AI conversion lift: read the fine print

The AI conversion lift is the gap between how often visitors arriving from an AI assistant buy and how often everyone else does, and it describes the composition of a measured cohort at least as much as it describes the quality of the traffic.

That distinction matters right now because the number is being quoted in budget meetings as though it were a stable property of AI referrals. It is not stable. It has moved further, and faster, than any behavioral explanation can carry on its own.

#The swing

Adobe Analytics tracks this across a very large sample, and the trajectory is public. In March 2025, AI traffic was 38% less likely to result in a purchase than non-AI traffic. Twelve months later, in March 2026, the same measurement showed AI-driven visits converting 42% better than non-AI traffic. By July 2026 the gap had widened again, with Adobe reporting AI traffic converting 60% higher than non-AI visits and generating 53% more revenue per visit.

Read as a behavior story, that says shoppers referred by AI assistants went from noticeably worse than average to substantially better than average inside a year. Shopper behavior does not usually move like that. Three other things did.

#The measured cohort is not the whole cohort

A session only enters the AI bucket if something survives the trip that names it as AI: a referrer header, or a tracking parameter in the URL. Everything else lands in Direct alongside bookmarks and typed URLs. Google's own definition of its GA4 AI Assistant channel makes the matching conditions explicit: the medium is exactly ai-assistant, or the referrer matches Google's maintained list of AI assistants. There is no rule that recovers an untagged, referrer-less click.

So the AI cohort in any of these reports is the attributable slice, not the population. That slice is skewed toward desktop web sessions clicking tagged links, and away from in-app taps and copy-paste navigation. Whatever the invisible slice converts at, it is not in the numerator.

The problem is not that the slice is partial. Plenty of useful metrics are partial. The problem is that the slice got bigger over exactly the period being compared, because assistants added link tagging. When coverage and performance improve at the same time, an outside observer cannot separate them. A March 2025 measurement and a March 2026 measurement carry the same label but do not describe the same set of sessions.

#The baseline is a blend

The comparison is against non-AI traffic, which Search Engine Land notes covers channels such as paid search and email marketing. That is an average across everything else a retailer does, weighted by however much volume each channel happens to have that month.

This is not a like-for-like intent match. A visitor who arrives after asking an assistant to compare three products has already done the comparison step. A blended baseline includes top-of-funnel display clicks, broad-match paid search, and promotional email to lapsed buyers. Some of the measured lift is the AI cohort sitting further down the funnel than the average of everything it is being compared to. Adobe's engagement figures point the same way: AI-referred visitors spend 48% longer on site and browse 13% more pages per visit. Those are the numbers of people who already decided to shop.

#The population is growing faster than anything else in the report

AI traffic to US retail sites rose 393% year over year in the first quarter of 2026. A cohort growing at that rate is almost entirely new people every period. The March 2025 group was small and dominated by early adopters testing a novelty. The 2026 group includes ordinary buyers who now use an assistant the way they used to use a search box.

Comparing a conversion rate across a population that turned over completely is a legitimate thing to report and a bad thing to extrapolate from. It tells you the channel matured. It does not tell you the channel is worth a fixed multiple of anything.

#What the number is actually good for

Direction, not magnitude.

The most credible part of the Adobe data is not the size of the lift. It is that July 2026 marked the eleventh consecutive month of AI traffic outperforming non-AI traffic. A sign that holds for eleven months on one consistent methodology is a real finding. The specific multiple is the fragile part, because it depends on the mix, the baseline, and the attribution coverage in that particular month.

Practically:

• Do not import a published multiple into your own business case. Measure your own AI cohort against your own baseline, in your own analytics, over your own months.
• Compare against a matched baseline where you can. Non-brand organic search is a closer intent match than a blended average that includes email.
• Segment by landing page. If half your AI sessions land on a product or pricing page rather than the homepage, that alone explains part of the lift and tells you which pages are carrying the channel.
• Track the trend line, not the ratio. If your own lift holds positive across many months, that is the finding. The month-to-month value will move for reasons that have nothing to do with your content.
• Expect your measured number to be conservative on volume and optimistic on rate, because the sessions that go missing are the ones you cannot label.

#The finding under the headline

The same Adobe research contains something more directly actionable than the conversion figure, and it got a fraction of the coverage. Roughly a quarter of the content on retailer homepages has not been optimized for LLMs, and around 34% of product pages cannot be properly accessed by AI.

That is a supply-side problem you can fix this quarter, and it sits upstream of every conversion debate. A page an assistant cannot fetch or parse never enters the candidate set, so it never gets cited, so it never sends the high-intent visit whose conversion rate everyone is arguing about. Machine accessibility is the precondition. The lift is downstream of it.

If you have to choose where to spend attention, spend it on whether your pages can be retrieved and read, and treat the conversion multiple as a directional signal that the channel is worth the work.