Santosh Pandey, co-founder and CEO of Christchurch agency Ridiculous Digital, on the AI shopping agents now researching and buying on behalf of New Zealand customers - why much of what lifts human conversion can quietly lower your chances of being chosen, and what to fix first.
Two shoppers land on the same product or service page. The first is a person. They notice the "only 3 left in stock", the countdown timer, the price with a line through it, the reviews. Some of it nudges them toward buy. The second shopper is an AI agent - ChatGPT, Perplexity, or one of Amazon's - sent to research or purchase on someone's behalf. And almost everything you built for the first shopper, the second one cannot see. Worse, a few of those tactics count against you.
That second shopper is still outnumbered by humans in New Zealand - but they're walking amongst us, and they're multiplying fast. Adobe's numbers from June's Prime Day event back this up: AI-referred traffic to US retail sites nearly doubled year on year - and, for the first time, it converted better than every other channel, 40% ahead of paid search, email and social. Twelve months earlier, it converted 23% worse. Getting your store ready is easier while it's early and awkward once it isn't. So it pays to understand what the second shopper reads, because it is not at all what a decade of conversion optimisation taught us to worry about.
The shopper you can't charm
At Ridiculous, we've been told we're a pretty charming lot. But AI agents are a tough crowd. I know you have spent years on conversion rate optimisation ops and this will break your heart: the tactics that move people don't reliably move an agent, and the more capable the agent, the worse they perform.
In a 2026 Harvard Business Review simulation of AI shoppers, researchers tested eight common promotional tactics across four leading AI models and more than 16,000 simulated shopping rounds. Scarcity badges, countdown timers, strike-through pricing, bundle framing - all unreliable, and several reduced the chance of selection. The detail that should stop you: the more advanced reasoning models penalised pressure tactics, reading them as a signal of low quality. The countdown timer that lifts your human conversion rate can quietly mark your product down in the eyes of the machine.
No, we’re not abandoning CRO. Of course not. For the first shopper, the flesh-and-blood human, that work is still exactly right - us gold-hearted Kiwis still buy on emotion. Love that. The problem is we've been optimising for one audience while quietly assuming the other one sees the same page the same way. It doesn't. Most Kiwi stores we talk to have refined the human UX for years and never once considered the agent reading the same URL with completely different eyes. If this is you, you’re not alone at all.
What the second shopper reads
Take away the sound and lights on the product or service page and what wins an agent over is short and unglamorous: structured product or service data it can parse, a competitive price, and - above all - independent validation.
In a 2025 controlled study of agentic e-commerce, researchers at Columbia and Yale built a sandbox marketplace and set frontier models - GPT-4.1, Claude and Gemini - loose to shop in it. The agents read reviews, of course - but they also mathematically quantified them, leaning hardest on two numbers: average star rating and total review volume. Price mattered as well, though how much, varied sharply from model to model. The product's own marketing claims carried little weight. The machine trusts what other people say about your product far more than what you say about it.
This is key. For the second shopper, your highest-value assets are not your headline copy or your hero banner – they are your reviews, your ratings, your specifications, and how cleanly all of it is structured. Ergo, how cleanly it can be picked up and consumed by the AI bot.
What the agent rewards:
High average star rating and review volume - the strongest signals by some distance
A competitive price - this depends on a number of things, of course
Structured, machine-readable product or service data and specifications
Freshness - agents favour the recently update
What the agent ignores or penalises:
Scarcity countdowns and urgency banners
Strike-through pricing theatre
Bundle framing
Persuasion copy aimed at human psychology
The work you can see is the work that doesn't count much with AI; the work that counts is mostly behind the scenes, underneath the surface. Fun times! 😊
No two agents read alike
There isn't one "AI shopper" to optimise for - each engine reads your website through its own blind spots and preferences. Perplexity leans heavily on third-party citations, especially Reddit threads and expert blogs. Amazon's Rufus synthesises review sentiment and the Q&A on a listing, all inside Amazon's own walls - walls Amazon is defending in court to keep rival shopping agents out. ChatGPT leans on Google Shopping's index and, increasingly, on product feeds merchants send straight to OpenAI. Microsoft's Copilot watches pricing through Bing. Google's Gemini weighs the intent signals in your product data and local inventory.
The pattern underneath is consistent: each agent cares about a slightly different mix of data and trust signals. The single product feed you send everywhere is optimised for one or two of them and overlooked by the rest. You do not need five separate strategies, but you do need to treat "showing up in AI" as several destinations that disagree about what matters - which is why testing where you appear, engine by engine, beats guessing. Please do read that again, the several destinations bit, because someone in the ELT will ask you this question 😊
Building the second layer
None of this means tearing the website down (although sometimes it might pay to do so!). It means building a second layer underneath the one you have - written for the machine, while the human-facing version keeps doing its job. Think of it like an aircraft part: beautifully finished, but if it fails inspection on the spec sheet, it never flies. Your product page can be gorgeous and persuasive and still fail the agent's inspection - unreadable data, a blocked crawler, stale content - and never make the shortlist. A 2026 Adobe analysis found roughly a quarter of retail homepage and category content was effectively unreadable to AI. Most webstores are failing this inspection, and they do not even know that it exists - service businesses fare no better.
The good news, if you're on a flat budget - and if you're a Kiwi marketer, you most likely are - is that most of the fix is hygiene on things you already own. None of it needs a brand-new website (unless your current one is five years old!). Your team can start this week: earn recent reviews and show them as crawlable text rather than locked inside a widget the agent may not be able to read, and check your robots.txt isn't quietly blocking AI crawlers like GPTBot - one line, and some stores are invisible. The rest is a clean brief for your developers: render product data into the page's HTML rather than JavaScript the agent may be unable to run, add proper product and review structured data, keep the feeds tidy. You are going to need to be specific, compared to "go do some AI stuff"!!!
The machine levels the field - our silver lining
All of that can sound daunting – let me give you a wee bit of motivation – and it’s a goodie! Since the beginning of time, us Kiwi marketers have been bullied by bigger brands with bigger budgets, able to do fancier things on their homepages and product pages. No more. Well - sort of.
The agent can't be charmed by any of that. And it can't be outspent. A small Ōtautahi Christchurch store with clean data, a solid GEO foundation and customer reviews can get recommended ahead of a national Auckland brand whose best product page won't render for an AI bot - and that Auckland brand can get ahead of the London one with worse GEO hygiene. For once - for once - the machine levels the field.
So go back to those two shoppers on your page. We've spent a decade learning to win the first one over. The second is already there, reading the same words through different eyes, quietly deciding who gets shown to the next buyer - and it can be won only one way: by being GEO ready, the most readable, most credible answer to the question it was asked. The first shopper made our last ten years. The second will make our next. Bonus? We get to push back against the big-budget-bully brands we've been up against all our lives.
Related reading:
If your customers are other businesses rather than shoppers, the same shift is reshaping how you get found - we've covered how AI now builds the B2B shortlist before a buyer ever makes contact, in our companion piece on AI and B2B buying.
Curious how your website reads to the second shopper - which agents can see your products or services and which can't? That's a Ridiculous special. Bring your busiest product page, or your most critical service page, and we'll show you what the machine sees.
Santosh Pandey is the co-founder and CEO of Ridiculous Digital, a Christchurch agency helping New Zealand businesses with Generative Engine Optimisation (GEO), digital strategy and execution, and paid media.
Disclaimer: This article is an opinion piece from the author named, and does not necessarily reflect the views or opinions of the Marketing Association.