Methodology
An explainable score, published in full
If a number tells you what to do next, you deserve to know how it was produced. Here is the whole model.
Explainable
Every point is traceable to a component, and every component to the responses behind it.
Versioned
Weights change only with a new model version, so historical scores stay comparable.
Evidence-led
Scores derive from recorded assistant responses, not opinion or estimated traffic.
Honest about limits
Assistants are probabilistic. We sample, we disclose, and we never promise rankings.
Score components
Model v1.0AI mention rate
The share of your tracked queries where your business was named in the AI assistant's answer. 40 of 50 queries mentioning you gives a component score of 80.
Average mention prominence
For each answer that mentions you, we score position (first named business scores highest), the amount of description given, and whether you are recommended or only listed. Those are averaged across all mentions.
Competitor share of voice
Your mentions divided by all business mentions across your tracked queries, rescaled so that an even split with your tracked competitors sits at the midpoint.
Website and entity clarity
Checks for structured data, a consistent business name, clear service and location pages, and whether your site states plainly who you serve and where.
Authority and backlinks
Volume and quality of independent sources referencing your business, weighted towards sources in your own sector and region.
Citation consistency
The proportion of directory and data-aggregator records where your name, address and phone number match your canonical details exactly.
Reviews and reputation
Blended review volume, average rating and recency across the review platforms most commonly surfaced in AI answers for your sector.
Content coverage
The proportion of your tracked query topics that have a substantive page on your website addressing that specific question.
How the method works
1. How prompts are selected
We start from the way customers actually ask for what you sell, not from keywords. For each business we generate questions across four intents: discovery (“best X near me”), comparison (“X vs Y”), qualification (“who does X for Z”) and problem-led (“how do I fix X”), covering each service and each location you have added.
The draft set is shown to you in the Query Library before testing. You can add, edit, prioritise or remove any question, so the score always reflects the questions you agreed matter. Changing the set changes the score, which is why the set is stored with every report.
2. Which AI systems are tested
Questions are submitted to leading consumer AI assistants, currently covering OpenAI (ChatGPT), Google (Gemini) and Anthropic (Claude) model families, with Perplexity included on plans that cover it. The exact platforms queried are recorded on each run.
Provider coverage evolves as models are released and retired. Because provider behaviour is outside our control, we record the platform and date of every observation rather than assuming continuity between runs.
3. How scores are calculated
Each recorded response is parsed for whether your business is named, where it appears in the answer, how it is described, and which other businesses appear alongside it. Those observations feed the weighted components above, which combine into a single benchmark out of 100.
Scoring is deterministic: the same observations always produce the same score. Language models record and describe; they do not decide the number.
4. Confidence levels
Every question is sampled repeatedly rather than asked once. Confidence reflects how consistent the outcome was across those samples: high where the same result appears in nearly every sample, medium where it appears in most, and low where results are volatile or the sample was small.
Low-confidence findings are labelled as such in the report and should be treated as directional. We would rather show you an honest confidence level than a falsely precise number.
5. Testing frequency
A one-off audit is a single snapshot taken at the timestamp shown. Intelligence membership re-tests the same approved question set on a recurring monthly cycle, so movement between runs is comparable and attributable to the same method.
Re-runs can also be triggered manually within fair-use limits, for example after making a recommended change.
6. Versioning
Weights and calculation rules are version controlled. The current scoring model is v1.0, published July 2026. Reviewed and version-controlled at each release. Historic versions are retained so past reports remain reproducible.
When a version changes we state it clearly, and reports produced under an earlier version continue to display the version they were scored with, so historical trends are never silently rewritten.
What every report records
Transparency is the product. Each audit carries the full audit trail needed to reproduce, challenge or re-run it.
Questions tested
The full approved question set, grouped by customer intent.
Platforms tested
Which AI assistants were queried on this run.
Testing date
The timestamp of every recorded observation.
Score calculation
Component weights and the resulting benchmark score.
Confidence level
How consistent the result was across repeated sampling.
Model version
The scoring model applied to the run (currently v1.0).
What we do not claim
- Assistant answers vary between runs, users and dates. We sample repeatedly and report the observed pattern, not a single answer.
- A score measures visibility across the questions you approved. Change the question set and the score changes with it.
- Visora does not guarantee placement, mentions or inclusion in any assistant, and does not guarantee traffic, enquiries or revenue.
- We have no access to the ranking or selection systems used by any AI provider. Where we explain a result, we report likely contributing factors, not confirmed causes.
- Competitor figures reflect appearance in your question set during our testing, not overall market share or business quality.
Independent benchmark
Visora provides a proprietary AI Visibility benchmark based on observed testing and published methodology. It is not an official ranking or score issued by OpenAI, Google, Microsoft, Anthropic, Perplexity or any other AI provider.
AI responses change over time
AI-generated responses change over time. Results vary depending on model version, prompt wording, date, location, user context, platform updates and the information available. Testing represents a snapshot taken at the time shown in the report.
See the method applied
The sample report shows the same components scored end to end on a demonstration business.