Measure it only when a specific business decision is waiting on the answer, and skip it when no decision changes either way.
That sounds obvious until you notice how much measurement in this category starts from what can be tracked rather than from what needs deciding. The result is a dashboard full of figures nobody was ever going to act on, bought with a budget that had to be defended to somebody.
Three studies published in October 2026 made that defence considerably harder, and also considerably more confused.
What the studies actually found
Marie Haynes examined twelve months of GA4 data across 100 sites, more than 145 million sessions. ChatGPT sessions came to 0.62 percent of the volume Google organic delivered, and all assistants together to 0.73 percent.
SEO Works examined 95 GA4 properties from September 2025 to August 2026. AI referral sessions came to 72,095 against 2,825,594 organic, which is 2.6 percent of organic volume.
Kevin Indig took a different approach, looking across twelve sites for hidden effects rather than counting referrals. He could not detect AI-driven direct traffic to long-tail pages, and found no reliable relationship between AI visibility and branded search or direct traffic.
All three point the same direction on identifiable volume. None of them agrees with the others on magnitude. Haynes and SEO Works are roughly three and a half times apart on what is close to the same quantity.
A number several times wide cannot carry a budget
If a plan to fund AI visibility work is viable at 2.6 percent of organic volume and collapses at 0.7 percent, the plan was never resting on evidence. It was resting on whichever study got read first.
This is not a criticism of the researchers. Both published their methods. The gap comes from real differences in period, volume thresholds, and how AI referrals were identified at all. Google’s own AI Assistant channel grouping only exists from June 2026, so any analysis spanning the twelve months before it had to match referring domains by hand or by pattern.
What it means practically is that the number is not the decision input people want it to be.
The two questions people run together
There are two separate things a company might want to know, and they have very different price tags.
The first is an acquisition question. How much attributable business arrived from AI assistants last quarter. This belongs next to paid search, organic and outbound, measured the same way, and if the identifiable share is low then treating it as a primary acquisition channel is hard to justify on current evidence. Low channel share on its own does not settle whether a given investment pays off, since that depends on cost, margin and what the work consists of. But it does settle whether this is where your pipeline comes from.
The second is a correctness question. Whether ChatGPT, Claude, Gemini or Perplexity is currently describing a requirement that matters to a buyer incorrectly. This is bounded work. It needs a small number of real buyer questions and somebody reading what comes back.
Conflating the two produces both failure modes in this category. A programme gets sold on the correctness worry and then judged on the acquisition number, which it was never going to move. Or the acquisition number gets checked first, comes back low, and the correctness question never gets asked.
The four fields that decide it
Before paying to measure AI visibility, name the decision the result could change. Four fields, one page.
The buyer question. Not a keyword and not a topic. An actual sentence a prospect would type or say while evaluating something like you. “Can this handle multi-entity reporting for an agency managing twelve clients?” is a buyer question. “AI reporting software” is not.
The decision it informs. What you would do differently depending on the answer. Build a dedicated page, or leave the documentation as it is. If nothing changes either way, you have found a question not worth measuring.
The evidence that would change the decision. Named in advance. What would have to appear in an answer, how often, across which assistants, before you acted. Two of four assistants returning the same wrong thing on the same question is a threshold. Deciding the threshold after seeing the data is how noise becomes strategy.
The stop or recheck condition. When you walk away, and when you look again. If the answers are correct on the first pass, stop and look again in a quarter. A check with no end condition is not a check.
An empty list is a real result
The most useful outcome of running those four fields is sometimes discovering there is nothing here worth watching this quarter.
That is a legitimate finding and it takes minutes to reach. It is also the finding that almost nobody selling in this category will hand you, which is worth factoring into whose framework you adopt.
A question that survives all four fields is worth a bounded look. A question that fails any of them is worth leaving alone.
Frequently asked questions
Does low AI referral traffic mean AI visibility does not matter?
It means AI assistants are not currently a major source of attributable traffic in the datasets reviewed. It does not establish whether an incorrect answer about your product costs you business, which referral data cannot measure either way.
Why do the AI traffic studies disagree with each other?
They used different time periods, different eligibility thresholds and different methods of identifying AI referrals. Google’s AI Assistant channel grouping only exists from June 2026, so earlier data had to be matched by referring domain.
What is the smallest useful AI visibility check?
One real buyer question, asked across ChatGPT, Claude, Gemini and Perplexity, with the answers read rather than scored. If the answers are correct, you are finished for the quarter.
Should conversion rate from AI referrals change my budget?
Be careful with it. Haynes reports ChatGPT referrals converting at 13.8 percent against Google organic at 3.02 percent, but the conversion comparison rests on the 55 of her 100 sites that had tracked GA4 conversions, and SEO Works reports a much smaller gap of 1.18 times pooled organic.
How often should an AI visibility check be repeated?
Set the interval before you run the first one. A quarter is reasonable for answers that came back correct, and thirty days is reasonable for rechecking after a source has been revised.
What does Axis Suite actually record?
Answer observations with the question, the engine, the date and the evidence context attached, so a later reading can be compared to an earlier one under known conditions.
ABOUT AXIS SUITE
Axis Suite by TrendAxis is the independent intelligence layer that explains what AI believes about your brand, why it believes it, and what decision that belief ultimately drives. See the recorded observations and the method behind them at axissuite.ai and in the Proof Center.