GEO Handbook, Part 6: Measure AI visibility and understand enquiries

Your business appears in an AI answer. A useful starting point. The business question follows: which customer questions bring you into view, is your service described correctly, and does that lead to suitable enquiries? Part 6 explains how to answer these questions with a manageable measurement plan, current market data, clearly defined metrics and a practical path from evidence to improvement.

By Kacper Ruta · GlasBox Studio, Malters/Lucerne · Research and source review: 20–21 October 2026.

01 — What the 2026 market data actually shows#

AI adoption has reached a meaningful scale. Published figures describe different things, however. Weekly users, monthly users and queries within a business sample cannot be added together or presented as market shares. The table therefore identifies each figure, its population and its date.

The Swiss study surveyed 1,590 adults across three language regions. Its announcement does not provide a separate response count for the supplementary AI question. The 42% figure therefore cannot establish search behaviour within a particular industry or demand for your service. [5]

Our interpretation: these figures justify including AI in a measurement plan. Customer questions and your own observations establish which platforms matter for your business. We keep the Claude and Perplexity usage samples within their stated context; they do not support an audience-size comparison with ChatGPT or Gemini.

02 — AI traffic: business value depends on context#

For July 2026, Adobe reports a conversion rate 60% higher for AI-referred visits than for non-AI visits on observed US retail websites. On travel websites, it was 1% lower. The report was published on 19 August 2026. [6]

These are observed differences between visitor groups, not an experiment establishing GEO’s causal effect. They do not reveal the absolute conversion rate of a Swiss service business. A 60% conversion rate is also a different quantity from a rate that is 60% higher. [6]

For your planning, examine your own path: who arrives, which page helps them, and which enquiries fit your service? Visibility and business outcomes each require their own evidence.

03 — Four layers: mention, source, visit and enquiry#

The layers can occur independently. An answer may link to your guide without naming your brand. A person may read a recommendation and call later. An AI system may retrieve your page without a person visiting your website. Combining these situations into one GEO score obscures where work is actually needed.

04 — Choose the right customer questions#

Unbranded questions test discovery: the prompt does not contain your company name. Branded questions test whether the AI describes your business correctly. Keep these groups separate. Adding your brand afterwards changes the test. The same applies to explicitly requesting source links: this can be a useful fixed protocol, but it must remain identical across comparable runs.

Use separate question groups for DE-CH and EN. Translating a prompt does not establish that English-speaking customers are seeking the same service. Prioritise proximity to a business decision and the evidence available: which question concerns a suitable customer shortly before making an enquiry?

05 — A measurement protocol you can repeat#

A manageable starting point could be 20 fixed questions, each repeated three times per round: 60 planned responses per surface and language. This is a practical workload, not a statistical quality threshold. Five surfaces would already mean 300 runs per language. Start with the environments that matter to your customers and choose a scope you can maintain.

A location named in the prompt does not establish the actual test location. A selected model is not necessarily the model used for every background search. Mark what the interface does not disclose as unknown. New conversations reduce the influence of previous messages, while account settings and personalisation can still matter.

Repeat the review on a fixed schedule and within comparable time windows. Retain every planned run, including unfavourable answers. A technical error does not count as an absent mention. For AI Overviews, also record whether an overview appeared at all. Its display rate and your presence within displayed answers are separate metrics.

06 — Keep ChatGPT, Claude, Gemini and Perplexity separate#

Anthropic documents web search and source links in Claude; in its new experience, search can happen automatically without a separate toggle. Record the behaviour you can observe. The OpenAI API, meanwhile, offers a separately configured web-search tool. API monitoring is its own measurement channel and can complement app observations. [11][13]

Our working rule: each surface receives its own time series. Changes to model, search mode or test tool are marked as events. An aggregate is useful only if its weighting is explicit and the individual results remain available.

07 — Metrics with an understandable denominator#

An inaccurate service promise remains a problem even when it sounds positive. Sentiment describes tone; quality review checks the claim. Record the specific sentence, its linked source and the correct information. Human spot checks remain useful when a tool performs the initial classification.

Share of voice is, at most, a share within your defined test set. One possible formula is your brand mentions divided by mentions of all predefined comparison brands, counting each brand once per answer. This is not market share. Changing comparison brands or questions changes the denominator.

08 — Read a time series: variation, percent and percentage points#

The following chart uses entirely fictional teaching data. Each round tests the same 20 questions three times on the same surface. All 60 runs per round are valid in this example. The purpose is to make the arithmetic visible; these are not GlasBox or client results.

From A to B, the mention rate rises by 10 percentage points. Relative to the starting value, this is about 33.3%: (40 − 30) ÷ 30. In C, it falls again. Publishing only B selects the most favourable observation. Showing every round makes stability discussable. A relative percentage increase is undefined when the starting rate is zero.

These 60 answers are not 60 randomly selected potential customers. Repetitions of the same question are related, and the question set was deliberately chosen. The results describe that set under those conditions. A small series establishes neither a population-wide share nor a reliable causal effect.

When runs fail, report planned and valid counts separately. Also compare the subset of fully comparable questions. Mark model changes, seasonality and website changes when they coincide. A before-and-after comparison documents a change; its cause requires separate investigation.

09 — What GA4, Search Console and Bing show today#

Documentation checked on 21 October 2026. GA4 includes the AI Assistant default channel for recognised traffic from AI assistants. AI Overviews and AI Mode are explicitly excluded and belong to Organic Search. For website visits, examine the session-scoped channel group alongside source/medium; origin information can be missing. [7]

Bing introduced AI Performance as a public preview on 10 February 2026. Its citation figures describe source presence in supported environments; they do not establish a position within every answer or a universal authority score. [9]

During setup, check an actual click and the resulting record. Document channel-rule changes. A break between an old referral report and a newer channel assignment can reflect a measurement change. Compare the same periods, dimensions and consent configuration.

10 — From a session to a qualified enquiry#

Define a suitable enquiry before reporting it. For a regional service business, this could mean a real project, a service you offer, a relevant location and a reachable contact. A button click, an opened form or an automated test message does not yet meet these criteria.

• Record each enquiry once and exclude tests, spam and duplicates.
• Retain identifiable technical origin where supported by your measurement setup and consent.
• Record a customer’s voluntary account of origin separately, such as “found you through ChatGPT”.
• Document qualification and any later sale in a CRM or protected enquiry register.
• Keep unknown origin as unknown; do not put contact details or message text into analytics events.

A visitor may first see an AI answer, later search for your name and only then enquire. Technical attribution and self-reported origin illuminate different parts of this path. Keep them as separate fields on the same lead so that one enquiry does not enter the total twice.

An initial monthly business report can answer three clear questions: how many suitable enquiries arrived, how many have an identifiable or customer-reported AI connection, and which became sales? Include unknown origins. Associate revenue only with a documented sale.

11 — Turn a finding into a specific improvement#

Specialisation makes the distinction visible: who does the work, for which problem, with what experience and within which limits? Creative work translates that knowledge into an understandable diagram, original photographs, a short film or a clear comparison. The format follows the customer’s question. A chart should explain something that prose alone makes harder to grasp.

The technical foundation stays concrete: accessible pages, useful internal links and structured data that matches visible content. Google requires no special AI files or additional schema types for AI Overviews and AI Mode. [8]

OpenAI uses OAI-SearchBot for search, while GPTBot serves a separate training purpose. Perplexity also distinguishes search crawling from user-initiated retrieval. A verified bot request in a server log establishes retrieval; a citation or human visit requires separate evidence. [10][12]

12 — A manageable starting plan#

• Set the scope: one service, relevant customer questions, languages and selected surfaces.
• Document the baseline: a fixed question set, repeated rounds, sources and factual errors.
• Choose one bottleneck: weigh business relevance, available evidence and implementation effort.
• Implement deliberately: improve expertise, presentation or technology; record dates and owners.
• Measure again and decide: repeat the questions and review website and enquiry outcomes separately.

A monthly report does not need a wall of metrics. One page with scope, trend, three material findings and the next tasks is a useful start. Decide in advance when to review the question set. New questions can be added while the existing comparison group remains a separate series.

13 — How the six parts fit together#

RAG remains a separate evaluation area: does the system retrieve the right approved documents, stay faithful to them and answer the question? Public GEO measurement asks whether external platforms find and visibly use your information. Both benefit from maintained sources, but their metrics should remain separate.

Sources, dates and method#

Research was conducted on 20–21 October 2026. The table identifies the verified publication date or data period; checking a source in October does not turn earlier observations into October data. Provider disclosures, usage samples, surveys and website analytics remain separate. We do not estimate undisclosed sample sizes or publication days.

The selection focuses on primary sources about usage and measurement. A newer product announcement is not automatically a new measurement of the same metric. Text, process diagrams and measurement recommendations are original GlasBox work. The series in section 8 is explicitly fictional.

14 — Glossary: terms and their place in the GEO Handbook#

These translations are working terms for this article. The definition in your measurement protocol remains decisive.