How we measure AI visibility.
This is our AI visibility measurement protocol for generative engine optimization (GEO) and answer engine optimization (AEO) work, not a results report. It separates website readiness from observed mentions, recommendations and citations. No results are implied until observations have been collected and validated.
Discuss your AI visibilitySeparate readiness, mentions, recommendations and citations
Readiness is a website assessment. A mention is the business being named in an answer. A recommendation is an explicit suggestion to consider or choose the business. A citation is a source link supporting the answer; it can exist without a recommendation.
Record these separately. A higher readiness score does not establish visibility or cause a recommendation.
Control the question and product surface
Version the prompt panel before testing. Record the exact prompt, intended market, question intent, product surface, model when exposed, search setting, account/personalization controls, tester location and timestamp. Start fresh conversations and record any unavoidable differences.
Consumer ChatGPT observations and API tests are separate conditions. An API response is not evidence of what a consumer sees in ChatGPT. Do not pool products, settings or markets as if they were one test.
Repeat runs and keep market segments distinct
Repeat each question three times per controlled condition and retain every attempt. Compare Dubai, US and global questions separately, then compare question intent within those markets.
Our internal company benchmark keeps a 42-question library with a 30-question active panel: 18 Dubai, six US and six global questions. Three runs create 90 attempted observations per condition. This internal research panel is not a client package allowance.
- Monthly recheck of up to ten agreed prompts across three agreed answer platforms
Errors are not successful no-mention answers
Classify each attempt as successful or failed and record the error type. Successful answers that do not mention Iyara remain in the valid denominator. Failed or blocked runs are reported separately and excluded from mention, recommendation and citation rates.
Calculate rates as qualifying successful observations divided by all successful observations in the same condition and segment. Publish both counts. If there are no valid answers, show Not measured rather than 0%.
An unordered answer has no rank. Record list position only when there is an explicit ordered list; do not infer that citation order represents a recommendation ranking.
Archive evidence and report limitations
Keep the original answer, source links, screenshots or exports, timestamps and classification notes in a private evidence archive. Review ambiguous classifications and retain the reasoning. Redact personal information before any approved publication.
Model updates, search availability, personalization and randomness can change answers. Before/after changes are observations, not proof that a website edit caused them. Referral analytics and self-reported discovery help assess demand but cannot capture every AI-assisted journey.
Primary references
Platform documentation explains access and search behavior; it does not promise placement.
Measure first. Improve what the evidence supports.
We can apply this protocol to an agreed market and question set, then scope the website work the observations support.
30-minute scoping callDiscuss a measurement scope