James Richardson.
James Richardson, Founder, Optimising
ChatGPT referrals to the Australian ecommerce brands we track grew by a median of 76.8 per cent in three months. AI-attributed revenue across the same client base sits at around 0.1 per cent of the total. Both numbers are true, both come out of our own tracking, and the gap between them is the measurement problem this whole industry keeps tripping over.
Scott Purcell argued in these pages at the end of June ("Marketing to the machines: good SEO is good GEO") that most GEO retainers are snake oil in a new bottle, and that the qualities which earn an AI citation are the ones that always earned a ranking. GEO, for anyone who has dodged the acronym so far, is generative engine optimisation: the business of getting brands into AI answers. On the fundamentals, he’s right. I’ve been making the same argument for eighteen years, most recently to anyone unlucky enough to be stuck next to me at an industry lunch.
But his proof of value is a publisher's proof. Man of Many's 1.3 million weekly AI-crawler requests and 38 per cent growth in AI-assistant referrals show that machines are reading and readers are following. What that data cannot show is the brand side of the transaction: what happens after an AI system recommends a brand and the customer never clicks anything at all. Answering that takes advertiser-side data, which is the data we happen to sit on.
I have argued publicly that AI visibility should be tested against branded search, direct traffic, conversions and revenue. Google and Bing now hand over presence data for free, so knowing you appear in AI answers is table stakes. The only thing left worth paying for is knowing what that presence is worth. So we built that test.
First, the obvious objection, because Purcell's piece already loaded it. Rand Fishkin says any tool that gives you a ranking position in AI is full of baloney, and SparkToro's data backs him: fewer than one in a hundred repeat runs of the same prompt return the same brand list. No argument from me, and it’s why we don’t measure positions at all. We track recommendation share of voice: the questions real buyers actually ask, weighted by how often they ask them, run repeatedly across four platforms over months, and scored on whether a brand is genuinely recommended rather than merely name-checked. One run in isolation tells you nothing. Hundreds of the same runs over months start to show a trend worth leaning on, and even then we treat the number as meaningless until commercial metrics move with it.
Two US studies explain why this connection is invisible in your analytics. Similarweb's panel research found that users who received a ChatGPT brand recommendation visited that brand at roughly 2.5 times the rate of a paired competitor over the following seven days, and 55.9 per cent of those visits arrived through search, not through an AI referral link. Profound linked more than two million real AI conversations to subsequent browsing and found only about 2.5 per cent of downstream visits carried an identifiable AI-referral parameter. The visit is measurable. The exposure that caused it is not. Neither study measured purchases, and association is not causation, but the mechanism is clear: the AI does the persuading, and your last-click report gives Google the credit.
Our own published Australian benchmarks show what that blindness looks like in practice. Across 24.8 million sessions and 115 businesses, our 61-business ecommerce subset recorded 41,335 AI sessions, 332 transactions and $48,705 in AI-attributed revenue. About 0.1 per cent of ecommerce revenue, converting at 2.9 per cent against 6.0 per cent for organic search. The study itself notes those totals are conservative, because many AI journeys hide where they came from and land in analytics as direct or organic. On the same client base this year, seven in ten brands grew or started receiving ChatGPT referrals inside three months, and nearly one in ten received their first ever ChatGPT visit.
Now put those two numbers side by side. Referrals climbing that fast beside a revenue line that small, while the US research says the pathway between them is mostly invisible, is not evidence AI visibility is worthless.
It is evidence the meter is broken, which is exactly why we built our own.
A peer-reviewed Marketing Science study adds caution before anyone overcorrects the other way. Across 973 ecommerce sites, traffic from AI assistants made up less than 0.2 per cent of all visits a year after ChatGPT launched, and ChatGPT itself accounted for around 90 per cent of it. That traffic converted below organic search but above paid social, and it performed notably better for complex purchases. Put plainly, AI traffic behaves differently depending on what you sell and how considered the purchase is.
So should you stop paying for GEO? Wrong conclusion. Stop paying for a visibility number with nothing attached to it, because on its own that’s just bad measurement. For a CMO, the useful test has three layers.
- Are we recommended for the questions real buyers ask?
- Is that recommendation followed by measurable branded demand?
- And does the movement reach orders, pipeline, revenue or gross profit?
Track all three, and keep repeating the prompts, because AI answers vary from one run to the next. Weight the prompt set by what customers actually ask, and strip out paid activity, promotions, PR and seasonality before AI gets any of the credit.
A visibility number on its own hasn’t shown you a commercial result. The misses matter as much as the wins: if visibility rose and nothing else moved, that finding belongs in the report, not the bin. Success stories alone make a sales deck, not a case.
AI visibility isn’t revenue. It gets you a shot at the consideration set, nothing more. Our job is to show when that exposure is followed by demand and money, and when it is not. If the number cannot be connected to either, it is still only a number.