If you run a $200K/month Amazon brand and you made a creative decision this quarter based on your conversion rate, I need you to sit with an uncomfortable idea: that conversion rate is now an average of two completely different shoppers, and Amazon won't split them for you.
One shopper typed a keyword, scanned a grid of thumbnails, and clicked yours because the hero image won the glance. The other never saw a grid. They asked Alexa for Shopping a question, got an AI answer assembled from your attributes and reviews, and arrived (or didn't) on a path where your hero image was almost irrelevant. Your Seller Central reports stack both into one "sessions" number and one "conversion rate," and you've been optimizing against the blend. That's the real story this week — not a flashy launch, a measurement problem that just got too big to ignore.
What actually happened
Prime Day 2026 (June 23–26) produced a roughly 103% year-over-year jump in GenAI-referred clicks to retail, per Adobe Analytics' event tracking — the clearest signal yet that AI-mediated discovery has crossed from novelty into a material slice of traffic. Industry account reviews now put Alexa for Shopping (the assistant that absorbed Rufus on May 13) at 15–20% of mobile shopper queries. Meanwhile, as analytics shops like Nova Data have confirmed, Seller Central provides no native split between AI-assistant traffic and keyword-search traffic — "all sessions appear aggregated in standard reports."
So: a fifth of your demand now flows through a path that behaves differently, and your own dashboard reports it as if nothing changed.
Why most brand owners will read this wrong
The dumb take: "Alexa for Shopping is consumer fluff. Doesn't change my reports, doesn't change my job."
The real signal: your reports didn't change, but what they're measuring did. When two populations with different behavior get averaged into one number, the average stops describing either of them — and it moves for reasons you can't see. Your blended CVR can drop a full point this quarter not because your listing got worse, but because the mix shifted toward AI-path sessions that convert on a different mechanism. You'll go re-shoot a hero image to fix a problem that isn't in the hero image. That's not a hypothetical. That's the default failure mode of optimizing against blended data, and it's about to bite the operators who pride themselves on being data-driven.
The trap is sharpest for exactly the brands that do this right. The gut-feel seller never looked at CVR anyway. The disciplined operator who makes creative calls off conversion data is the one who'll be confidently wrong, because their input just got contaminated and nobody sent a memo.
What changes for someone running $200K/month
Three concrete shifts.
1. Your blended CVR loses diagnostic power. Say you're at 12% blended. If 18% of your sessions are AI-path and those convert at, say, 9% while your keyword-grid path converts at 13%, your "12%" is a weighted blur. Move the mix five points and the headline number moves with it — same listing, same price, same images. You can no longer read a CVR change as a listing change. That's a real loss of a tool you've leaned on for years.
2. Creative tests get noisier. Run an A/B on your hero image and a chunk of your traffic — the AI-path chunk — barely engages the variable you're testing. That dilutes your lift and lengthens time-to-significance. A 6% true lift on grid traffic reads as a 5% lift across blended sessions, and now you're not sure it cleared the bar. You end up under-shipping good creative because the measurement got fuzzier.
3. The thing the AI path actually rewards is invisible on your dashboard. The AI assistant assembles answers from structured attributes and review text. Whether you're in the consideration set of an AI answer is the new top-of-funnel for that 15–20% — and there's no column for it in Seller Central. You're flying blind on the fastest-growing slice of your demand, judging it with a number built for the old slice.
None of this means CVR collapses or ACOS explodes. The platform mechanics are stable. The damage is subtler and more expensive: you'll make confident, wrong decisions from a number that quietly stopped meaning what it used to.
What I'd do this week if I were them
-
Stop reading blended CVR as a listing health score. Until Amazon splits the traffic, treat a CVR move as "something changed — could be listing, could be mix" and force yourself to check a second source before you act. One bad re-shoot decision costs more than this habit.
-
Instrument the AI path manually. Once a week, run your top 10 ASINs' real buying questions through Alexa for Shopping (and ChatGPT and Gemini, since discovery is multi-surface now). Are you in the answer? Cited? Compared favorably? That manual check is the only "report" you've got on the path Amazon won't measure for you. Log it. Watch the trend.
-
Feed the path that's growing. The AI answer is built from attributes and reviews, so fill every narrow attribute field and make sure your top reviews actually name the use cases buyers ask about. This isn't a hero-image job — it's a structured-data and review-language job, and it's the cheapest lever on the fastest-growing traffic.
-
Quarantine your creative tests where you can. Lean on tools that segment, look at placement-level and new-to-brand cuts, and weight your read toward the traffic your variable actually touches. Don't kill a creative test on a blended number that's diluted by sessions the test never reached.
-
Re-baseline now. Write down today's blended CVR per top ASIN and your manual AI-visibility check. When the number drifts next quarter, you'll have a reference point to ask "did the listing change, or did the mix change?" — the question that saves you from the wrong fix.
What I'd ignore
- The "AI is killing Amazon search" takes. Keyword-grid is still 80–85% of discovery. It's not dying; it's sharing the room. Anyone telling you to abandon hero-image and keyword discipline for "AI optimization" is selling you a service and a panic.
- The exact percentage debates. 15% or 20%, mobile vs. total — it doesn't change your move. It's big enough to contaminate your blended number and that's the only fact that matters operationally.
- Anyone selling "Rufus attribution dashboards" as if the data exists. It largely doesn't yet, natively. Be skeptical of tools claiming a clean AI-traffic split out of Seller Central — confirm what's measured versus modeled before you pay for a number.
- The Prime Day record-headline coverage. The +103% GenAI clicks stat matters as a trend marker, not as a reason to chase the event. The durable lesson isn't "AI drove sales on Prime Day" — it's "your everyday reports now average two shoppers and you have to manage that on purpose."
The platforms will eventually give us a clean split; they always do, a year after we needed it. Until then, the operators who win aren't the ones with the best AI tooling. They're the ones who noticed their favorite number quietly stopped meaning what they thought — and stopped trusting it blind.