I took 10 blank Amazon accounts and gave each one a persona — a pregnant mom, a vegan, a night-shift nurse, a 68-year-old gardener, a college kid cramming for exams — then asked Alexa for Shopping the same question: what’s the best pain killer for me?
I set up ten verified-blank Amazon accounts — no purchase history, no profile — and asked each the same thing: what's the best pain killer for me?
All ten blank sessions returned education-only answers. Zero product cards. With nothing to personalize on, Alexa explained pain-killer categories and recommended nothing.
I declared a persona in the chat — and all ten answers personalized. The pregnant mom got safety framing and Tylenol-only cards. The vegan got Genexa and Tiger Balm. The diabetic got cautions before products. The 68-year-old retiree got an easy-open arthritis cap.
A persona-flavored follow-up — leg-day soreness, night-shift back pain — changed all ten answers. The fitness enthusiast's branch flipped to whey protein and BCAAs. The retiree's moved to Voltaren and Salonpas. Said context moved it 10 out of 10; the silently seeded cart was never referenced in any core answer.
In these 40 sessions, the declared persona and the spoken follow-up moved every answer, and no core answer referenced the silently seeded cart. For CPG brands, the evidence says the AI shelf is won in the chat, at the moment of the declared need.
Ten branches, one per persona — scroll sideways through the whole tree.
Lifting &
running
Whey · BCAAs
Yard work &
lifting
Advil · Biofreeze
Night-shift
nursing
Framing, then 0 cards
Lifting
large dogs
Compare cards
Dorm &
exam stress
1 card
Pregnancy
Tylenol-only
Gardening
& arthritis
Voltaren · Salonpas
Type 2
diabetes
Cautions first
Vegan &
clean label
Genexa · Tiger Balm
Classroom
tension
Compare cards
Each persona lived on its own blank account and ran the same four-stage session: anchor, declaration, silent cart, follow-up. Nothing carried over between accounts.
The same pain-reliever question, no identity given. The baseline everything else is measured against.
The persona is stated in the chat — "I'm vegan" — and the question is asked again.
Unrelated products added, never mentioned. No core answer referenced the seeded cart — though five card lists did shift between stages without naming a seed, so this isn't a clean zero.
A persona-flavored symptom, said out loud — leg-day soreness, night-shift back pain, exam headaches.
The ten personas: a dad into yard work and DIY, a pregnant mom of three, a fitness enthusiast, a college student, a 68-year-old retiree who gardens, a dog owner, a night-shift nurse, a vegan, a type-2 diabetic, and an elementary teacher.
Each branch is one account's four stages, from the session transcripts.
With no declared identity, Alexa wouldn't name a single product across ten accounts. The same question, asked by nobody in particular, got nothing.
The moment each person said who they were, the answer changed — safety framing, clean-label filters, cautions. The chat is the targeting layer.
A spoken follow-up — leg-day soreness, night-shift back pain — changed every answer a second time. The chat re-targets live.
Optimize for the moment someone says "I'm pregnant," "I'm vegan," "my back hurts after night shifts" — that's where the recommendation is decided.
Three days before I ran this experiment, PacVue’s Tommy Burton — seven years at Amazon Ads on DSP and Amazon Marketing Cloud — told the AdTechGod podcast what brands are watching:
When you’re going in and typing in and having a conversation with Alexa, shopping — how many brands show up and which brands show up — that is a key challenge that a lot of people are watching.Tommy Burton, VP Global Partnerships, PacVue (ex-Amazon Ads) — AdTechGod Pod, September 15, 2026. The episode was sponsored by Amazon Ads.
You have Alexa for shopping on Amazon. That’s a conversational experience within the shopping journey. A lot of these things are really early, and that’s kind of where it’ll be really interesting to see how things start to progress.Tommy Burton, same episode.
I see it on Amazon with Rufus that they’re experimenting with, where you search something and then in the answer responses — oh, do you want this? Do you want that? It’s kind of like trying to figure out what your desires are, to personalize that experience to you.Melissa Burdick, founder and president of Pacvue — live from Groceryshop 2024, The CPG Guys, December 2024.