For most of 2025 the reasonable position on AI-referred retail traffic was that it was growing fast and converting badly, which is the signature of curiosity rather than intent. That is no longer the position.
Adobe reports AI-driven traffic to US retail sites up 393% year on year in the first quarter of 2026, following a 693% year-on-year increase over the 2025 holiday period. Growth at that rate would be interesting on its own. What changed is the quality.
In March 2025, AI-referred traffic converted 38% worse than non-AI traffic. In March 2026, it converted 42% better. Adobe also reports these visitors engaging 12% more, spending 48% longer on site, viewing 13% more pages, and generating 37% higher revenue per visit.
Separately, Salesforce estimated that AI agents influenced more than 20% of global online retail sales during the 2025 holiday season.
How AI-referred retail traffic performs against non-AI traffic
Adobe AnalyticsA reversal of 80 points in twelve months. Adobe also report engagement up 12%, time on site up 48%, pages per visit up 13% and revenue per visit up 37%.
View as table
| March 2025 | March 2026 | |
|---|---|---|
| Conversion rate, relative to non-AI traffic | -38% | 42% |
What that changes for product data
A visitor who arrives through an assistant has usually had a set of constraints applied on their behalf before they ever see your page. They are further down the decision than organic traffic, which is a plausible reason for the conversion reversal.
It also means the filtering happened somewhere you cannot see, against whatever attributes were available. If a constraint could not be evaluated against your product, your product was not in the set that got shown. There is no partial credit and no benefit of the doubt: an agent applying a filter treats a missing value as not qualifying, because that is the safe behaviour for it.
A shopper might read your description and infer the screen size. An agent asked to filter on it drops you and moves on.
The honest limit of this argument
We cannot tell you what share of agent traffic you are currently losing to unrecorded attributes, and neither can anybody else, because the filtering is not observable from your side. Anyone quoting you a recovery figure for this is guessing.
What is checkable is the input. You can establish which constraints buyers in your category apply, which of those your catalogue can answer, and how many of the gaps are recoverable from text you already publish. That is a measurement rather than a projection, and it is available without waiting for anything.