267: The New Buyer You're Not Optimizing For (AI Agents, Product Listings & What the Research Says)
What if the next person to buy your product isn't a person at all?
AI shopping agents — tools like ChatGPT's Agent Mode, Google's "Buy for Me," and Amazon's Rufus — are already live and purchasing products on behalf of real customers. They don't browse. They don't respond to clever branding. They evaluate, decide, and execute.
In this episode, I share an AI-generated audio conversation I created with NotebookLM based on my article "The New Buyer You're Not Optimizing For." Two AI hosts break down new research from Columbia and Yale Universities on exactly how these agents make buying decisions — and what digital creators need to do about it right now.
"You're no longer competing for attention in a crowded list of results. You are competing to be the singular output of a language model."
What You'll Hear in This Episode
Key Research Findings
The conversation covers a working paper from Columbia and Yale Universities that ran thousands of buying scenarios across eight product categories to map the hidden biases of AI shopping agents. Here are the headline numbers:
The change? Moving the primary use case keyword to the front of the product title. A listing called "SUNMORY Floor Lamps for Living Room" became "SUNMORY Office Floor Lamp" — same product, completely different result.
The 5-Step Listing Audit for Digital Creators
The episode walks through Dr. Copp's practical audit framework for digital product creators selling on Shopify, Etsy, course platforms, and beyond:
- Step 1 — Rewrite your titles for intent clarity. Lead with what your product does and who it's for. Brand names go last. "The Brand Builder Bundle" becomes "Canva Social Media Templates for Wellness Coaches."
- Step 2 — Systematize review collection. A 0.1 increase in rating can be the deciding factor between you and a competitor. Build an automated email trigger 7 days after purchase requesting a review.
- Step 3 — Earn platform trust badges. Labels like "Bestseller," "Instructor Recommended," and "Most Popular" act as positive signals for AI agents. Know your platform's criteria and make earning them a goal.
- Step 4 — Justify premium pricing with explicit feature language. AI agents don't respond to charm pricing, but they will filter out expensive products that don't clearly state their differentiators. Spell out what makes yours worth more.
- Step 5 — Run a quarterly agent test. Each time major AI models update, their decision-making logic shifts. Every quarter, test your top three products against ChatGPT, Claude, and Gemini — and adjust based on what you find.
Read the Full Article
This episode is based on the full article published on the Creator's MBA blog. It includes the complete 5-step audit, a breakdown of the Columbia and Yale research, and a source citation for the working paper.
→ The New Buyer You're Not Optimizing For — full breakdown, research citations, and the complete 5-step listing audit for digital creators.
"What is your AI Agent Buying? Evaluation, Biases, Model Dependence and Emerging Applications for Agentic E-Commerce" — Allouah, Besbes, Figueroa, Kanoria & Kumar (Columbia University & Yale University, Working Paper, December 2025). Read the paper →
Is Your Creator Business Built to Be Found?
Take the free Creator Business Scorecard and get a personalized breakdown of where your business stands across all five stages of the Creator Growth Flywheel.
Take the Free Scorecard →Frequently Asked Questions
AI shopping agents are tools like ChatGPT's Agent Mode, Google's "Buy for Me," and Amazon's Rufus that purchase products on behalf of users. Unlike traditional search engines that surface options for humans to browse, AI agents evaluate listings and execute transactions directly — making them a completely new type of buyer.
Start by rewriting your product titles to lead with intent-matching keywords rather than brand names. Then systematize review collection, pursue platform trust badges, ensure your pricing is justified with explicit feature language, and run a quarterly test using the major AI tools to see if your products surface.
Yes — dramatically. Research from Columbia and Yale found that moving the primary use case keyword to the front of a product title led to an 80 percentage point increase in selection by GPT-4.1. The AI reads the first few tokens of a title and makes its match decision almost immediately.
AI agents were trained on human feedback data, where testers consistently skipped sponsored results in favor of organic results. The AI learned that "Sponsored" correlates with lower relevance, so it now actively deprioritizes sponsored listings — even if they would otherwise be a strong match.
A quarterly review of your top three products is the recommended approach. Each time a major AI model updates, the weighting of factors like keywords and trust signals can shift significantly. Test your listings against the latest versions of ChatGPT, Claude, and Gemini each quarter and adjust accordingly.
Pin this and save for later
The New Buyer You're Not Optimizing For
Creator's MBA Show — Transcript
[00:00:00] DESTINI:
Hi there, and welcome back to the Creator's MBA Show. I am Dr. Destini Copp, and before we dive in, I want to give you a quick heads up about this episode because it is a little bit different from my usual format.
I recently published an article on my website called "The New Buyer You're Not Optimizing For," and it's about something I think every digital product creator needs to understand right now — and that is AI shopping agents.
Tools like ChatGPT's Agent Mode and Google's "Buy for Me" are already purchasing products on behalf of real customers, and the research on how they make those decisions is absolutely fascinating. So I ran the article through NotebookLM, and what came out was this really engaging conversation between two [00:01:00] AI hosts breaking down the key findings in the article, and I thought it was really worth sharing.
It's an easy listen, and the insights are immediately actionable. So that's what you are about to hear. After you listen, I'd love to know what you think — and if you want to read the full article with the five-step audit, I'm going to link to that in the show notes.
Alright, here's the conversation.
[00:01:30] HOST 1:
I want you to imagine something for a second. You have spent years — literally years — perfectly crafting an online storefront. You've dialed in the copy, you've uploaded all these high-resolution images, and painstakingly gathered glowing reviews. Doing everything right.
HOST 2: Exactly.
HOST 1: You've built this brand aesthetic that feels entirely your own. You are doing everything right by the traditional playbook. But then you realize that the customer walking through your digital door isn't a human anymore. It's an [00:02:00] algorithm. And that algorithm completely ignores your carefully curated aesthetic.
It does not window shop. It does not browse. It just executes.
Welcome to the deep dive. Today our mission is to unpack a massive paradigm shift happening right now in e-commerce: the rise of AI shopping agents. We are exploring how these tools evaluate products, how they make final purchasing decisions on our behalf, and really why anyone selling anything online needs to completely rethink discoverability — because it fundamentally rewrites the rules of digital commerce.
So to guide us, we are examining this brilliant April 2026 article by Dr. Destini Copp from Creator's MBA. [00:03:00] It's titled "The New Buyer You're Not Optimizing For," and she breaks down a fascinating December 2025 study from researchers at Columbia and Yale University.
HOST 2: The researchers basically ran thousands of test scenarios across multiple product categories, just to map out the hidden buying biases of these agents — which is wild when you see the results.
HOST 1: It really is. But before we get into the tactical side of how you actually optimize a digital storefront for machines, we need to establish the shift itself. We are moving away from an era of search engines and entering an era of decision engines — and that distinction is everything.
Look at the tools already live or rolling out to mainstream users right now: ChatGPT's Agent Mode, Google's "Buy for Me," Amazon's Rufus. [00:04:00] They are everywhere.
HOST 2: This raises an important question. How does an agent actually differ from traditional search?
HOST 1: Well, traditional search surfaces a ranked list of options for a human to look at. The human synthesizes the information, gets distracted, clicks around, and makes the final call. An AI agent eliminates those options entirely. It evaluates the landscape based on its own internal biases and simply executes the transaction.
It's like hiring a hyper-literal personal shopper who absolutely refuses to window shop. They walk into the marketplace with an incredibly rigid checklist, and they just don't care about the mood lighting or the background music or how witty your packaging is.
HOST 2: They don't care at all. If your product does not explicitly match the strict mathematical parameters of their prompt, you simply do not exist to them. And that is a terrifying prospect if your entire marketing strategy relies on catching a human buyer's eye.
You're no longer competing for attention in a crowded list of results — you are competing to be the singular output of a language model. Which brings us to a major question: if an algorithm is doing the shopping, how do you even know what it wants?
[00:05:00] HOST 1:
Does that mean clever, brand-forward naming is completely dead?
HOST 2: Mostly, yes.
HOST 1: Are we just naming things like "Rectangle Box for Stuff" to appease the machines?
HOST 2: You're actually not far off. The biggest takeaway from the Columbia and Yale study is the hypersensitivity these models have to specific words — and more importantly, the exact positional order of those words.
HOST 1: Wait. The order matters that much?
HOST 2: It's the difference between a sale and total invisibility. The researchers tested this across eight different product categories.
HOST 1: Okay, give me an example.
HOST 2: Let's look at their primary one. It was a physical product listed as "SUNMORY Floor Lamps for Living Room." SUNMORY is the brand. Now to a human brain, that string of text is perfectly fine — we instantly parse that SUNMORY is the brand and the object is a lamp. Our brains naturally filter out the brand name to get to the core object.
But an LLM doesn't read like a human. When the researchers slightly tweaked that title to "SUNMORY Office Floor Lamp," the results were staggering. [00:06:00] That single positional title change led to an 80 percentage point increase in selection by GPT-4.1.
HOST 1: Wait — let me stop you there. 80 percentage points?
HOST 2: Yes. An 80-point jump.
HOST 1: That feels like an absurd leap just for putting the word "office" in front of "floor lamp."
HOST 2: So if the buyer asks for an office lamp and the very first tokens the AI encounters are a brand name it doesn't care about, followed by a vague category like "floor lamps" — it just drops it. The mathematical probability of matching that listing to the prompt plummets instantly.
HOST 1: That is wild.
HOST 2: The AI needs to see intent clarity in the very first few tokens. If the primary use case doesn't hit immediately, the model simply moves on. So it's not reading your beautifully crafted 500-word product description.
HOST 1: Not at all. It isn't going to read all that to find a hidden gem the way a patient human might — it filters ruthlessly by early intent tokens.
And this wasn't an isolated quirk with ChatGPT either. Did they test other models? [00:07:00]
HOST 2: Yes. The study showed Gemini 2.5 Flash saw a 52-point increase in selection with that exact same title tweak. And Claude Opus 4.5 saw a 41-point increase.
HOST 1: So the primary use case literally needs to be at the front of the line.
[00:07:30] HOST 2:
Okay, so hacking the tokens gets you past the initial gatekeeper. You match the prompt, you make the cut — but making the initial cut is only half the battle. How does an emotionless algorithm decide what is actually a good or reliable product among the survivors?
HOST 1: That is where trust signals come in. And to understand this, you really have to look at how these agents were built in the first place.
HOST 2: Training data.
HOST 1: Exactly. They are trained on unfathomable amounts of human data. Because of that, they inherently mimic human trust patterns.
HOST 2: I see the logic — but what does that look like in practice? An AI doesn't feel relief when it sees a five-star review.
HOST 1: No, it doesn't feel anything. It assigns a massive mathematical weight to the exact metrics that humans value. The researchers found that basic human trust metrics heavily sway the AI's final selection. A mere 0.1 increase in a product's rating measurably raises its likelihood of being chosen. [00:08:00]
HOST 2: Wait, really? Just a 0.1 delta?
HOST 1: So the jump from a 4.6 to a 4.7 is a huge deal to the algorithm. It can literally be the deciding factor between two functionally identical products.
HOST 2: And it extends way beyond just the star rating.
HOST 1: Absolutely. The models heavily index platform-specific trust badges — labels like "Bestseller," "Our Pick," or "Highly Recommended." These act as incredibly effective social proof signals for the AI, because in the AI's training data, human buyers consistently chose products with those badges. The algorithm is just mathematically formalizing our own psychological biases.
HOST 2: So it is mirroring human behavior.
HOST 1: Yes — but with one glaring, highly counterintuitive exception.
HOST 2: Oh, I have a feeling this is going to upend a lot of marketing budgets.
HOST 1: It already is. The study revealed that labeling a product as "Sponsored" actively [00:09:00] reduces its chances of being chosen by an autonomous agent.
HOST 2: Here's where it gets really interesting — because historically, paying for a sponsored tag has been the ultimate VIP pass. You pay the platform, you get boosted to the top of the search results where the human eyeballs are, and for human buyers, that visibility works.
HOST 1: For an AI agent, that "Sponsored" tag acts as a negative signal.
HOST 2: Why? I thought platforms design their algorithms to prioritize the very ads that make them money.
HOST 1: They do for search. But these decision agents underwent rigorous reinforcement learning from human feedback — RLHF. During that training phase, human testers routinely skipped over sponsored results to find the most organically relevant, high-quality answers.
HOST 2: Because humans don't like ads.
HOST 1: Exactly. The AI learned that "Sponsored" heavily correlates with lower organic relevance or paid interference — so it baked that behavioral pattern into its own [00:10:00] mathematical weight.
It's like paying to wear a neon sign to a VIP club that says "Hey, I bribed my way in" — and instead of letting you through, the AI bouncer actively kicks you out because it's programmed to only let in authentic VIPs.
HOST 2: That's a great way to visualize it. The "Sponsored" tag literally becomes a negative coefficient in the AI's relevance scoring.
HOST 1: Dr. Copp points out that while running paid placements might still keep you visible to the humans who are manually browsing, it will actively penalize you when an agent is doing the shopping.
[00:10:30] HOST 2:
So what does this all mean? We've established that the agent filters by exact early tokens, demands high review velocity, and aggressively rejects paid placements. How does a listener actually apply this — especially a digital creator selling courses, templates, or software on platforms like Shopify or Etsy? They aren't selling floor lamps. How do they audit their own storefronts based on this data?
HOST 1: It requires a fundamental shift in [00:11:00] philosophy. Creators need to move from designing for persuasion to designing for precision.
HOST 2: Precision.
HOST 1: Dr. Copp outlines a highly practical set of audit steps specifically for digital creators. Let's walk through that playbook.
Should a creator literally be surveying their email list to figure out if their specific audience uses ChatGPT more than Claude or Gemini?
HOST 2: Absolutely — because as the data shows, different models weight these positional tokens and trust signals differently. So they aren't all identical under the hood.
HOST 1: Not at all. Claude Opus 4.5, for example, is known to process nuanced, longer-tail context slightly better than earlier GPT models. Whereas Gemini might heavily favor bulleted feature lists.
Knowing what your buyer's agent is running on gives you a baseline. But regardless of the model, step one of the audit is always the same: rewrite your titles to prioritize intent clarity over brand identity.
HOST 2: So if I'm selling a digital product called "The Brand Builder Bundle" — [00:12:00]
HOST 1: You're essentially invisible to an agent. Completely invisible. The AI doesn't know what "Brand Builder" means in a functional sense. You need to strip away the cleverness. Change it to something like "Canva Social Media Templates for Wellness Coaches" — lead with the software it uses, the exact output it provides, and the specific niche it serves. The brand name, if you must include it, goes at the very end.
HOST 2: Clarity over cleverness. Got it. What is step two?
HOST 1: Step two is systematizing your review collection. We know that a 0.1 rating increase can flip an algorithm's decision. The tiny margin matters — so you cannot just hope people leave reviews. Creators need to implement automated sequences, like an email trigger seven days after a digital download, specifically requesting a rating.
HOST 2: Because if you aren't doing this, you are starving the algorithm of the exact trust signals it's looking for.
HOST 1: Exactly. And step three involves those trust badges we discussed. [00:13:00] If your platform offers any sort of organic labeling — like "Instructor Recommended," "Most Popular," or "Top Rated" — you need to figure out the platform-specific criteria to earn them and make that your primary goal. The AI is actively scanning the metadata for those exact strings of text as a proxy for human trust.
HOST 2: I have to ask about pricing, which I assume is step four — because humans are notoriously irrational about pricing. We see a 1% price drop, like a course going from $100 to $99, and our brains process it as a massive bargain. Does the AI fall for the $99 trick?
HOST 1: It doesn't — and this is where the evolution of the models is incredibly apparent. The researchers tested identical products where one was priced just 1% lower. Older AI models used to get tripped up by this, viewing the mathematically lower number as the objectively better value. But newer models resist it entirely. [00:14:00]
HOST 2: How do they resist it?
HOST 1: They evaluate price contextually. Newer models have significantly larger parameter counts and much deeper reasoning capabilities — so they look at historical price data, feature sets, and market averages. The $99 trick does nothing.
However, they will ruthlessly filter out significantly more expensive items if those items lack clear tokenized differentiators.
HOST 2: Wait — explain "tokenized differentiators."
HOST 1: Let's say you are selling a social media template for $100 and a competitor sells a functionally similar one for $20. The AI evaluates the feature tokens of both listings side by side. If your text doesn't explicitly justify why yours is premium — like it includes video tutorials, a commercial license, or one-on-one support — the AI just assumes it's an inefficient purchase and drops you. It filters you out before the human ever even sees it.
HOST 2: So you don't have to race to the bottom on price, but you absolutely have to spell out your premium features using the exact terminology the AI is scanning for. [00:15:00]
HOST 1: Precisely. You have to feed the model the justification.
[00:15:15] HOST 2:
Okay, I am putting myself in the shoes of a creator right now. I've spent my entire weekend rewriting my product titles for precision. I've set up automated review flows. I've explicitly justified my premium pricing. I am optimized.
HOST 1: That's great. But the landscape is a moving target. These models update constantly, and the decision-making logic shifts significantly between versions. The study explicitly noted that the weights applied by GPT-4.1 look very different from earlier iterations.
HOST 2: This sounds exhausting. Are creators supposed to make a full-time job out of chasing the algorithm every single time GPT-5.1 or a new Gemini drops? If the rules of the game change every three months, my storefront could break overnight.
HOST 1: It's a completely valid fear. But if we connect this to the bigger picture, it becomes a system rather than a scramble. Dr. Copp frames this adaptation within what she calls the Creator Growth [00:16:00] Flywheel.
HOST 2: Tell me about the flywheel.
HOST 1: Optimizing for AI agents isn't meant to completely replace your traditional SEO or your social media strategy — at least not yet. It is simply a new, highly specific layer in the Attract stage of your business.
HOST 2: So it's an addition, not a replacement.
HOST 1: Right. And to manage the constant model updates, Dr. Copp recommends abandoning the idea of being perfectly optimized at all times. Instead, you adopt a quarterly testing schedule.
HOST 2: So just checking in every few months.
HOST 1: Yes. Every quarter you take your top three highest-revenue products, open up the latest versions of ChatGPT, Claude, and Gemini, and act as your own target customer. You prompt the AI to solve the exact problem your product solves. If your product isn't the one the agent recommends, you analyze what the winning product had — what tokens or trust signals it used — and then adjust your listings. [00:17:00]
HOST 2: A localized quarterly test on just your top products. That is significantly more manageable than overhauling your entire catalog every time OpenAI holds a press conference.
HOST 1: It really is. And here's the urgency behind doing this now: most businesses still treat AI shopping agents as a futuristic concept. They think it's years away. But it is literally rolling out to mainstream interfaces today.
Creators who systematize this quarterly checkup right now are going to build a compounding advantage. Your competitors are still trying to figure out why humans aren't organically discovering their "Brand Builder Bundle." You have already captured the AI's attention by explicitly stating what your product is.
And when agent-based shopping becomes the absolute default — which, let's be real, is happening incredibly fast — you will already be hard-coded as the preferred choice in the algorithm's weights. You build a moat before anyone else even realizes there's a castle to defend.
[00:17:45] HOST 2:
Let's summarize this for everyone listening, because this is a fundamental shift in how we think about digital commerce. [00:18:00]
The core message is that AI agents are a completely new species of buyer. They do not browse. They do not appreciate your branding. They demand precise intent matching in the very first few words of your title. They rely heavily on review velocity, and they trust organic social proof over paid placements.
In this algorithmic world, absolute precision beats clever persuasion every single time.
HOST 1: But the silver lining is that the foundation of a great product hasn't changed. If you have a genuinely good offering that solves a specific problem, you just need to translate its value into the language the machine understands. The stakes have simply evolved.
The algorithm is now the gatekeeper to the human wallet. Spending an afternoon rewriting your titles today could literally be the reason the AI selects your product six months from now. View your storefront through the eyes of the machine — it is the only way to survive the shift to decision engines.
HOST 2: As we wrap up, I want to leave you with a thought — something this study touched on but left kind of open-ended. [00:19:00]
We talked extensively about how these AI agents are hardwired through their training data to actively deprioritize and reject sponsored products. They view paid placements as inherently less trustworthy than organic results.
So think about this: what is going to happen to the multi-billion dollar advertising revenue models of tech giants like Google and Amazon when human browsing goes entirely extinct? If autonomous agents refuse to click sponsored links, does that entire foundational economic model of the internet just collapse?
Think about that the next time you see a sponsored tag at the top of your screen.
Thank you so much for joining us on this deep dive. Go audit those titles, systematize those reviews, and we will catch you next time.
[00:19:45] DESTINI:
Thanks for listening all the way to the end. I hope you enjoyed this episode today. If you love the show, I'd appreciate a review on Apple Podcasts or your favorite podcast platform. Have a great rest of your day — and bye for now.
Transcript for Creator's MBA Show. Original article: "The New Buyer You're Not Optimizing For" by Dr. Destini Copp — https://www.destinicopp.com/blog/the-new-buyer-youre-not-optimizing-for Audio overview generated with NotebookLM.