Reading time: 13 minutes
To measure brand visibility in AI, count how many answers mention the brand, how many cite its site and how much traffic that brings. Because answers change with every run, measurement uses fixed, repeated prompts. This guide shows the full protocol for agencies and SEO teams.
In short
- A one-off check measures nothing: according to SparkToro, the chance that ChatGPT repeats the same list of brands is below 1 in 100.
- The metric that holds up is the presence rate across many repeated prompts, because the position in the list changes almost every time.
- Mentions and citations move separately, and each engine favors one of them. So the dashboard needs both.
- To measure brand visibility in AI, a minimum protocol per client has 50 prompts, 5 runs per engine and 4 engines per month.
What to measure for AI visibility: presence, citation and recommendation
A brand shows up in an AI answer in three ways: mentioned by name, with its site cited as a source, or recommended as an option. These are different metrics, so the dashboard needs all three to measure brand visibility in AI, plus the traffic that reaches the site. If the concept is new to you, start with the guide on what GEO is.
| Metric | What it counts | How to calculate it |
|---|---|---|
| Presence rate | Answers that mention the brand | Answers with the brand ÷ total × 100 |
| Citation rate | Answers that cite the brand’s site as a source | Answers citing the domain ÷ total × 100 |
| Recommendation rate | Answers that suggest the brand as an option | Answers recommending the brand ÷ total × 100 |
| Share of voice | The brand’s share among the tracked competitors | Brand mentions ÷ mentions of the brand and competitors × 100 |
| AI traffic | Visits and conversions from assistants | GA4, Search Console and Bing Webmaster Tools |
Count a recommendation when the answer suggests the brand as an option, in a list or with a verb like “consider”. A passing mention, however, counts only toward presence. Tools on the market use similar formulas under different names, so compare numbers only within the same tool.
What is the difference between a brand mention and a citation in AI?
A mention is when the answer says the brand’s name. A citation is when the answer uses the brand’s site as a source, with a link. ChatGPT tends to cite the site without saying the name, but Gemini tends to do the opposite. That is why the dashboard separates the two rates, engine by engine.
Indeed, a Semrush study across 14 countries found that 61.7% of a site’s appearances were citations without the brand being named. In Gemini, the brand was named in 83.7% of appearances, while in ChatGPT the site was cited in 87%. In Brazil, brands were named in only 18% to 22% of answers, versus 50% in India and Sweden. But the study does not disclose the sample by country.
Most citations also come from outside the brand’s site. In the May 2026 study by Muck Rack, about 84% of citations came from earned media, such as news outlets, research and encyclopedias. In the dashboard, therefore, list the domains the AI cites when it talks about the client’s category. Each one is a PR or content target. See which sources ChatGPT cites most and how to earn presence in them.
Why a one-off check can’t measure brand visibility in AI
The same question produces different answers on every run. In other words, the brands, the order and the cited sources change. Each engine also consults its own sources. As a result, only a set of prompts repeated many times shows real brand visibility in AI and separates a trend from chance.
For example, SparkToro asked 600 volunteers to run 12 prompts 2,961 times in ChatGPT, Claude and Google’s AI. The chance that ChatGPT or Google’s AI would repeat the same list of brands was below 1 in 100. Rand Fishkin, SparkToro’s cofounder, summed up what this means for tools on the market:
“Any tool that gives a ‘ranking position in AI’ is full of baloney.”
Rand Fishkin, SparkToro (2026)
Sources also rotate. According to SISTRIX, which tracked 82,619 prompts for 17 weeks, 56% of the domains cited by AI Mode were new each week.
How to measure brand visibility in AI in 8 steps
The protocol fits in eight steps. The first three happen once a quarter: engines, competitors and the prompt list. The others, however, repeat every month, from collection to report. Thus, with a frozen list and a standard spreadsheet, each month is comparable with the previous one.
- Choose four engines: one starting point is ChatGPT, Gemini, AI Overviews and AI Mode. Swap one of them for Perplexity, Copilot or Claude if the client’s audience uses those tools more.
- Define 3 to 5 competitors for share of voice.
- Write 40 to 60 prompts and freeze the list for a quarter.
- Prepare the collection: a new chat for each run, no personalization, access from Brazil and prompts in Portuguese.
- Run each prompt 5 times per engine over the month.
- Calculate presence, citation, recommendation and share of voice, with the margin of error.
- Export GA4, Search Console and Bing Webmaster Tools data for the same period.
- Compare with the baseline and list the actions for the following month.
The spreadsheet has one row per run: date, engine, prompt, run number, brand named, brand recommended, site cited, competitors named, domains cited, sentiment and the full answer.
How to build the prompt list
To measure brand visibility in AI, the list should reproduce the questions customers ask before buying, in Portuguese and with local context. We recommend 40 to 60 prompts per brand, in four intent groups. Pull the questions from Search Console, the sales team, support tickets and Google’s “People also ask” box.
- Category: unbranded questions, such as “qual o melhor software de gestão para clínicas” (“what is the best management software for clinics”).
- Problem: the use case, such as “como reduzir faltas de pacientes em consultas” (“how to reduce patient no-shows”).
- Comparison: the brand versus competitors, such as “marca A ou marca B para clínicas pequenas” (“brand A or brand B for small clinics”).
- Brand: pricing, complaints and alternatives. This group is for checking accuracy and sentiment.
For businesses with a physical address, build prompts by city and neighborhood, as in the guide to GEO for local business. Similarly, stores should include shopping prompts with a price range, as the guide to GEO for ecommerce shows.
How many times should you run each prompt?
To measure the presence rate across the prompt set, we recommend 5 runs per prompt in each engine, spread over the month. For conclusions about a single prompt, however, use 10 to 15. According to SparkToro, mapping every brand the AI recommends for a single question takes far more: 60 to 100.
After all, sample size sets the margin of error. The table shows the margin for a presence rate at 95% confidence. The real margin falls within the range, because repeated runs of the same prompt resemble one another.
| Answers per month | Collection | Margin of error |
|---|---|---|
| 30 | 10 prompts, 3 runs, 1 engine | ± 18 to ± 31 points |
| 250 | 50 prompts, 5 runs, 1 engine | ± 6 to ± 14 points |
| 1,000 | 50 prompts, 5 runs, 4 engines | ± 3 to ± 7 points |
Two practical rules come out of the table. First, more prompts reduce the margin more than more repetitions do. Second, with 1,000 answers a month, only treat a change of more than 4 to 10 points as real, and only after the next month confirms it.
How to control the collection
Controlling the collection means fixing everything that changes the answer except time. That is, use new chats with no personalization, the same country, the same language and the same engines. Also, record the date and engine of each answer. Otherwise, the dashboard shows changes in the collection rather than in brand visibility in AI.
- Personalization: ChatGPT can use the user’s location and memories when it searches, according to OpenAI. So use a temporary chat in ChatGPT and an incognito window in the other engines.
- Location and language: collect from Brazil and ask in Portuguese, because both change the cited sources.
- Interface, not API: to measure what consumers see, collect in the interface. After all, in an academic audit, the sources from ChatGPT’s API and from its interface overlapped by only 12%.
- AI Overviews: record separately the prompts where the summary does not appear, and then calculate presence only on the ones where it did.
How to track the traffic your brand gets from AI
You track AI traffic in three free sources. GA4 shows visits and conversions from assistants. Search Console shows impressions in AI Overviews and AI Mode. Bing Webmaster Tools shows citations in Copilot. Finally, server logs complete the picture.
Google Analytics 4
Since May 2026, GA4 has had the AI Assistant default channel, with visits from assistants such as ChatGPT and Gemini. However, clicks from AI Overviews and AI Mode are not part of it and stay in Organic Search.
To control the list of sources, create a custom channel group, as the GA4 help explains. In Admin, open Data display and then Channel groups. Create an AI channel with the condition Source matches regex, using the regex below. Then, under Reorder, place it above Referral.
^(chatgpt\.com|chat\.openai\.com|(www\.)?perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai)$The full step-by-step process, with the exploration and what the channel leaves out, is in the guide to AI traffic in GA4.
Still, some traffic arrives with no source. According to Cloudflare, clicks in the Claude app arrive without the referrer header. As a result, analytics records those visits as direct traffic.
Search Console and Bing Webmaster Tools
In 2026, Search Console added the Generative AI performance report. According to Google’s help page, it shows the site’s impressions in AI Overviews and AI Mode, without clicks or queries.
The AI Performance report in Bing Webmaster Tools, on the other hand, shows citations by page. It also shows the searches the AI ran to find the cited content. It covers Copilot and selected partners. To read these numbers alongside the drop in clicks, see the SEO and AI guide.
Server logs
In the logs, ChatGPT-User, Perplexity-User and Claude-User indicate visits triggered by user questions, according to the documentation from OpenAI and the other companies. So filter the log by these bots and count the pages fetched per week. If you lack access to the raw log, use the CDN logs. On Cloudflare, for example, AI Crawl Control shows the status of each AI crawler.
How to measure sentiment and accuracy
You measure sentiment and accuracy by reading what the AI says when it answers the prompts from the “brand” group. Classify each mention with fixed criteria and note every wrong piece of information about price, product or company, along with the source the AI cited. The fix usually lies outside the site, for instance in a directory or an old news story.
- Positive: the answer recommends the brand or cites a strength.
- Neutral: the answer only names or describes the brand.
- Negative: the answer points out a problem, complaint or caveat.
Errors are common. In the Searchable report, for example, brands themselves flagged as incorrect 9.2% of the 32,556 facts that AI tools stated about them.
How to present AI visibility results to the client
The monthly report compares brand visibility in AI with the baseline, always with the same prompts, engines and competitors. It also shows the margin of error, explains the changes that fall within it and closes with the actions for the following month.
| Block | Metric | Data source |
|---|---|---|
| Visibility | Presence and recommendation by engine | Prompt collection |
| Competition | Share of voice against 3 to 5 competitors | Prompt collection |
| Authority | Site citations and top third-party sources | Prompt collection |
| Reputation | Sentiment and errors found | Brand group prompts |
| Traffic | AI sessions and conversions | GA4 |
| Google and Microsoft | AI impressions and citations | Search Console and Bing |
To turn this report into a recurring contract, see how to price GEO services at an agency.
When does manual collection stop being practical?
Manual collection stops being practical when the volume passes a few hundred answers a month. With 15 clients, for example, the minimum protocol to measure brand visibility in AI adds up to 15,000 answers a month. At that scale, collection moves to either a GEO tool or an in-house script. Before you choose, ask the vendor these four questions:
- Is collection done through the interface or the API?
- From which country, in which language and with how many runs per prompt?
- Does the tool separate mention, citation and recommendation?
- Does it export the full answers for auditing?
What we found in Portuguese-language articles on brand visibility in AI
We ran 26 searches on how to measure brand presence in AI and analyzed the 8 Portuguese-language pages that appeared most often. Six warn that the AI’s answer changes with every run. But only three say how many times to run each prompt, with recommendations from 3 to 60 runs. None includes the regex for the AI channel in GA4.
Frequently asked questions
Yes, while the volume is small. That is because manual prompt collection, GA4, Search Console and Bing Webmaster Tools cost nothing. When the volume passes a few hundred answers a month, collection moves to a tool or an in-house script, as the GEO tools comparison shows.
No. Clicks from AI Overviews and AI Mode stay in GA4’s Organic Search channel. The AI Assistant default channel, created in May 2026, groups visits from assistants such as ChatGPT and Gemini. Impressions in Google’s AI features, on the other hand, are in Search Console’s Generative AI performance report.
Because, according to Cloudflare, clicks in the Claude app reach the site without the referrer header. Without that header, analytics records the visit as direct traffic. Server logs complete the picture, because Claude-User indicates visits triggered by user questions.
Want to know how your brand shows up in AI answers today?
NotHuman is opening spots for agencies and brands to track how they appear in AI answers. Join the waitlist and we’ll let you know when a spot opens up for you.
Main sources
- SparkToro, study on how inconsistent AI brand recommendations are (2026).
- SISTRIX, stability of the sources cited in AI search results (2026).
- Semrush, ghost citations study across 14 countries (2026).
- Google Analytics, custom channel groups (2026).
- Google Search Console, Generative AI performance report (2026).
