How to Measure Your AI Visibility: A 5-Step Guide

Measuring AI visibility means asking AI assistants (Gemini, ChatGPT, Claude and Google AI Overviews) the questions your buyers ask, then counting how often your brand appears in the answers, where it appears, and who appears instead. According to the G2 2025 Buyer Behavior Report, "generative AI chatbots" are now the number-one influence on B2B supplier shortlists, ahead of review sites, supplier websites and salespeople (source). If AI answers are building shortlists, you need to know whether you are on them. This guide covers the five measurement steps, the share-of-voice formula, the metrics that matter, and how to do it yourself with a spreadsheet.

What does it mean to measure AI visibility?

Measuring AI visibility means treating AI answers as a measurable channel: you define a fixed set of buyer questions, put them to the engines, and record every brand mentioned in the responses. It is closer to a share-of-voice study than to rank tracking, because AI assistants do not have rankings.

That distinction matters. Answers are probabilistic: SparkToro research found less than a 1-in-100 chance that ChatGPT gives the same list of brands in any two responses to the same query (source). You cannot ask "what position do we rank?" There is no fixed position to hold. What you can measure is the share of buyer questions where your brand gets named, and how that share compares with your competitors'. That is the number the five steps below produce.

Step 1: Define the buyer questions worth measuring

The foundation of any AI visibility measurement is a fixed set of realistic buyer questions: the conversational, multi-constraint questions real people put to an assistant, not keyword strings. "Project management software" is a keyword; "best project management tool for a 10-person agency that bills hourly" is a question a buyer actually asks. Measure the second kind, because that is what the engines are actually answering.

Aim for 20–30 questions, and make most of them non-branded: questions that describe your category and your buyer's constraints without naming your brand. Non-branded questions are the real test of visibility, because the customers you have not won yet do not know your name. If an assistant only mentions you when the buyer already asks about you by name, you are not being discovered: you are being looked up. That is why Skavora deliberately generates roughly 80% non-branded and 20% branded questions for every scan, a split enforced in code, not left to chance. The reasoning is set out in full in how Skavora builds its question set.

A quick checklist for a sound question set:

• Written the way people speak, with real constraints (team size, budget, location, use case)
• Roughly 80% non-branded, 20% branded
• No platform names and no competitor names in any question: naming either contaminates the measurement
• Fixed and reusable, so later runs are comparable

Step 2: Run the questions live against the AI engines

Put every question to the engines live, with web access enabled, and save the full answers. Live answers with grounding (the engine searching the web before it responds) are what buyers actually see; a model answering from its training data alone is a snapshot of the past, not a measurement of the present.

Practical rules for a clean run:

• Use a fresh session per question: no chat history, so earlier answers cannot colour later ones
• Never tell the engine which brand you are measuring; the question should read exactly as a buyer would type it
• Record the complete answer text, every brand named, and every source cited
• Run all questions in one sitting per engine, so the results describe one moment in time

For our worked example, "best project management tool for a 10-person agency that bills hourly", you would put that question to Gemini, ChatGPT and Claude in turn, then search the same phrasing on Google and record the AI Overview if one appears (Google shows them on some queries and not others), logging every tool each answer recommends.

Step 3: Count mentions, position and share of voice

For each answer, record three things: whether your brand was mentioned, where it appeared in the list, and every other brand named alongside it. These three raw counts generate every metric that follows.

AI share of voice is the headline figure. It is your brand's slice of all brand mentions across the full question set:

If your 25 answers name brands 100 times in total and your brand accounts for 12 of those mentions, your Share of AI Voice is 12%. The power of the metric is comparative: it tells you not just whether you appear, but how much of the conversation you hold against the brands the engines actually named, which are not always the competitors you expected.

Coverage is simpler: the percentage of questions where your brand appears at all. Track it separately for non-branded and branded questions. In our worked example, if the agency-billing question names five tools and yours is third, that is one covered question, one mention, and a position of 3, all logged from a single answer.

Position matters because answers are usually lists. Being named first in a shortlist of five is not the same as being the last name in it, and a single average position hides that difference unless you record it. Record the position at every mention and average it across the questions where you appear.

Step 4: Classify the sources behind the answers

Grounded AI answers usually cite the pages they drew on, and those citations tell you where the engines get their information about your category. Classify every cited source into a type (review and comparison sites, retailers, marketplaces, editorial and press, Reddit, YouTube, forums, reference sites, competitor sites, your own site) and count which types dominate.

This step is what turns measurement into a plan. The Google-AI disconnect is well documented: an Ahrefs study of 15,000 prompts found only 12% of URLs cited by ChatGPT, Gemini and Copilot appear in Google's top 10 for the same query (source). On that evidence, your Google rankings are a weak proxy for which sites feed AI answers in your category; the citations in the answers themselves are the direct evidence. Once you can see that, say, review sites and Reddit threads dominate the answers to your buyers' questions, you have a shortlist of places worth earning coverage on next. We break the whole ecosystem down in which sources AI engines actually cite.

Step 5: Track the results over time

Tracking AI visibility over time means re-running the same fixed question set on a regular cadence (monthly is a sensible default) and comparing coverage, share of voice and the source mix with previous runs. One measurement is a snapshot; only a series of them shows direction.

The volatility is real, and it is the reason tracking matters: the Semrush AI Visibility Index, tracking 2,500 prompts across Google AI Mode and ChatGPT, found that 40–60% of AI-cited sources change from month to month (source). A brand that dominates this month's answers can fade from next month's. Keep the question set fixed between runs: change the questions and you are measuring a different thing, not a trend.

Can you measure AI visibility manually?

Yes, a spreadsheet and a spare afternoon will do it, and doing one manual round is a genuinely useful exercise. The DIY version looks like this:

1. Write 20–30 buyer questions in a spreadsheet column (80% non-branded)
2. Paste each into each engine in a fresh session, with web access on
3. Log per answer: mentioned (Y/N), position, every brand named, every source cited
4. Compute coverage, average position and Share of AI Voice with the formula above
5. Repeat next month and compare

The honest caveats: it is slow (25 questions across four engines is 100 answers to run and code by hand), brand-counting is error-prone when answers use partial names, and consistency between runs depends on your discipline.

Skavora automates the scan itself: it researches your brand live, generates the questions, runs them with grounding, and returns the scored report. Scans run on demand (it is not always-on monitoring), so the cadence stays in your hands; re-running is a minute's work rather than an afternoon's. The free check covers one engine, with no signup: run a free AI visibility check and compare it with your spreadsheet. The £7.99 report runs the same questions across all four engines with a cross-engine benchmark, and lets you edit the questions or write your own (up to 25) and re-run.

Which AI visibility metrics matter most?

Five metrics cover most of what AI visibility measurement needs to tell you: visibility score, non-branded coverage, Share of AI Voice, average position and sentiment. One-line definitions:

• Visibility score, a single 0–100 figure summarising how visible your brand is across the full question set, weighted toward non-branded coverage.
• Non-branded coverage, the percentage of questions that never name your brand where the engines mention you anyway; the truest signal of discovery.
• Share of AI Voice, your brand's percentage of all brand mentions across every answer in the set.
• Average position, where your brand typically sits in the answer when it is mentioned; first-named and last-named are different outcomes.
• Sentiment, how the answers characterise your brand when they mention it: recommended, neutral, or caveated.

If you track only two, track non-branded coverage and Share of AI Voice: the first tells you whether new buyers can find you, the second tells you who owns the conversation. For what each of these five numbers points you to do next, section by section, see how to read an AI visibility report and act on it.

Measuring AI visibility comes down to five repeatable steps

• Fix 20–30 conversational buyer questions, ~80% non-branded
• Run them live, grounded, in fresh sessions, never telling the engine which brand you are measuring
• Count mentions, positions and Share of AI Voice
• Classify the cited sources into types and note which dominate
• Re-run monthly with the same questions and track the trend

Measurement is only the first half of the loop. If the numbers come back low, the tactics with published evidence behind them are set out in how to get ChatGPT to recommend your brand.