Document
Methodology
Question set
v1 · 5 Jul 2026
Extraction
v2 · 22 Jul 2026
Market
Australia
Assistants
ChatGPT · Perplexity · Gemini · Claude

The method in full

How the Southlume Index measures AI visibility

The Southlume Index measures who AI assistants recommend across Australian categories – which businesses get named when real buyers ask real buying questions, and how consistently. This page is the method in full.

One rule governs everything we publish: every number carries a date, a question-set version, and a market. If a figure can’t tell you when it was measured and how, we don’t publish it.

fig. m1 · suppose we ask 20 timesillustrative of method
0 · never1 · identicalthe reading we publish

Ask once and you get a point. Ask twenty times and a range forms – the grey band. We publish the middle, and we print the band beside it so you can see how much to trust it.

download · png · svg

1the method

The method, in one table

QuestionsFixed buyer questions, plain and non-branded, per category – “best mortgage broker in Sydney”, asked the way a customer would
Question setVersioned like software – currently v1, 5 July 2026. If a question changes, the version changes; nothing is reworded quietly
AssistantsChatGPT · Perplexity · Gemini · Claude (as at v1)
MarketAustralia at launch – the method reads any market the assistants answer
MeasuresMention rate – how often a business is named. Citation rate – how often its site is linked as a source. Never blended into one score
SamplingTwo stages – a screening pass across a category, then a deeper pass where precision matters most. Sample sizes publish with each edition

question set v1 · 5 Jul 2026 · market Australia

A business can be named often and cited never – we’ve measured the gap, which is why they stay separate.

2signal vs noise

How we tell a real change from noise

Ask an AI the same question four times and the recommendations can change substantially – we’ve measured that too. Any single reading is partly the model’s own randomness, so every figure we publish carries its stability band: the range randomness alone would produce. A change only counts as movement when it clears the band. A mention rate that rises inside its band is wobble, not a win – and we say so.

More repeat runs narrow the band; that’s what the two-stage sampling buys. On our charts, orange means one thing only: moved beyond its stability band.

3what we count

What counts as a business

An entity is a business, not a string of text. Assistants emphasise headings and half-sentences alongside the names they recommend, and write one business several ways in the same answer – so counting text overstates how many were named and understates how often one was held.

CountedA named business or practitioner – the thing a customer could book
Not countedProse, headings, addresses, opening hours, dates, bare place names. Each exclusion publishes with its reason
Variant spellingsMerged – but only where the names share a distinctive word, so two clinics sharing a generic or geographic term stay separate
IdentityA business website domain where the assistant links one (~3% here), the name everywhere else – name-matched counts carry that caveat

entity extraction v2 · 22 Jul 2026

Extraction is versioned like the question set – this is v2, 22 July 2026 – and counts published under v1 are carried beside the current ones in every dataset, not replaced.

4the record

Dated, versioned, and published in full

Every reading is stamped – date, question-set version, extraction version, market – and editions are numbered, so any figure can be cited precisely.

what publishesthe commitment
Movements that clear the bandReported as real, with the band drawn on the bar
Movements inside the bandReported as wobble – named as such
Quarters where nothing movedReported as exactly that. A boring result is still a result
Our own visibilityLogged from a baseline of zero – the method applies to us first
5limits

What the numbers do not say

Zero means: not found on these questions, on this date, in this market – never “invisible everywhere, always”. The questions are a sample – a disciplined one, but a sample.

Prompt checks are a sample; referral analytics are the ground truth of actual AI-driven visibility. A zero here is ‘not found on these prompts’, never ‘invisible’.

FAQ

Frequently asked questions

How does Southlume measure AI visibility?
By putting a fixed, versioned set of buyer questions to ChatGPT, Perplexity, Gemini and Claude, counting who gets named and who gets cited, and publishing the results as rates with stability bands, per Australian category.
What is a stability band?
The range an AI’s own randomness produces when the same question is asked repeatedly. A change that stays inside the band is noise; a change that clears it is real movement. Every figure we publish carries one.
How often does the Index publish?
Editions are dated and numbered rather than promised to a calendar. Edition 1 is in measurement now – leave an email and you’ll have it the day it publishes.
Which AI assistants does Southlume check?
ChatGPT, Perplexity, Gemini and Claude, as at question-set version 1. If the set changes, the version changes with it.
What’s the difference between mention rate and citation rate?
Mention rate counts how often a business is named in AI answers; citation rate counts how often its website is linked as a source. They move independently and we never combine them.
Can a score be zero?
Yes. Zero means “not found on these questions, on this date” – never “invisible”. The distinction is the caveat above, and it’s permanent.

The method, in use

Edition 1 is in measurement now.

notify me · edition 1