If AI engines now answer the question before anyone clicks, then clicks and impressions miss most of what is happening. Brands need a different set of numbers to know whether they are being found, described accurately and recommended.
This guide sets out the KPIs we use in GEO Compass audits, and the benchmarks from the GEO Compass dataset that put them in context.
From old KPIs to new ones
| Old KPI | New KPI for AI search |
|---|---|
| Share of voice | Share of model: how often AI answers name you |
| Search ranking | Answer position: where you appear within the answer |
| Media impressions | Answer sentiment: how you are framed |
| Message pull-through | Answer accuracy: whether the facts are right |
| Backlinks | Citation share: how often your sites, and the sites you influence, are cited |
| Clicks and downloads | AI referrals and conversions: visits and actions that come from AI engines |
None of the old metrics disappear. Search still drives traffic. But they no longer tell you what an AI engine says to the customer who never clicks.
1. Share of model
The share of answers, across a defined set of questions, that name your brand. It is the headline AI visibility number.
Measure it separately for:
- Branded and unbranded questions. Being named when someone asks about you is easy. Being named when they ask about the category is the real test. In the benchmark, the focus brand is named in 27% of unbranded answers on average.
- Each engine. Gemini, Claude and Google AI Mode name focus brands in about 35% of unbranded answers; Perplexity in 18%. See how each engine compares.
- Each market. Focus brands are named in more than half of unbranded answers in Indonesia and Malaysia, and under a fifth in the UK.
GEO Compass benchmark. All markets: 27%.
2. Answer position
Where you appear in the answer when you are named. An answer that lists five options and puts you fifth is worth less than one that leads with you. Track average position for each question set, and how often you are named first.
3. Answer sentiment
How the sentences about your brand are framed: positive, neutral or negative. Across the benchmark, 5% of answers that discuss brands frame them negatively. That rises to 10% for brand trust and reputation questions and 15% for health and medical questions, against under 3% for recommendation questions. If your category sits in trust or health territory, watch sentiment closely.
4. Answer accuracy
Whether the facts are right. Prices, product details, leadership, locations, policies and history. AI engines repeat outdated or wrong information confidently. Accuracy is checked against a fact sheet you control, and every error is traced to the source page that caused it.
5. Citation share
How often your own sites, and the third-party pages you influence, are cited. Two numbers matter:
- Own-site citation rate. In the benchmark, a brand's own site is cited in 73% of branded answers with sources, but only 12% of unbranded ones.
- Source mix. Which types of sites shape the answers in your category: earned media, reviews, social, reference, government, marketplaces. This tells you where to invest. Brand websites supply 31% of citations overall; earned media 20%.
6. AI referrals and conversions
Visits and conversions from AI engines. In your analytics, set up a channel group for referrals from chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, claude.ai and similar domains. Volumes are still small for most sites, but they are growing, and these visitors often arrive further along in their decision.
Getting the measurement right
- Use a fixed question set. Pick 50–200 questions real customers ask, covering branded and unbranded intents, the themes that matter (recommendations, comparisons, price, trust) and each market and language.
- Run them repeatedly. AI answers vary between runs. Daily or weekly collection, reported monthly, smooths out the noise.
- Record the model version. Engines change models without notice. In the benchmark, one model update changed the share of citations going to earned media by 12 points within the same studies.
- Benchmark against your category. Use the GEO Compass explorer to see what is typical for your industry, engine and market before setting targets.
Turning metrics into action
Each metric points to a lever:
- Low share of model in unbranded questions → win the third-party sources engines cite for the category.
- Poor position → strengthen comparison content and reviews.
- Negative sentiment or poor accuracy → find and fix the cited sources.
- Low own-site citation → restructure your pages so the facts are extractable.
See the GEO checklist for the practical steps.
Frequently asked questions
- What is share of model?
- Share of model is the share of AI answers, for a defined set of questions, that name your brand. It is the AI-search equivalent of share of voice. It should be measured per engine, market and question type, because the averages hide large differences.
- What is a good AI visibility score?
- It depends on the category and question type. In the GEO Compass benchmark, the brand a question set was built around is named in 27% of answers to category (unbranded) questions on average, and its own website is cited in 12% of those answers that use sources.
- How often should AI visibility be measured?
- Continuously, with monthly reporting. AI answers vary from run to run and models change frequently, so a single snapshot is unreliable. Track a fixed set of questions daily or weekly and report trends monthly.
Figures marked as GEO Compass data come from the anonymised GEO Compass benchmark of AI answers collected by We AI & Data Advisors. Related: AI search in Australia · AI search in Singapore · AI search in Malaysia · AI search in Indonesia · AI search in India · AI search in Europe