Australian AI Search Visibility Index framework
Define the market and the observation unit
An index needs a declared category, Australian market boundary, eligible brand set and collection window. A brand's inclusion criteria should be fixed before the answers are observed.
The observation unit is one answer to one frozen question under recorded conditions. The answer, not a later summary produced by another model, is the evidence.
Build questions around real commercial decisions
The prompt set should cover category discovery, problems, comparisons, use cases, eligibility, risk and brand verification. It should not be rewritten after collection to make a preferred company look stronger.
Branded questions test factual accuracy and reputation. Non-branded questions test category discoverability and recommendation. They answer different questions and should not be merged.
Keep mentions, citations and recommendations separate
A mention means the brand appears in the answer. A citation means a visible source link or attribution points to the brand or a page about it. A recommendation means the answer positively selects or shortlists the brand for the stated need.
One answer can contain more than one state. A citation can support a negative statement. A mention can be factually wrong. Reporting the states separately keeps those differences visible.
Score factual accuracy independently
Check material claims about the organization, service, product, people, location, eligibility, price and policy against an inspectable public source. Record accurate, inaccurate, unsupported and unavailable rather than forcing every statement into true or false.
Preserve answer variance instead of hiding it
Generated answers can change between runs. The framework therefore retains repeated observations and reports the distribution rather than presenting one answer as a permanent rank.
A later edition must use the same eligibility, questions and classification rules or label the break clearly. Method changes can improve the index, but they destroy a clean time comparison unless both versions are retained.
Report the useful metrics before any composite score
The first report should show prompt coverage, mention rate, citation rate, recommendation rate, factual-error rate, source-domain distribution and the number of repeated observations. A composite index can be added only when its weights and missing-data treatment are public.
No visibility metric proves revenue. Where referral visits, qualified enquiries or sales can be observed, report them separately and preserve the attribution limit.
Turn each gap into a source action
An incorrect official fact becomes a source correction. A missing category answer becomes a page or entity question. A competitor citation becomes a source-quality comparison. An unsupported recommendation becomes a proof and corroboration question. The index is useful only when it identifies the next verifiable change.
What this framework does not claim
It does not claim permanent visibility, complete coverage of every private or personalized answer, statistical confidence that has not been calculated, platform preference, a causal ranking factor or a revenue result.
It also does not publish an Australian brand leaderboard yet. Any future edition must publish its eligible set, collection window, sample, definitions, weights and limitations beside the results.