How the measurement works

Method

Everything here is fixed before an edition runs, and published figures change only through a marked restatement

The questions

650 questions across three markets — NDIS, aged care and allied health. Half are phrased as a search query and half the way a person actually types, carrying the same buying intents, so the difference between the two can be measured rather than assumed. It is a large difference: asked conversationally, these systems name about a quarter fewer vendors.

Every question is Australian-scoped. The set is fixed for an edition. If we change it, that edition is not comparable to the one before it, and we say so on the page rather than quietly restating the trend.

Questions that name a vendor are asked but not scored. The “alternatives to X” and “X versus Y” kind are real buyer queries and we keep the answers, but a vendor named in the question is named back almost every time, and only a handful of vendors are ever asked about that way. Leaving them in would hand a score to whoever happens to appear in the wording. The published figures are computed without them.

We publish the composition but not the questions themselves. A published list would be a target: a vendor could write content aimed at those exact strings and move their score without becoming any easier for a real buyer to find. The full set can be inspected by an independent auditor of the index — not a vendor, and nobody acting for one — read-only, without a copy, under a non-disclosure agreement.

From asked to scored

The engines

Four, named on every edition: ChatGPT, Perplexity, Google AI Mode and Google AI Overviews.

The model is part of the instrument — changing one changes which vendors are named — so it is fixed for an edition and stated on the page. In the current edition ChatGPT runs on gpt-5.1 and Perplexity on sonar. The two Google surfaces offer no model selection; the answer is whatever Google served that day.

ChatGPT and Perplexity are asked through their own official APIs. The two Google surfaces are not, because Google publishes no API for them — they reach us through a disclosed third-party SERP-data provider. Every edition names which engine came from which, and why that distinction matters. We do not scrape any provider’s consumer app — that breaches their terms, and we would rather have a smaller measurement than an unusable one. What these surfaces return is not identical to what a person sees in the app, and we say so on every edition.

Not every question produces an answer on every engine. Google renders no AI Overview for most queries in this category — in the August 2026 edition that surface answered 298 of the 400 asked that month. A question that returns no answer is recorded as no surface, never as a score of zero, and every edition prints how many of each. It is the difference between “AI did not name you” and “AI did not answer”, and conflating them would flatter the leaders and punish everyone else.

The score

Each question is asked three times per engine. For every question-and-engine pair we record whether a vendor was named, and where it appeared.

What a change mark means, and what is never marked

Asking every question three times gives three independent readings of the same month. Their spread is what this instrument does when nothing has changed, and it differs by vendor: in the August 2026 edition one vendor named in 23.9% of answers carried four times the spread of another named in 25.2%. So the threshold is each vendor’s own, taken from that edition’s three passes — never one figure applied to everyone, which would hide a small vendor’s real move and print a large vendor’s wobble as news.

A month-over-month figure is a difference between two editions, and each side carries its own error, so the threshold is widened for that. We check it rather than assert it: the edition’s own repeat runs are shuffled into pairs that differ by nothing but chance, and the share the threshold wrongly calls movement is printed on the edition page. It is aimed at one in twenty.

Rank movement is never shown. A rank is a position among every other vendor, so it moves when they move and when the index adds a vendor — neither of which is this vendor being named more often. What is marked instead is the change in how often a vendor was named, which nobody else’s result can move.

Presence is the share of runs naming the vendor. Prominence weights rank — but rank among the vendors this index tracks, not position in the answer. A vendor named after two products we do not track still counts as first.

Question-and-engine pairs are averaged, not individual runs, so a question we happen to ask more often cannot outweigh one we ask less.

One consequence is worth stating plainly, because it looks like a bug and is not: adding a vendor to the registry moves the scores below it, even when nothing about AI has changed. So every edition is rescored against the current registry before any month-over-month figure is published, and where a prior edition is restated we report the registry effect and the real change separately.

A brand name appearing only inside a web address is counted as a source, not as a naming. Being linked and being recommended are different events, and an index that scores them the same is measuring search engine optimisation.

Two numbers we publish that most people would leave out

Answer stability. AI systems do not give the same answer twice. We publish the share of the time that a vendor named at all was named again on every repeat of the same question. In August 2026 that was 35%. Visibility in these systems is a probability, not a fixed position, and you deserve to know how firm the ground is before acting on a number. Where an edition did not sample enough runs to measure it, we print “not measured” rather than a figure.

Deflection rate. The share of answers that named no tracked vendor at all and pointed at a review directory instead. If that number climbs, AI is getting less useful for this category — including less useful for us.

Coverage

We track 108 vendors. Not “all major vendors” — 108, listed in full in every edition. Inclusion is decided by one published test: does the product contain something only a care provider would ever use? Ownership, size, and whether a vendor has ever been a customer play no part. If a vendor is missing, tell us and we will add it to the next edition.

Corrections

Raw answers are kept for three years, so any figure we have published can be traced back to the text it came from. If you think a figure is wrong, write to corrections@namedindex.com with the edition and the vendor. We will re-check it against the stored answers, correct it if we are wrong, and note the correction in the following edition. We do not quietly edit published editions.

What counts as being named

What this is not

A vendor can be excellent and invisible, or ordinary and everywhere. We measure one thing: how often AI systems name you when your buyers ask.

This is not a review site, and it is not a ranking of product quality. Nothing here is a recommendation to buy anything, and no position in this index should be read as a judgement about whether a product is any good. We have not evaluated a single one of them.

It also measures what AI answers, not what buyers do next. We make no claim about how often Australian buyers ask these systems, or what they do with the answer.

Disclosure

We sell AI-visibility services, so a reader can discount the index accordingly. Position in this index cannot be purchased. Clients are named in the methodology of each edition published during the engagement and for twelve months after it ends; where a genuine legal or contractual restriction prevents naming one, that edition gives the number withheld instead. No vendor was a client of ours in the August 2026 edition, and none is currently withheld.

The same tool produces the public index and the client reports, from the same stored answers, so a client cannot be shown a different number from the one you see. We do not guarantee placement in any AI system’s answers; no published research demonstrates that any technique reliably produces one.