“Which One Should I Get?” Is Turning Into “Just Order It”
Asking ChatGPT or Claude which product to buy stopped being a story a while ago. And everybody does it. But that’s still the “help me compare” stage — you get the shortlist, and you take it from there.
I do it too, usually late at night, usually for something I don’t actually need (nobody needs a fourth pair of running shoes).
But the next step isn’t comparing. It’s choosing, and then buying — and the moment a model starts doing that, the definition of “the buyer” quietly changes underneath you, which means the thing you’re optimizing stops being the ad or the hero image and starts being the doorway into the purchase itself.
So how close is any of this, really?
A few groups have put numbers on it, and the numbers are worth sitting with.
The Forecasts Are Loud. The Present Is Quieter.
PwC’s write-up of ShopTalk 2026 collects the forecasts in one place, and they don’t agree with each other at all — which is itself the useful part.
eMarketer puts 8.8% of ecommerce transactions through AI agents by 2029. Merkle goes much further: up to 50% involving agents or answer engines (a looser bar — “involved in” isn’t the same as “bought it for you”). Strategy& lands somewhere in between, expecting up to 15% of European ecommerce spend, with adoption running as much as four times faster than traditional ecommerce managed.
So take the low number and this is still a channel you’d put someone on; take the high one and it’s the channel. (When serious forecasters land five-fold apart, what they’re mostly telling you is that the thing hasn’t settled yet.)
But the consensus right now isn’t “the machine buys everything”. It’s a lot more modest than that. 60% of consumers who get a recommendation go and research it independently anyway — so the models are still working mostly as discovery and research tools, not as checkout buttons.
Which means have taken the front half of the purchase rather than the whole thing. Still a lot, though — the front half is where the shortlist gets made, and you can’t win a comparison you were never entered into.
The Real Shift: The Buyer And The User Come Apart
Stefano Puntoni at Wharton makes a point that sounds obvious for about four seconds and then stops being obvious. Marketing has always quietly assumed the person who buys and the person who uses are the same person. In an agent world, they’re not.
Call it the . A human still uses the thing. But if a model holds the entrance — the comparing, the shortlisting — then you owe an explanation to two very different audiences at once, and they don’t want the same explanation.
Copy built purely to move somebody emotionally tends to lose a machine comparison, because there’s nothing in it to compare; copy built purely for the machine reads like a spec sheet and loses the person. (Anyone who’s written a product page knows how fast those two briefs start pulling apart.)
And I think that tension is the real competitive question of the next couple of years. It isn’t one you settle by picking a side.
Swap The Model, And The Recommendation Swaps Too
Here’s the part that made me sit up. Kamruzzaman, Nguyen and Kim compared brand recommendations across GPT-4o, Llama-3-8B, Gemma-7B and Mistral-7B at EMNLP 2024 (one of the big natural-language-processing conferences). Same job, four models.
The pattern was blunt. Asked about high-income countries, the models recommended luxury brands 88–100% of the time; asked about low-income countries, non-luxury brands 84–98% of the time. Gemma-7B recommended luxury brands 100% of the time across every category for high-income countries.
Same question, different machine, different shelf.
Then there’s a second bias sitting on top of that one. Laurito and colleagues, writing in PNAS in 2025, documented what they call : GPT-4 preferred AI-written product descriptions 89% of the time, where human raters preferred them only 36% of the time.
So the agent era isn’t one shelf with one ranking. It’s more like a hiring loop where the interviewer keeps changing and the criteria change along with them — which is a genuinely different structure from the one-Google-to-rule-them-all decade, and (maybe I’m wrong here) I don’t think the old instincts port over cleanly.
Conclusion: Build The Shelf For The Model And The Shelf For The Person
If agents are moving into the front half of the purchase, a brand needs more than an emotional story. It needs information a model can actually compare. The mistake I’d worry about is blending the two — one page that hedges between a feeling and a spec and lands neither.
So here’s the one thing I’d start this month: run the same buying question past ChatGPT, Claude and Gemini every week, and log which brands come back and what reasons get attached to them. (Keep the wording of the question identical each time; the whole point is that the models differ, not the prompts.) Do that for four weeks and the per-model tilt stops being a theory and turns into a chart you can hand someone. That’s your read on the machine-facing shelf, and it’s cheap.
Then the human-facing one. On the product page, put the primary facts where both audiences can reach them — price, specs, return conditions, terms of use — stated plainly rather than implied by a lifestyle shot. A person skims it; a model can quote it. Same page, two readers.
And where you point that effort first depends on what you sell, because the payoff isn’t even across a catalog: I dug into which categories AI referrals actually work on over here.
Sources
- Puntoni, S., “AI Is Upending Marketing on Two Fronts”, Harvard Business Review, 2026-02-23, hbr.org
- PwC, “AI Commerce Is Here: Key Takeaways from ShopTalk 2026”, 2026-03, pwc.com
- Kamruzzaman, M., Nguyen, H.M. & Kim, G.L., “Global is Good, Local is Bad?: Understanding Brand Bias in LLMs”, EMNLP 2024, aclanthology.org
- Laurito, W. et al., “AI-AI bias: Large language models favor communications generated by large language models”, PNAS, Vol. 122(31), 2025, pnas.org
