Measuring ChatGPT · Perplexity · Gemini · Claude
Who AI recommends in Australia – measured.
When your customers ask ChatGPT, Gemini or Perplexity who to buy from, a short list gets named – and it changes minute to minute.Southlume measures who makes it: the same questions, repeated on a schedule, with the working shown.
The questions are fixed. The schedule is fixed. What moves is the answer.
23 businesses in four answers. More than half named once, never again.
names per answer · 11 · 15 · 4 · 9
One buying question, put to ChatGPT four times in one sitting, July 2026. Of the 23 businesses it named, 13 appeared once and never again. Only one business was named in all four answers – and ChatGPT wrote its name four different ways doing it. The recommendation survived; the name barely did.
Ask again three days later and the same assistant agreed with itself more than twice as often. Not even consistency is consistent. So every number we publish carries its range – the stability band, on every chart we make.
Read the study – including where we re-checked our own numbers →
The Southlume Index –
who AI recommends, category by category
Rates with ranges, dated, misses published beside wins – where you stand, per assistant, per category. Edition 1 is in measurement now – no ranking announces before it can stand behind itself.
That email doesn't look complete.
Measured on ourselves, first.
An independent measure, the way TV audiences are read by a ratings house rather than the networks – nobody can pay to appear in it, move up in it, or keep a bad quarter out of it. So the instrument points at us first. We publish our own reading from the day it turned on, starting where every brand starts:
log entry 000 · baseline