The method, published before the numbers

The index covers Canadian businesses in categories where an AI assistant is plausibly consulted before a purchase: professional services, home services, healthcare-adjacent consumer services, software, and retail with a physical footprint.

Selection is from public sources only, and the list gets published alongside the first results so the frame can be audited. Brands are not invited, cannot apply, and cannot pay to be included.

Prompts are written as a buyer would phrase them, not as a keyword. That means full questions with the follow-up included, because the follow-up is usually where a recommendation actually gets made.

Each category gets a fixed panel. The panel is written before any run happens and then frozen, and every subsequent edit is logged with the date and the reason. Silently adding prompts a brand has started winning is the most common way this kind of index gets quietly corrupted.

ChatGPT, Claude, Gemini, Perplexity and Google's AI surfaces, each recorded separately rather than averaged into one score. They draw on different sources and update at different rates, so a blended number would hide the only useful information.

Runs happen from a clean, logged-out state with no memory or personalisation, on a fixed schedule, with the raw response stored. Personalised sessions produce answers specific to one account, which is interesting and not measurable.

Four measures, kept separate because they move independently. Presence: whether the brand is named at all. Share of voice: how often it is named against the competitors named instead. Accuracy: whether what the engine says about it is factually correct. Framing: whether it is the recommendation or the alternative mentioned in passing.

Accuracy is scored against the brand's own public information. Where a claim cannot be checked against a public source, it is recorded as unverifiable rather than scored as wrong.

Answer engines are non-deterministic. The same prompt returns different answers between runs, which is why single results are meaningless and why the design depends on repeated runs against a frozen panel.

The engines change without notice and do not publish their selection logic. A shift in a result may reflect a brand's work, a model update, or a change in the retrieval layer, and this method cannot always distinguish between them. Where a jump correlates with a known model release, it gets flagged rather than claimed.

Geography is inferred from the prompt wording rather than from real geolocation, since runs are not distributed across Canadian cities. Results should be read as what an engine says about a Canadian query, not as what a user in Halifax specifically sees.

LumiRank sells the discipline being measured, which is a conflict worth stating plainly. Client brands will be marked as such in any published result, and the underlying run data will be available so the scoring can be checked independently.