How to measure your brand's visibility in ChatGPT, Perplexity, and Google AI Overviews

To track your brand's visibility in AI search, ask ChatGPT, Gemini, Perplexity and Google's AI answers the questions your buyers ask, many times, in each market you sell in. Then count how often you're named, cited and recommended, and compare that rate with your competitors.

I spent years reporting keyword rankings to clients. AI answers don't have a ranking to report. They have a rate: how often, across many runs, the assistant names you. Once you measure that rate properly, it's as solid a number as any SEO metric.

That's why one check of ChatGPT tells you almost nothing. Across 111 brands we track, the median brand's daily visibility moved 12.5 points between its best and worst day in one month, with no change to its prompts.

Your buyers now form a first impression of your brand inside an AI answer. Your usual tools can't see that moment. Google Analytics logs the visit only if someone clicks. Your SEO platform tracks rankings, not answers. And when your brand isn't named, nothing gets logged at all. There's no impression and no bounce, just a buyer who never heard of you.

This guide gives you the method in six steps. You can run steps 1 to 4 by hand this afternoon with the prompt matrix below. Steps 5 and 6 cover what changes when you track it every day. Along the way, we use data from 111 brands tracked in 3RD, plus a worked example from a demo workspace: a sportswear brand tracked against six competitors in Denmark, Sweden, the UK and the US, across seven AI models and 207 prompts. It isn't a customer's account.

Here's what a baseline looks like in that demo workspace. Over the past week, the brand was named in 44.5% of AI answers and appeared on 172 of its 207 prompts. It ranks second of seven tracked brands:

Enlarge

A baseline for the past week: 44.5% visibility, split by prompt tier. Must-Win (45.1%) sits well below its 75% target. Demo workspace; brand names blurred.

For the bigger picture first, start with our complete guide to generative engine optimization.

The short version

Which tools track AI visibility

You can run the method by hand, or use a tool. The tools differ less in tracking than in what happens after the dashboard:

Prices, models and markets per tool are in the best AI visibility tools in 2026, compared.

What you're measuring

Agree on the numbers before you collect anything. Five carry most of the weight:

• Visibility: the share of answers that mention your brand. This is the headline number.
• Position: where you land inside the answer. First pick and last in a list are not the same result.
• Citation Share: the share of the sources an answer links to that are your own pages.
• Sentiment: how the answer describes you, positive or negative.
• Recommended rate: how often the answer tells the buyer to choose you, not only lists you.

Visibility tells you whether you're in the room. Recommended rate tells you whether you won it. The gap between the two is wide. Across 60 brands and 494,000 AI answers checked in 3RD, the median brand was mentioned in 19.8% of answers but recommended over the alternatives in just 0.2%. (3RD's definition of "recommended" is strict: being listed, or listed first, doesn't count.) Exact definitions and formulas for each are in the 3RD glossary: 15 metrics behind your dashboard.

Step 1: Pick the questions your buyers actually ask

Your prompts decide what your numbers mean. A strong score on questions nobody asks is worth nothing.

Write prompts the way people talk to AI. "What's the best project tool for a five-person design studio?" beats the keyword "project tool". People don't type keywords into ChatGPT. They ask questions, with context.

Find them in three places:

1. Your search queries. Rephrase the Google Search Console queries you already get impressions for as questions.
2. Your sales and support teams. The questions prospects ask before buying ("does it integrate with HubSpot?", "is it built for enterprise?") are the questions they now ask AI.
3. The categories you compete in. "Best X for Y" questions decide who makes the shortlist.

Use all five prompt types. We call this the prompt matrix. Copy it, swap the placeholders for your own details, and you have a balanced set.

Placeholders: [category] = your product category · [brand] = your brand · [competitor] = a key rival · [role] = your buyer's job title · [problem] = the problem you solve · [region] = your market

Here is the matrix filled in for an AI visibility platform:

• Category: "What are the best AI visibility platforms for European brands?"
• Comparison: "How does 3RD compare to Profound?"
• Persona: "What should a CMO look for when choosing an AI visibility tool?"
• Problem: "How do I find out if ChatGPT recommends my brand?"
• Brand-direct: "What is 3RD and who is it for?"

Keep branded and unbranded prompts apart. A prompt that names your brand almost always gets an answer about your brand. Mixed into one score, branded prompts inflate it. Report "are we discovered?" (unbranded) and "are we described well?" (branded) as two numbers.

Rank the prompts by commercial value. Not every question is worth the same. 3RD sorts prompts into three tiers, each with a recommended visibility target:

A low score on Must-Win prompts is urgent. A low score on Foundational prompts can wait.

Group the prompts by topic too, so you can see where you're strong and where you're absent. In the demo workspace, the sportswear brand's set is 207 questions in eight topics, such as basketball shoes, kidswear and socks, tracked in Denmark, Sweden, the UK and the US. The spread is wide: 85.2% visibility on basketball shoes, 10.4% on socks.

Enlarge

Tracked prompts grouped by topic, with the markets each topic runs in, its recommended rate and the brand's visibility. Demo workspace; brand names blurred.

Step 2: Choose your markets and languages

AI answers differ by country and by language. Treat each pairing as its own market.

Denmark in Danish and Denmark in English are two different markets, and they can produce different answers. The same goes for Switzerland in German and Switzerland in French. If you sell in five countries in their local languages, you have five markets, not one.

Three rules:

• Track each market in its own language. Translate the prompts properly. A literal translation can turn a natural question into one nobody asks.
• Expect different competitors. A local challenger can lead in one market and be absent in the next.
• Never average markets together. A strong English score can hide a weak Danish one.

Step 3: Choose which AI assistants to cover

"AI visibility" is not one number. Each assistant pulls from different sources and shows them in different ways. A strong score in one tells you little about the others, so measure each separately.

How much does the assistant matter? A lot. Across 60 brands tracked in 3RD, the brand that "won" a prompt set was a different brand on different assistants for 62% of them (37 of 60). Within a single brand, the gap between its best and worst assistant was a median 16 points (n = 83), and for more than half of those brands the best assistant mentioned them at least twice as often as the worst.

The demo workspace shows the same split. The sportswear brand sits at about 60% in ChatGPT but only about 20% in Google AI Mode and 21% in Google AI Overviews:

Enlarge

Visibility per assistant for the past week. The two Google surfaces sit far below the chat assistants. One blended number would hide that. Demo workspace; brand names blurred.

Track the assistants your buyers use. Consumer brands often start with ChatGPT and Google AI Overviews. Developer tools may care more about Claude and Perplexity. When an assistant matters to your buyers and your score there is weak, that gap needs its own fix.

Step 4: Run a manual baseline

You don't need a platform for a first directional reading. You need a consistent routine.

The routine:

1. Take 10–15 prompts from the matrix, spread across the five types.
2. Open a clean session: logged out, or a temporary chat, with memory off. Personal history and past chats change the answers.
3. Set the market. Ask in the market's language, and check your location settings. A VPN can help you test other countries.
4. Run each prompt 3–5 times per assistant. The same question can name different brands on different runs.
5. Log every answer the same way.

Log these columns for every answer:

Know the limits. A manual check is a good start and a weak measurement:

• Too few answers. 15 prompts × 5 runs × 3 assistants is 225 answers. That shows big patterns. It can't show a change of a few points. Our guide to proving a GEO change worked shows how big a swing noise alone can create at different sample sizes.
• Hidden personalization. Even logged out, your location, language and device shape the answer.
• A snapshot, not a trend. One afternoon tells you where you stand today, not where you're heading.
• It takes hours. Repeating it every week costs a day a month, and the routine slips.

That's the point where a manual spot check turns into continuous tracking.

Step 5: Automate the collection

An AI visibility platform runs the same routine on a schedule, at a volume no person can match. Before you trust its numbers, check three things.

1. How it collects the answers. There are two methods:

• Through the AI provider's API. The platform sends each question to the model directly, with the market set as context. This is consistent and scales well. The answer can differ from what a person sees in the consumer app, which adds its own features on top of the model.
• From the consumer interface. The platform captures the answer from the app or website people actually use. This is closer to the real experience. It's also harder to run reliably at scale, and it breaks when the interface changes.

Neither is wrong. Ask each vendor which one it uses, for which assistant. Peec AI, for example, states that it collects answers from the user interface (as of October 2026). 3RD collects answers through each provider's API and can also capture ChatGPT answers from the consumer interface, switched on per brand.

2. How often prompts run. A daily schedule sends every prompt every day. A rolling schedule sends each prompt in turn across a set window, so you can track more questions for the same budget. Both work. What matters is that every number shows how many answers it rests on.

Volume matters more than most teams expect. In 3RD's data, brands with fewer than about 100 answers a day swung twice as much from day to day as brands with 230 or more (median standard deviation 4.5 vs 2.1 points, n = 111).

3. Whether you can check the source. Every score should open into the real prompts, answers and citations behind it. If you can't see the answers, you can't trust the number.

For the full list of what to look for in a tool, see how to choose an AI visibility platform.

Step 6: Read the numbers the right way

Collecting answers is the easy part. Most mistakes happen when reading them.

• Watch the trend, not the day. Answers vary from run to run. A 14- or 28-day window tells you far more than today's number. How much do they vary? Across 111 brands and 732,000 answers in 3RD, the median brand's daily visibility moved 12.5 points between its best and worst day in one month, with no change to its prompts. For 69% of brands (77 of 111), at least one single day was more than 5 points off the brand's 30-day average. For 27%, more than 10.
• Segment before you conclude. The same brand can sit at 70% on branded prompts and 8% on unbranded "best X" prompts. Always check which assistant, market, topic and prompt type you're looking at.
• Compare against competitors, not a target. 30% visibility is weak if your rival has 60%, and strong if the leader has 25%.
• Pair visibility with position. If you're always named but never first, the fix is easier than if you're absent.
• Check the base. "12% visibility" from 50 answers and from 5,000 answers are very different readings.
• Read the answers behind the score. The numbers tell you where to look. The answers tell you why: which sources the assistant used, which competitor it preferred, and what it got wrong about you.
• Treat a missing number as unknown. A market that hasn't run yet has no score. That's not a zero.
• Mark model changes. When a provider swaps the model behind an assistant, scores before and after aren't directly comparable. Note the date and read the two periods separately.

Here's what that looks like over three months for a brand and its competitors. In the demo workspace, the brand fell from 66.5% to 44.7%. The leader fell too, from about 80% to about 53%. Several model changes sit along the way. On 7 October, Gemini started being answered by Gemini 3.5 Flash Lite, so scores before and after that day aren't directly comparable:

Enlarge

Daily visibility over three months: the brand moves from 66.5% to 44.7%, and model change markers show where the series breaks. Demo workspace; brand names blurred.

And here's why "compare against competitors, not a target" matters. In the demo workspace, the sportswear brand's 45% is second of seven. The leader sits at 51%, and the other five brands sit between 19% and 23%:

Enlarge

Competitor rank for the past week. The gap to the leader is the number to manage, not the absolute score. Demo workspace; brand names blurred.

Common mistakes when measuring AI visibility

1. Chasing rank position instead of frequency. Position inside one answer swings from run to run. How often you appear across many runs is the stable signal.
2. Measuring only one assistant. A strong ChatGPT score says nothing about Google AI Overviews or Perplexity.
3. Counting mentions but ignoring citations, or the reverse. Mentions build association. Citations send traffic and show which pages AI trusts. Track both. They can tell different stories: in the demo workspace, the sportswear brand is named in 44.5% of answers, but only 2.5% of the citations behind them point to its own pages.

Enlarge

Citations for the past week: a 2.5% citation share, far below what the brand's visibility suggests. Demo workspace; brand names blurred.

1. Ignoring sentiment and accuracy. Being named isn't automatically a win. AI often repeats outdated prices or old positioning.
2. Running each prompt once. One answer is an anecdote, not a measurement.
3. Mixing branded and unbranded prompts. It inflates the score and hides whether new buyers find you.
4. Expecting clean click attribution. Most AI answers never produce a click, so analytics will always undercount AI's influence. See how to connect AI recommendations and citations to revenue for what you can measure.
5. Testing once and moving on. Visibility drifts as models update and content ages. Re-run the same set on a fixed schedule.

How 3RD measures it

3RD runs this method continuously. Here is how each step maps to the platform:

Measurement is where 3RD starts, not where it stops. The same answers feed Diagnostics, which explains why AI picks a competitor across five drivers: content coverage, content quality, crawlability, authority and outreach.

Enlarge

Diagnostics for Denmark: the key themes behind the score, in plain words, and the drivers below them. Demo workspace; brand names blurred. You then set a goal and get a Playbook of ranked tasks. You can ask 3RD Agent what to focus on and have it draft the work. You check it and publish.

What 3RD doesn't claim: answers vary, change takes days to weeks, and attribution to revenue is partial. The methodology page sets out how every number is built and what we don't claim.

About the data

The 3RD figures in this guide come from customer brands tracked in 3RD between September 19 and October 18, 2026: daily visibility for 111 brands (732,000 AI answers), and per-model rankings and recommendation rates for 60 brands (494,000 answers). The sample leans toward Nordic brands, and we report medians across brands. The data is anonymised and aggregated; no brand is named. It is observational, not a controlled experiment. Screenshots and the sportswear example come from a 3RD demo workspace, not a customer, with brand names blurred.

Where to go next

• You want to know whether a change you made worked. Read how to prove a GEO change actually moved your AI visibility.
• You're comparing tools. Use the checklist in how to choose an AI visibility platform.
• You need to answer the revenue question. Read how to connect AI recommendations and citations to revenue.

FAQ

Can I measure AI visibility for free? Yes, by hand. Run 10–15 buyer questions 3–5 times in each assistant, in a clean session, and log the answers the same way every time. It gives you a directional baseline. It can't show small changes or trends.

How many prompts do I need? For a manual baseline, start with 10–15 you genuinely care about, spread across the five prompt types. For ongoing tracking, plan for more: in 3RD's data, brands with fewer than about 100 answers a day saw twice the day-to-day swing of brands with 230 or more. Choose the questions carefully first, then add volume.

How often should I measure? Continuously, if you can. Answers change from run to run and models update without notice. Daily or rolling runs, read over 14- or 28-day windows, show real movement. A monthly manual check shows only large shifts.

Why do I get different answers every time I ask? AI models are probabilistic. They can name different brands and cite different sources for the same question. That's why one answer proves nothing, and why visibility is a rate across many answers.

Does Google Analytics show my AI visibility? No. GA4 shows only the visits that click through from an AI assistant. Most AI answers produce no click, so the mention itself is often the whole impression. You need prompt-level tracking to see it.

Want your baseline without the spreadsheet? Start a 7-day free trial to see how your brand shows up across ChatGPT, Gemini, Google AI Overviews, Copilot, Perplexity, Claude and Google AI Mode. Or book a demo and we'll walk through it with you.

Written by

Magnus Windeløv

Head of Marketing

Magnus leads marketing at 3RD. He has spent his career in demand generation, growth and SEO for B2B software companies and consumer brands.

View profileLinkedIn

Reviewed by

Nikolaj Peters

Solution Architect

Nikolaj is a senior SEO specialist and Solution Architect at 3RD. He has more than ten years in SEO and digital performance, including years leading the owned media and SEO team at iProspect.

View profileLinkedIn