Public methodology
How we measure AI visibility without false certainty
AI answers vary. A reliable method preserves context, distinguishes mentions from citations and draws conclusions only after several valid measurements.
- Repeatable: The same prompt set and market context are used at several points in time.
- Domain-specific: Only a genuine source URL belonging to the client counts as a proven citation.
- Limited but honest: The result describes the measured sample, not every possible AI answer.
What do we measure?
We measure whether a brand is mentioned, whether a client page is used as a source and how often this happens within the fixed prompt set. These signals are stored separately.
- Brand mention
- Domain-specific source citation
- Cited page
- Prompt category and intent
- Model and measurement date
Quality rules for a valid citation
A source must belong to the same client, site, monitor and recent snapshot. Lookalike domains and a brand name without a client-linked source URL do not count as proven citations. A matching path is additional evidence and is not required when the domain is correct.
- Exact client domain or valid subdomain
- Snapshot within the technical 72-hour window
- Linked to the same client, site and monitor
- Preserve the source URL and measurement time
- No historical backfill presented as new evidence
Measurement limitations
Results vary with model updates, personalisation, location, language and source index. The measurement indicates trends within a fixed sample and is not a complete count of all answers.
- No guarantee of repetition for every user
- No direct revenue attribution
- Models may present sources differently
- A baseline measurement does not prove that a brand appears nowhere
The measurement unit and context of each AI measurement
The basic unit is one valid execution of one versioned prompt on one named system, with one recorded measurement time and one market context. Each execution receives a unique run ID. A change to prompt text, language, location or system selection creates a new prompt set version.
For each run, we save the exact prompt, the system, the displayed model or product name when available, the date, time zone, language, available location and login context, the answer, the displayed source URLs and any error or exclusion reason. Personalisation that is not visible remains an explicit limitation.
- Unique run, prompt, monitor and site IDs
- Exact prompt text and prompt set version
- System, measurement time, time zone, language and location context
- Answer, source URLs and redirect destination
- Classification, error status and exclusion reason
Classify mentions, citations and errors separately
A brand mention means the agreed brand name or a verified variant appears in the answer. An own-domain citation requires at least one visible source URL on the verified domain or a subdomain allowed in advance. A target-URL citation is more specific and refers to the canonical page chosen in advance.
Timeouts, missing answers, hidden sources, lookalike domains and untraceable redirects are not silently counted as zero mentions. They receive an invalid or uncertain status. Each ratio therefore shows its numerator, denominator and the number of invalid or missing runs.
- Brand mention
- Own-domain citation
- Target-URL citation
- External entity citation
- Invalid or uncertain
Report raw counts, ratios and sample changes
A trend comparison uses the same prompt set version and, as far as possible, the same measurement context. A single baseline is a snapshot. With a small sample, even two measurement rounds may be insufficient for a stable conclusion, so raw counts remain visible alongside ratios.
Changes to the system, prompt set or coverage are disclosed in the comparison. We do not report an average AI position when a system does not show a stable, comparable ranking. The existing 72-hour rule is described solely as a technical matching window, not as a universal validity period for a citation.
- Brand mention ratio with numerator and denominator
- Own-citation ratio with numerator and denominator
- Target-URL ratio with numerator and denominator
- Prompt coverage and number of invalid runs
- Changelog for prompt set and classification rules
Frequently asked questions
Is a brand mention the same as a citation?
No. A citation includes a client-linked source reference; a brand name without such a source is only a mention.
How many measurements are needed?
At least several unique measurements are needed to establish a trend. The exact frequency depends on prompt volume and market volatility.
Can AI visibility be linked to revenue?
Not automatically. Referral traffic and website actions can be measured in addition, but a citation alone is not a sale.
Why are invalid runs not counted as zero mentions?
Because an error, timeout or missing answer proves nothing about visibility. Invalid runs are counted separately so coverage remains visible.
Is the first source in an answer position one?
Not automatically. Source order and presentation differ by system. SEOJuice reports a position only when the system shows a meaningful, reproducible order and the definition has been established in advance.
How do you handle personalisation?
Login status, language and available location context are recorded. Personalisation that is not visible cannot be ruled out completely and remains a limitation.
Can I compare two different prompt sets?
Only with clear segmentation. Changing prompts changes the sample; do not present a difference as though only brand visibility had changed.
Primary sources
By SEOJuice editorial team, Linktool B.V.. Content last updated: .
View your baseline