Measuring visibility in AI search: share of model voice

Why rankings no longer tell the whole story

For twenty years, search visibility meant positions on a results page. You could check where you ranked for a term and roughly predict the traffic. AI assistants break that model. When someone asks ChatGPT, Perplexity, Gemini or Copilot “which agencies in Copenhagen do event production”, the answer is a short composed paragraph naming a handful of companies, sometimes with links and often without.

There is no position 1 to 10 in that answer, and no public tool that reports how often you appear. Yet for many buyers it is now the first shortlist. If your company is never named, you are not in the running, however well you rank in the classic results.

Share of model voice borrows from the old advertising idea of share of voice. Instead of measuring how much of the conversation in media you own, it measures how often you are part of the answers that AI systems give to the questions your buyers ask. It is the main metric behind generative engine optimisation, and it can be tracked with nothing more than a spreadsheet and some discipline.

Building a prompt set that reflects real buyers

The prompt set is the foundation. If it is wrong, the numbers are meaningless. Aim for 30 to 100 prompts, written the way buyers actually phrase questions to an assistant, which is usually longer and more conversational than a Google search.

• Category prompts: “best wedding planners for a destination wedding in Greece”, “who does corporate event sound in Copenhagen”.
• Problem prompts: “my website is not showing on Google, who can fix it in Denmark”.
• Comparison prompts: “agency or freelancer for Google Ads in Copenhagen”.
• Brand prompts: “what does [your company] do”, “is [your company] good for event production”. These test whether the model describes you accurately.

Write the prompts in each language your buyers use. A Danish prompt and its English equivalent can return quite different companies. Take the questions from your sales inbox and search data rather than inventing them, and freeze the set once it is agreed. Changing prompts each month makes trends impossible to read. If you need to add prompts, add them as a new group and report it separately for a few months.

Running the prompts and recording the answers

AI answers are not deterministic. The same prompt can name different companies on two runs a minute apart, and results shift with the model version, whether web search is used, the user’s location and any memory or personalisation on the account. A sound method reduces that noise rather than pretending it is not there.

1. Use clean sessions: logged out where possible, or a dedicated account with memory and personalisation switched off.
2. Note the assistant, the model or mode, the date and the location you ran from. Denmark-based results matter for a Danish business.
3. Run each prompt three to five times per assistant per measurement round, and record each run separately.
4. For each run, record whether you were mentioned, where in the answer, whether your own pages were cited, the tone of the description, and every competitor named.
5. Repeat monthly, on roughly the same dates.

Doing this by hand across five assistants and fifty prompts is a few hundred runs a month, which is tedious but manageable for a small team. A number of commercial trackers now automate it through the assistants’ interfaces or APIs. They save time, but check how they sample, because API responses do not always match what a user sees in the app.

Which numbers to calculate

From the recorded runs, a handful of metrics give a clear picture. Keep them simple enough that someone outside marketing understands them.

Break each metric down by assistant, by language and by prompt group. A company can be visible in Perplexity, which searches the web on every query, and almost absent in an assistant answering from training data alone. That split tells you whether the work is mostly about fresh, citable pages or about a wider and slower reputation across the web.

Traffic and crawler signals to put alongside it

Share of model voice measures presence. You also want to know whether presence turns into visits and enquiries. Three data sources help, each with gaps.

• GA4 referrals: visits from chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai show up as referral traffic. Create a custom channel group for them so they are not lost among other referrers. Many people read the answer and never click, and some clicks arrive with no referrer, so this undercounts real influence.
• Search Console: at the time of writing, clicks and impressions from Google’s AI Overviews and AI Mode are included in the ordinary web search figures without a separate filter. Rising impressions with falling clicks on question queries can be a sign that AI results are answering for you.
• Server logs: crawler user agents such as OAI-SearchBot, ChatGPT-User, PerplexityBot and ClaudeBot show which AI systems fetch your pages and how often. Copilot relies largely on Bing’s index, so Bingbot activity matters too.

Add one simple question to your enquiry form: “How did you hear about us?”, with an AI assistant option. Self-reported answers are imperfect, but they catch the buyers who were influenced by an answer and then searched for your name directly.

Reporting it so people act on it

A monthly report should show the trend, not just the latest figure. Plot share of model voice per assistant over at least three months before drawing conclusions, since month-to-month swings of several percentage points can be pure noise with a sample of a few hundred runs.

Pair every number with an explanation and an action. If a competitor is named in most answers about event sound, look at what the cited sources say about them, and what is missing about you. If brand prompts produce wrong facts, such as an old address or a service you stopped offering, correct those facts on your own site and on the third-party pages the assistants cite. Our search performance reporting puts AI visibility next to rankings, traffic and enquiries in one view, so the board sees one picture of search rather than two competing ones.

The actions themselves tend to fall into a few groups. Where your pages are cited but your company is described vaguely, rewrite the opening of those pages so each one states plainly what you do, for whom and where; this is the core of our citation-ready pages work. Where you are never cited at all, the gap is usually off-site: missing directory listings, no reviews that mention the service, or no independent page that names you alongside the companies the assistants already trust. Where the problem is one assistant only, check whether its crawler can reach the site.

Be honest about what the metric cannot do. It is a sample of answers to questions you chose, not a census of everything people ask. Used consistently, it still shows clearly whether your business is becoming part of the answer, and where to work next.