Short answer
Measure GEO with four metrics: mention rate (answers that name you), recommendation rate (answers that list you as a candidate), citation rate (answers that cite your website) and AI-referred enquiries. The first three come from a fixed question set tested at intervals; the fourth comes from source tracking. AI answers vary, so call it sampling, not statistics.
The four metrics, with formulas
Fix the definitions first, so every month is comparable with the last.
| Metric | Definition | Formula | What it tells you |
|---|---|---|---|
| Mention rate | The answer to an unbranded buyer question names your brand | Answers that name you ÷ valid answers | Whether you are in the room when buyers ask about your category |
| Recommendation rate | The answer lists you as a recommended or compared supplier | Answers that list you ÷ valid answers | Whether you are treated as a candidate, not just named; record the position too |
| Citation rate | Your website appears among the sources the answer cites | Answers citing your site ÷ valid answers | Whether the AI is relying on your own material |
| AI-referred enquiries | Enquiries whose source is confirmed as an AI assistant | Enquiries in the period, and their share of all enquiries | Whether there is a business result |
A few notes:
- Valid answers are those that returned without an error or interruption; failed ones stay out of the denominator.
- Mention rate counts unbranded questions only. If you ask "tell me about Company X", the AI will almost always name it, and mixing those in inflates the number.
- Record the three AI metrics per platform and then average them. Platforms retrieve from different sources, and an average can hide a platform where you are at zero.
- Our AI visibility audit reports mention rate, citation rate and average position; the definitions are in how the audit works.
How to design the question set
The question set decides whether the test means anything. Four rules:
- Use the target market’s language and the way buyers phrase things. If US buyers ask in English, test in English, and include specifications, certifications, minimum order quantity, lead time and application, not just the category name.
- Cover different intents: supplier recommendations, comparisons, specification and selection, certification and compliance, purchasing terms. Write several of each.
- Mostly unbranded questions, plus a few branded ones to see whether the AI describes you correctly. 20 to 30 questions is a workable start, grouped by buyer role (engineer, purchaser, distributor).
- Never reword once fixed. If the wording changes every month, the numbers cannot be compared. New questions are tracked separately and kept out of the old denominator.
Example phrasing (replace the brackets with your product):
| Intent | Example |
|---|---|
| Supplier recommendation | Who are reliable manufacturers of [product] in China that export to the US? |
| Comparison | What is the difference between [process A] and [process B] for [application]? |
| Specification | Which [product] meets [standard or certification] for [use case]? |
| Certification | Which suppliers of [product] have [certification]? |
| Purchasing terms | Who supplies [product] with low minimum order quantity and short lead time? |
How to test: platforms, region, repeats, records
- Platforms: ChatGPT (with search on), Gemini and Perplexity, plus Google AI Overviews if you like. Record each separately.
- Region and language: use the target market’s language and, where you can, a network in that market. Use a logged-out or fresh account with personalised memory off, and start a new conversation every time so earlier turns do not colour the answer.
- Repeats: ask each question at least 3 times on each platform and use the proportion, not a single result. More repeats means less noise.
- Records: date, platform and mode, region, the full question, the full answer, cited links and where your brand appears. Keep the raw answers so someone else can check them.
The record format looks like this (a format example only, not real audit data):
| Date | Platform | Region and language | Question | Mentioned | Position | Cited source |
|---|---|---|---|---|---|---|
| YYYY-MM-DD | Platform A | US / English | (full question) | Yes / No | #N / not listed | (cited URL) |
How to track AI-referred enquiries
There is no single exact number; you assemble it from several signals.
- Track the source in GA4: look at session sources in the acquisition report and create a custom channel group that puts chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, claude.ai and similar into "AI assistants". Links in ChatGPT search answers often carry utm_source=chatgpt.com, which you can also filter on.
- Add a source question to your forms: "How did you find us?", with an option such as "AI assistant (ChatGPT, Gemini, Perplexity and so on)" and a free-text field.
- Tag it in the CRM: record the source channel when the enquiry arrives, and have sales confirm it with the buyer in the first conversation. The confirmed answer wins.
- Watch branded search: look at impressions and clicks for queries containing your brand name in Search Console. A buyer who sees your name in an AI answer may go on to search for it. It is a supporting metric and cannot stand alone.
Keep in mind that many AI referrals arrive with no source information and are logged as direct traffic, so the AI traffic in GA4 is a floor. Form answers and sales confirmation fill the gap. The visitor analytics we set up for clients lists AI assistants as their own source channel and recognises those parameters.
Limits of this kind of testing
- It is sampling, not statistics. AI answers change with model version, account, region and time, and repeated runs of the same question differ. We do not have large samples and do not run significance tests, so we never say "statistically significant".
- Read trends, not single moves. Use the same question set and method for several months in a row, and draw no conclusion from changes within normal noise.
- The question set is biased. People wrote the questions, and they may miss real buying scenarios. Update them from buyer feedback.
- A mention is not an enquiry. Mentions and citations are process metrics; the end point is AI-referred enquiries and orders.
How to read the results
Re-test monthly, review the trend every three months, and compare with competitors. Did mention rate in the priority question clusters rise above the baseline? Are the cited sources your site or third parties? Has the share of AI-referred enquiries moved? If one platform stays at zero, go back to the technical foundation; see how AI picks suppliers.
How each metric is accepted in a pilot is covered in what a 90-day GEO pilot delivers. To see your own baseline first, run a free audit. Our export GEO service also re-tests monthly on these definitions. For the concept itself, read the glossary entry on GEO.
Frequently asked questions
- What is the difference between mention, recommendation and citation rate?
- Mention rate asks whether your name appears in the answer. Recommendation rate asks whether you are listed as a candidate supplier, and records the position. Citation rate asks whether your website is among the pages the answer cites. Being mentioned does not mean being recommended, and being recommended does not mean your own site is cited.
- How many times should each question be asked?
- At least 3 times per platform, and read the proportion. More repeats reduce noise, but it is still sampling and cannot be called statistically significant.
- GA4 shows no AI traffic. Does that mean there is none?
- Not necessarily. Some AI referrals carry no source and are logged as direct. Add a source question on your forms and have sales confirm it, and treat the GA4 figure as a floor.
- How often should we re-test?
- Monthly, with the same question set and method. Judge the trend over three months or more, and avoid rewording questions in between.
Further reading
- Service: Export GEO
- Service: Overseas website
- Glossary: What is export GEO (Generative Engine Optimisation)?
- Glossary: What are AI crawlers, and should you allow GPTBot, ClaudeBot and PerplexityBot?
- Lesson: 3.1 Overview of the six-step loop: diagnosis, modelling, content, sources, reputation and monitoring
- Lesson: 10.2 Case study: stating the basis of Aikang Medical’s 47 and 95 points objectively
See how AI describes you today
The free AI visibility audit asks ChatGPT, Perplexity and Gemini questions as an overseas buyer would, and checks your website’s technical foundation.