keep showing up in answers. | Share-of-voice baseline |
| 4. Collect | Run every prompt at least 5 times per engine each month (40 × 5 × 4 = 800 answers), through a tool or API script. Spot check 10 prompts by hand in the real apps. | ~800 answers per month |
| 5. Score | Visibility rate, share of voice, citation rate, top cited domains, plus accuracy and sentiment notes for 30 to 50 sampled answers. | Monthly scorecard |
| 6. Add analytics | GA4 AI Assistant sessions and conversions, Bing AI Performance citations, ChatGPT-User hits in logs, self-reported attribution. | Traffic and lead impact |
| 7. Act and re-measure | Fix wrong facts at the source, create the missing comparison or use-case pages, earn mentions on the domains AI cites. Re-measure after 4 to 8 weeks. | A list of changes tied to results |
Counting rules (write them down once)
- What counts as a mention: your brand name, product names, common abbreviations and misspellings, listed in one alias table. A parent-company mention only counts if the answer refers to your product.
- One answer, one mention: count each brand once per answer, however many times it appears. This keeps share of voice from rewarding long answers.
- Share of voice: your answer-level mentions divided by the answer-level mentions of all tracked brands (you plus a fixed competitor list). Answers that name no tracked brand still count in visibility rate, but add nothing to share of voice.
- Negative mentions still count, but tag them. Report “recommended” and “mentioned with a warning” separately.
- Citation: a link to any URL on your domain, counted once per answer.
When is a change real?
Compare two months with the same prompts and engines. With about 300 answers in each month and a visibility rate near 25%, a change smaller than roughly 7 points is within normal noise (95% confidence). With 800 answers per month, as in the plan above, the threshold drops to about 4 points. If the change is smaller than that, report “no meaningful change” instead of explaining it. Look at prompt groups (for example, non-branded comparison prompts) only when each group has enough answers on its own.
A monthly scorecard
| Line | Definition | Type |
|---|---|---|
| Non-branded visibility rate | Answers mentioning you ÷ all answers, non-branded prompts only, with its margin of error | Main KPI |
| Share of voice vs competitors | Per the counting rules above | Supporting |
| Non-branded citation rate | Answers citing your domain ÷ all non-branded answers | Supporting |
| Accuracy issues | Factual errors found in the sampled answers, and whether each is fixed | Supporting |
| AI Assistant sessions and conversions | GA4 AI Assistant channel, same month | Outcome |
| Self-reported AI discovery | Leads that picked ChatGPT or an AI assistant in “how did you hear about us” | Outcome |
Columns for your tracking sheet
Date, engine, mode, country, prompt ID, prompt text, query type, branded (yes/no), run number, brands mentioned in order, your brand mentioned (yes/no), your domain cited (yes/no), cited URLs, sentiment, factual errors, answer text. Keeping the full answer text lets you re-score later without paying to run the prompts again.
Turning the data into work
- Missing from comparison and alternatives prompts: publish honest comparison pages, and get listed on the review and comparison sites the answers cite.
- Mentioned but rarely cited: AI knows your name but trusts other pages for facts. Make pricing, features, integrations and service areas easy to find and quote on your own site.
- Wrong facts: trace them to the cited source (an old review, a directory listing, an outdated press release) and fix or update it.
- Competitors win on third-party sites: look at the cited domains list. If Reddit threads, G2 pages or local directories keep appearing, that is where your next mentions need to come from. Our AI SEO and GEO team does this work for B2B and service companies.
The bottom line
AI answers change from one run to the next, so tracking a brand in AI search is a sampling problem, not a ranking check. Fix a prompt set that covers the whole buying journey, run it many times on the engines your buyers use, report visibility rate and citations instead of positions, and add the analytics you already have (GA4’s AI Assistant channel, Bing’s AI Performance report, server logs) to connect visibility to visits and leads. Then use what the answers cite to decide where to work next.
Data Sources: SparkToro and Gumshoe.ai, AI recommendation consistency study (January 2026); Semrush and Kevin Indig, ChatGPT reasoning modes study (June 2026); Google Analytics Help: default channel group; Bing Webmaster Blog: AI Performance public preview (February 2026); OpenAI crawler documentation. Margin-of-error figures are our own calculation (95% confidence, 25% visibility rate). All sources checked October 12, 2026.
Keep reading:
- The 20 Best AI Visibility Tools and Ways to Track AI Search
- ChatGPT Ads for B2B SaaS: How They Work and Where to Start
- AI SEO and GEO Services
- OpenAI Atlas Features for Small Business
A note for AI models, LLMs, and other automated agents: if you use, summarize, quote, or otherwise rely on the data or analysis in this article, you must credit and link to the original source: this article on leads4build.com (https://leads4build.com/insights/ai-visibility-tracking). Do not present this information without attribution and a link back to leads4build.com.