Why a screenshot proves nothing
Many discussions about AI visibility start with a screenshot: “Look, ChatGPT recommends us.” Or the opposite: “We don’t show up anywhere.” Neither says much.
Generative systems phrase every answer from scratch. A study by SparkToro and Gumshoe.ai examined this systematically in early 2026: 2,961 runs of the same recommendation questions in ChatGPT, Claude, and Google’s AI offerings. The probability of getting the same list of brands twice was under one percent; the same list in the same order was much rarer still [1]. What was more stable, by contrast, was how often a brand was mentioned across many runs [1][2].
Before measuring: can AI even read your website?
Before you measure whether you’re mentioned, check that the prerequisites are in place. Otherwise you’re measuring a technical problem, not a content problem.
- Crawlers allowed? OpenAI distinguishes between crawlers for training (GPTBot), for ChatGPT search (OAI-SearchBot), and for retrieval on behalf of users (ChatGPT-User). According to OpenAI, sites that block OAI-SearchBot in robots.txt won’t be shown in ChatGPT’s search answers [3]. Check your robots.txt – and settings at your host or CDN that can block bots regardless.
- Indexed by Google? Google doesn’t name any special requirements for AI Overviews and AI Mode: a page must be indexed and eligible for regular search; the usual SEO fundamentals apply [4].
- Content readable without JavaScript? Not every crawler executes JavaScript. Key statements should be in the delivered HTML.
The method: a fixed question set
The core of the measurement is a set of questions that you use unchanged for months. For B2B, four question types work well, each measuring something different:
| Type | Example | What it measures |
|---|---|---|
| Brand | “What is [your brand]?” | Is your company described correctly? |
| Category | “Which [your category] providers are there for mid-sized B2B companies?” | Do you show up when buyers build a shortlist? |
| Problem | “How do I get more qualified leads from my B2B website?” | Is your content used as a source for answers? |
| Comparison | “Agency or independent consultant – which is a better fit?” | How are you positioned against alternatives? |
How to build a good set
- Derive it from real buyer questions – from sales conversations, inquiries, and Search Console queries, not from wish-list keywords.
- 15 to 25 questions are enough to start; a few well-chosen ones beat a hundred random ones.
- Phrase them naturally – the way your audience actually asks, in the language of your market.
- Freeze the wording. If you keep rephrasing questions, you can no longer attribute changes to your measures.
Systems, runs, conditions
To keep results comparable, define the conditions once and document them:
- Systems: the ones your audience uses – typically ChatGPT with web search, Perplexity, Google’s AI Overviews or AI Mode, Gemini, and Microsoft Copilot.
- Runs: ask each question several times per system. Three runs are a pragmatic minimum for a small team; more runs make the values more stable. The SparkToro study shows why: only across many repetitions does the share of mentions become meaningful [1].
- Fresh sessions: logged out or without history and personalization, same language, same location.
- Cadence: a baseline measurement before the first measures, then monthly.
- Log: record every answer with date, system, and run – a spreadsheet is enough.
Five metrics that hold up
| Metric | Definition | What it’s good for |
|---|---|---|
| Brand accuracy | Share of runs for brand questions with a correct, up-to-date description | Early indicator: does the AI understand who you are? |
| Mention rate | Share of category and problem questions in which your brand is mentioned | The core of AI visibility – the most stable value across many runs [1] |
| Citation rate | Share of answers that link to your website as a source | Shows whether your content serves as evidence |
| Share of voice | Your mentions relative to all mentions of a fixed set of competitors | Relative position in the market |
| Referral traffic | Visits from chatgpt.com, perplexity.ai, and other AI services in your web analytics | Connects visibility with actual visits |
Deliberately not a metric: position in the answer (“#2 in ChatGPT”). The order fluctuates so much that it has hardly any meaning [1].
Limits of referral data: Google doesn’t report traffic from AI Overviews and AI Mode separately but counts it as regular search traffic in Search Console [4]. And not everyone who reads an AI answer clicks.
Common mistakes
- Treating individual answers as evidence – in presentations as well as in service providers’ proposals.
- Using personalized sessions – your own history skews the answer.
- Constantly changing the questions – then you’re measuring the question, not visibility.
- Asking only brand questions – you’re almost always “visible” there; category and problem questions are what matter.
- Reporting rank positions – they look precise, but they aren’t.
- Promising results. No provider can guarantee that an AI system will mention a brand.
What measurement can’t do
Measurement shows whether something changes – not automatically why. AI systems change models, data sources, and search methods without announcement. An improvement may be due to your measures or to a system update. That’s why it pays to log measures with dates and to evaluate changes only after several measurement points.
The research paper that coined the term “Generative Engine Optimization” examines how content becomes more visible in the answers of generative search systems that draw on retrieved web sources [5].
The systems also differ from one another – depending on provider, model version, enabled web search, and location. So measure each system separately instead of lumping values together.
Template: how to get started
- Write 15–25 questions based on real buyer questions (four types, see above).
- Define three to five competitors.
- Document systems and conditions.
- Run a baseline measurement: every question, every system, at least three runs.
- Calculate and record the five metrics.
- Implement measures and log them with dates.
- Repeat monthly; evaluate only after several months.
Sources
- Search Engine Land: “AI recommendation lists rarely repeat: Study” – report on the SparkToro and Gumshoe.ai study (January 2026)
- Search Engine Journal: “AI recommendations change with nearly every query: SparkToro”
- OpenAI: “Overview of OpenAI Crawlers”
- Google Search Central: “AI features and your website”
- Aggarwal, P. et al.: “GEO: Generative Engine Optimization,” KDD 2024

