AI search visibility: what evidence connects it to commercial value?

Steve New

AI-referred traffic is beginning to look commercially interesting. Shopify's Q2 2026 data shows AI-referred sessions to its merchants growing 197% year on year, with orders roughly tripling. Once those visitors reached a product page, they converted about 80% better than visitors referred by organic search. Adobe found a similar pattern in US retail in July, where AI-referred traffic converted around 60% better than non-AI traffic.

It is tempting to turn numbers like those into a simple story: become more visible in AI answers, receive better traffic, make more sales. But the evidence between those steps is less tidy because two different problems can pull the apparent effect in opposite directions.

Referral analytics can miss journeys where somebody sees an AI recommendation and later reaches the business through search, direct traffic or another route. At the same time, the people who do click directly from an AI answer may already be unusually far through their decision, so their high conversion rate cannot simply be treated as the causal effect of AI.

AI can therefore look smaller in attribution data while the visible referral cohort makes its contribution look stronger than we have established.

If I want to know whether AI visibility is commercially valuable, that is the inference problem I need to solve.

AI referrals show only part of the journey

Shopify's data gives a useful clue as to why its AI-referred shoppers convert so well. More than half of AI-referred sessions start directly on a product page, compared with about 20% of organic-search sessions. The conversion advantage is particularly large in specification-led categories, where buyers have more comparison work to do before choosing. Shopify describes this as more of the decision happening before the click.

That explanation is plausible, but an 80% conversion difference does not mean AI created 80% more purchase intent.

Somebody who has already asked an AI system detailed questions, narrowed the alternatives and clicked a specific recommendation is not equivalent to somebody beginning with a broad search query. The higher conversion rate may partly reflect who reaches the site through that route and how far they had progressed beforehand.

That does not make the cohort less valuable. If identifiable AI referrals consistently produce strong customers, the business should care. It does mean that the conversion rate tells me about the customers arriving through the channel before it tells me how much additional intent the channel created.

Similarweb tracked US desktop journeys after people received ChatGPT brand recommendations across finance, travel and beauty. It excluded people who had recently visited the brand and those who had already named it in their prompt. Within seven days, users receiving a brand recommendation were 2.5 times more likely to visit that brand than an unrecommended competitor.

Most of those eventual visits did not look like AI traffic in conventional analytics. Search accounted for 55.9% of the visits Similarweb classified as AI-influenced, while only 8.8% arrived through an AI referral. Direct and other referral channels accounted for most of the rest.

The study is observational, so I would not read the 2.5x figure as a clean estimate of visits caused by the recommendation. It does show that an AI exposure can precede a later brand visit without the eventual session retaining an obvious connection to AI.

A buyer might see a recommendation in ChatGPT, close the conversation and search the company name that evening. Standard analytics records an organic visit, and the earlier AI exposure disappears from the referral trail.

Taken together, the Shopify and Similarweb findings create the more interesting problem. The directly identifiable referral cohort can be unusually selected, while a potentially important part of the wider AI-mediated journey is not identifiable as AI traffic at all.

There is no fixed “AI traffic quality”

Adobe's data is another reason not to turn the current conversion premiums into a property of the channel.

In May 2026, AI-referred traffic to US travel sites converted 28% worse than non-AI traffic. By July, the difference was about 1%. In US retail during July, meanwhile, AI-referred visitors converted around 60% better than non-AI traffic.

Shopify also finds its largest advantages in categories where buyers need to work through more specifications and comparisons before reaching the merchant.

The observable performance therefore varies substantially by category and over time. What the customer is trying to decide is one plausible reason, but the current evidence does not justify treating AI referrals as an intrinsically superior or inferior class of traffic.

There is also a limit to how far I would generalise the magnitudes. These are large commercial datasets, but they are not experiments designed to estimate one causal effect that should transfer to every market. Most of the evidence here comes from ecommerce, US retail and travel, or US desktop journeys in a small number of consumer categories.

I would carry the selection and attribution problem into other buying journeys much more readily than I would carry an 80% or 60% conversion uplift.

Start with the commercial claim

The evidence I need depends on what I am trying to say.

If the claim is simply that the business is becoming more visible in AI answers, I need a defensible way of measuring the relevant prompts, engines and market. If the claim is that AI platforms are sending traffic, referral sessions can answer much of the question.

The evidence burden changes once I want to say AI is influencing discovery beyond the visitors it refers directly. Referral analytics are no longer enough, so I might add customer self-attribution, assisted journeys, panel evidence, branded discovery or another way of observing what happens after the exposure.

If the customers attributed to AI appear commercially strong, I can examine their conversion, lead quality, order value and later customer behaviour. That tells me something useful about the cohort.

A stronger claim is that our work on AI visibility created additional commercial value. Now I need to know what happened because of the intervention that would not otherwise have happened, alongside what the work cost and whether the customers it produced were valuable.

These questions sit on the same path, but evidence for one does not automatically answer the next. Referral revenue may be useful without establishing incrementality, and a strong conversion rate among attributed customers can justify action without telling me how much of their apparent advantage AI created.

Imperfect metrics are unavoidable. The mistake is allowing the commercial claim to become stronger while the evidence underneath it stays the same.

Attribution still leaves the counterfactual unanswered

Suppose I can identify customers who encountered the business through AI somewhere in their journey and attribute revenue to them.

That is considerably better evidence than a mention count, but it still does not tell me what would have happened without the AI exposure or intervention.

Some of those customers may have found the company anyway. Existing demand can contribute to AI visibility just as AI visibility may contribute to later demand. A brand that is already prominent may appear frequently in AI answers, receive more branded searches and sell well for overlapping reasons.

This is a familiar problem in experimentation. Kohavi, Tang and Xu argue that useful short-term metrics need to be measurable while also having a credible relationship with the longer-term objective they are meant to represent. A metric can be highly responsive and easy to optimise while still being a poor guide to the result the business ultimately cares about.

AI visibility is vulnerable to the same failure. A company could increase mentions by covering more informational prompts, publishing more easily cited material or becoming visible for questions with little commercial relevance. The visibility score improves without necessarily producing more valuable demand.

That does not make visibility a vanity metric. It makes it a proxy whose relationship with the commercial result needs to be checked rather than assumed.

The evidence burden should match the decision

For a modest, reversible investment, I would not require an incrementality experiment.

If a business can identify customers attributed to AI who look commercially valuable, the relevant pages are inexpensive to improve and the work is easy to reverse, referral revenue and customer quality may be enough to justify the investment. Better attribution would increase confidence, but it may not change the decision.

I would set a higher bar before moving a large acquisition budget, building an expensive AI-content programme or making a strong claim that AI has become a superior growth channel.

At that point I want to know more about the journeys direct attribution misses, whether changes in visibility correspond with changes in valuable demand, how good the resulting customers are and what probably would have happened without the intervention.

The downstream economics matter as well. A high conversion rate can still produce mediocre results if the orders have lower margins, the leads are poor fits or the customers do not remain valuable after the initial transaction.

So I would still track mentions, citations and share of voice. I would track identifiable AI referrals and what those visitors do. Where the volume and decision justify it, I would add evidence for the journeys direct attribution misses and follow the customers far enough to see whether the apparent advantage survives.

I do not need one definitive “AI ROI” number before acting. I need to know what claim the evidence can support.

A visibility increase, an AI-influenced journey, a high-converting attributed cohort and incremental profit are not four names for the same result. I would rather know which one I can support than give all four the same number and call it attribution.

Sources