How Geovium measures AI visibility

AI answers vary between scans, assistants and countries, so a single answer tells you little. This article describes, step by step, how Geovium turns many individual answers into metrics you can follow over time. The definitions are the same ones the product uses; our methodology page has a shorter summary.

1. Start from the questions your customers ask

Each workspace tracks a set of questions that real customers might ask an AI assistant about your category: recommendations, comparisons, "best X for Y" questions and local searches. Questions are grouped into topics and carry an importance level (high, normal or low), so the ones that matter most to your business can count more. You also choose the competitors to track and, when location matters, the markets (country, region or city) the questions are about.

2. Ask every assistant, on a schedule

Geovium sends each tracked question to each assistant it covers: ChatGPT, Gemini, Claude, Perplexity and Grok. Questions go through the providers' APIs, so answers can differ from what a signed-in user sees in a consumer app with personal history or settings. Checks repeat on the schedule you choose, daily or weekly; because the same questions are asked again and again, changes show up as trends rather than one-off observations.

3. Read every answer the same way

Each answer is read into the same structure. For your brand, Geovium records:

• whether you are mentioned at all;
• whether you are mentioned early, in the first paragraph;
• whether your own website is named or linked;
• your position when the answer is a list;
• the tone of what is said about you;
• whether you are presented as trustworthy.

It also records which tracked competitors are mentioned and in what order, which sources the answer cites, and whether the answer is in the same language as the question.

4. Count only complete answers

Some checks do not produce a usable answer: the provider is busy or unreachable, the answer is cut off, the assistant declines, or the response is empty. These are left out of scoring instead of being counted as zero, so an outage never looks like a drop in visibility. Two separate metrics keep them visible:

• Coverage: the share of questions that received a complete answer.
• Error rate: the share of checks where an assistant could not answer.

5. Score each answer from 0 to 100

Each complete answer gets a visibility score built from the signals above:

Position, tone and trust only count when you are actually mentioned, and the total is capped at 100. An answer that names neither your brand nor your website scores 0.

Competitors are scored with the same formula (mention, early mention, website, list position), so the numbers are comparable. Tone and trust are measured for your brand only.

6. Roll answers up into comparable scores

• Assistant score: the average score of that assistant's complete answers in the period.
• Overall visibility score: a weighted average of the assistant scores. Weights are set per workspace and are recalculated over the assistants that actually returned complete answers, so an assistant whose checks all failed is left out rather than pulling the score toward zero.
• Weighted visibility score: the same calculation with question importance applied: high ×1.3, normal ×1.0, low ×0.7.

7. Look beyond the score

The score is a summary. Other metrics explain what is behind it:

• Visibility: the share of complete answers that mention you.
• Share of mentions: when an answer mentions you or a tracked competitor, how often it is you.
• Link rate: the share of answers that mention you and also name or link your website.
• Source link rate and official source: how often answers cite any source, and how often they cite your official website.
• Position stability: how much your place in answer lists moves between answers.
• Language mismatch: answers written in a different language than the question.
• Sentiment: the tone of answers that mention you.

8. Compare like with like

Every change is measured against the immediately preceding period of the same length, with the same filters. If the previous period has no complete answers, no change is shown rather than a misleading one. Results are grouped by when the scan finished, and scans that are still in progress or that failed are never reported.

Limitations

No measurement of AI answers is perfect, and we would rather be clear about what can move the numbers:

• Providers update their models, which can shift answer patterns overnight.
• The intent behind a question can drift over time.
• Results vary by location, freshness of information and exact phrasing.
• Tracked questions are a sample of what people ask, not a record of every real conversation.

This is why we recommend reading trends across several periods instead of reacting to a single scan.