An AI visibility baseline should let a colleague reproduce the count and explain what changed between reviews. Start with a stable set of customer questions and explicit counting rules. A percentage without those details cannot tell the team whether its brand improved, the sample changed or some answers failed to arrive.
Define the audience and keep a stable question set
Choose one market and a recognisable buying decision. A small packaging supplier might track questions about low-volume custom orders for European consumer brands. Include discovery, comparison and purchase constraints. Separate questions that name your brand from those that do not: they answer different business questions and should not silently contribute to one discovery score.
Give each question an ID, its exact wording, intended audience and date added. Preserve a core set across reviews. Keep new-market or experimental questions in a separate group until you decide to revise the baseline. Near-identical rewrites should not outnumber the distinct decisions the sample is meant to represent.
The calculation below uses 20 open category questions asked twice in one AI search environment. All 40 scheduled responses arrive. This deliberately small, invented dataset demonstrates the arithmetic; it is not a recommended universal sample size or an estimate of the market.

Record the conditions behind each response
Use a fresh conversation for each scheduled response. Record the platform, available model or product identifier, search setting, language, region, time and question ID. Preserve the answer and the cited URLs. If the platform does not expose a setting, mark it unknown rather than assuming it stayed fixed. A saved screenshot can assist review, but it should not replace the full question and answer.
| Record field | What to retain |
|---|---|
| Question and run | Stable question ID, exact wording, run ID and timestamp. |
| Environment | Platform, visible model, search setting, language and region. |
| Collection status | Returned answer, refusal or technical failure; link retries to the original run. |
| Observed evidence | Answer text, matched brand name, recommendation wording and exact source URLs. |
Define each metric before calculating it
For this example, presence means that an eligible answer names the brand at least once. Count the answer once even if it repeats the name. Match known aliases to the same brand, and inspect ambiguous names against the context and linked site. Keep recommendation status separate: an answer can mention a brand only to say it is unsuitable.
- Brand mention rate = eligible answers mentioning the brand / all eligible answers. Twelve mentions across 40 eligible answers gives 30%. This is answer-level presence, not the percentage of buyers who saw the brand.
- Own-site citation rate = eligible answers citing at least one page on your site / all eligible answers. Eight of 40 gives 20%. Several links in one answer still count once for this measure. Mention and citation are different observations.
- Question coverage = fully observed questions with at least one brand mention / all questions with the required completed runs. If the 12 positive answers came from six fully observed questions, coverage is 6/20, or 30%. Exclude incomplete questions from both parts of this calculation and disclose their count. Always show the repeat count.
These are explicit working definitions, not a claim that every tool uses them. A share-of-voice metric may use all tracked brands’ mentions as its denominator. Preserve the tool’s documented definition when comparing reports. Do not compare rates from different denominators as if they measured the same outcome.
Also report collection completeness. A timeout is not an answer without a brand mention. Define how refusals are handled before inspecting the results, keep the policy stable, and disclose exclusions. If missing runs cluster in one question type, recover them or limit the comparison; changing the denominator alone does not remove that bias.
Compare the same questions before adding the new ones
Suppose the next review repeats the 20 core questions twice and also adds 10 new questions, each asked twice. Every planned response arrives in both periods. The example below shows why an unchanged headline score can hide an improvement on the original set.
| Question group | First period | Second period | What the comparison says |
|---|---|---|---|
| Original 20 questions | 12/40 answers; 30% | 16/40 answers; 40% | A 10-percentage-point rise on the matched set. |
| 10 newly added questions | Not tracked | 2/20 answers; 10% | An initial reading, with no earlier baseline. |
| All tracked questions | 12/40 answers; 30% | 18/60 answers; 30% | The scope changed; this is not a like-for-like trend. |
Turn the baseline into a reviewable decision
A useful report states the scope before the conclusion: “Across the same 20 questions and 40 completed responses per period, mentions rose from 12 to 16. The new group starts at 2 of 20 responses.” Then inspect which questions changed and whether the brand was recommended accurately. Do not claim that a page edit caused the increase from these counts alone.
Compare platforms separately before combining them. If you need a combined figure, declare the weighting and keep it stable. A platform with more scheduled runs would otherwise dominate the average. Likewise, changing language, geography or brand-matching rules should start a labelled comparison rather than silently rewriting the earlier baseline.
In AskScope, use the question-level answers and citation records to investigate movements in the overview. Keep a short review note identifying the affected buyer need, the proposed page change and the next comparable check. Measure identifiable visits and enquiries separately. The baseline describes observed AI answers; it does not measure every customer’s exposure or prove revenue impact.


