🤓 AI Can Predict What You'll Buy—With 90% Accuracy
PLUS: ChatGPT’s major memory upgrades; AI companions go mainstream, and more
New research just found that Large Language Models (LLMs) can predict human purchase decisions—with up to 90% accuracy—based on nothing but the product description.
Not only that, but they can explain those decisions in language that sounds a whole lot like what you’d hear in a focus group.
The models can even offer detailed emotional reasoning for why they’d buy (or skip) a product.
“I’d probably buy it if the scent isn’t too strong and it feels eco-friendly.”
“Seems kinda bougie for this kind of product.”
“The ease of use and safety are appealing, but I’d want to know more about its effectiveness and any potential side effects”
How They Did It
The researchers used a method called Semantic Similarity Rating (SSR)—a prompt structure that gets the model to talk like a person first, then match that answer to real survey scales (like 1–5 stars or “very likely to buy”).
That shift, from rating to explanation, made all the difference.
These models are trained on trillions of words. And they’re fluent in the subtle ways people talk about preference, taste, hesitation, and dealbreakers.
So, instead of guessing which option is “better,” the model surfaced something more useful: the language people use to justify what they want, and why.
And because the models explained the ‘why’ behind their thinking, the researchers got sharper and more detailed reasoning that’s often richer than human open-ended answers.
When researchers compared GPT-4’s responses to real consumer data across 57 product tests and 9,000+ human answers, the model’s predictions closely matched what actual people chose.
Why This Is a Big Deal
For teams who live by consumer insight—product development, brand strategy, marketing, PR, comms—this changes the equation for early-stage testing.
Faster insights, earlier in the process, and at lower cost: Gauge potential reactions to product ideas, messaging or creative directions before spending on panels.
Richer feedback: Because the output is language‑based, you see the why behind reactions, not just the score.
More room to experiment: Test fast and often, even for scrappy ideas that wouldn’t normally get research budget.
This doesn’t replace human research (yet). But it gives you a smarter way to run early litmus tests—with more nuance, speed, and flexibility than you’d expect.
A Few Caveats
It captures intent, not behavior. It predicts what people say they’d do, not what they’ll actually do.
It works best in categories with lots of available public data (like personal care). Don’t assume it’ll translate to niche categories, B2B, or culturally-specific products.
It’s not yet reliable for narrow targets (e.g., “women 35–49 in rural Midwest”).
Bias is baked in. The AI’s training data reflects online voices (forums, reviews, social media) and cultural, socioeconomic, and platform biases.
The researchers say this works best as an early-stage input, not a final decision-making tool.
But even with those limits, it’s a meaningful shift.
It shows that language models can reflect how people make decisions in ways that are accurate, measurable, and scalable.
And remember: this is just GPT-4.
Not the more advanced GPT-4o or the newer reasoning and agentic models.
The next generation of models will push these capabilities even further.
And the better they get at understanding how we think, the more useful they’ll be across every part of the business.
📈 Work with Me
AI Advising and Consulting: Strategic guidance to build sustainable AI strategies and adapt to shifting audience behavior
AI Training: Practical training that helps teams use AI more strategically, effectively, and safely—so it becomes a trusted thought partner, not just a task assistant
What You Need to Know About AI This Week ⚡
Clickable links appear underlined in emails and in orange in the Substack app.
🧠 ChatGPT got new memory upgrades.
OpenAI rolled out major improvements to ChatGPT’s memory, and if you use it regularly, this matters more than you might realize.
Memory is what gives ChatGPT context.
It remembers details from past conversations—like the type of work you do, projects you’re working on, your long-term goals—and uses them to shape future responses.
Even small details can influence the direction, quality, and relevance of ChatGPT’s responses.
These upgrades give you more control over that context and make it easier to keep your conversations useful, personalized, and on track.
👉 Here’s what’s new for Plus ($20/month) and Pro ($200/month) users:
Smarter memory management: ChatGPT now decides which memories to keep “top of mind” based on how recent they are, and how often you talk about a topic. Older and less relevant ones move to the background, so the model stays focused on what matters most.
More visibility and control: You can now see which memories are active, sort or search through them, prioritize or deprioritize specific ones, and even view and restore past versions. You can also turn off auto-management entirely. (Though why would you do that. Don’t do that.)
No more memory limits: Some users were hitting capacity when trying to save new memories, but now you’ll have more room to keep adding context without running out of space.
As ChatGPT leans further into personalization, memory (and how you manage it) will be the key driver of response quality.
I spend hours with clients digging into memory (including Custom Instructions), because it really is that important.
💡 Pro Tip: Before diving into these memory upgrades, make sure your Custom Instructions are set.
They’re the most consistently applied layer of context and memory ChatGPT uses for every single conversation and take priority over all other saved memories.
If you haven’t touched yours in a while (or never set them up), the two posts below walk through what to do—and why it matters.
ChatGPT’s interface has changed a bit since I wrote them, but the guidance still holds up.
I’ve changed the formatting of my newsletter since then too, so you’ll need to scroll down to the middle of these editions to find the right section.
OpenAI also has a really helpful FAQ page which should address most of the basics.
🛍️ You can now shop Walmart in ChatGPT.
Walmart has joined Etsy and over a million Shopify merchants like Glossier, Skims, and Spanx in integrating ChatGPT’s Instant Checkout, allowing users to browse and buy its products without ever leaving the chat.
The products will include nearly everything available on Walmart’s website, except for fresh food. And Walmart+ members will still get their benefits such as free shipping when making purchases through ChatGPT.
The retail giant’s stock was up 5.2% the day following the announcement.
For the full breakdown of how OpenAI’s Instant Checkout works and why it matters) check the weekly update from this earlier edition.
😦 AI companions are going mainstream.
A new survey finds that one in five high schoolers report they or someone they know has had a romantic relationship with AI. 42% have used AI for companionship.
Meanwhile an analysis by ChinaTalk this month found that AI chatbots designed for sex and romance have about 29 million monthly users globally.
And OpenAI will allow verified adults to have erotic conversations starting in December.
🖍️ Sometimes AI is just... fun.
I’ve been messing around with Google’s Nano Banana image generator. Here’s what happened when I fed it a photo of myself and asked for a line drawing.
You can now access the model through both Google search and Notebook LM instead of having to sign into Gemini.
Here’s the exact prompt I used:
Create a photo-style line drawing/ink sketch of the faces identical to the uploaded reference image. Keep every facial feature, proportion, and expression exactly the same. Use blue and white ink tones with intricate, fine line detailing, drawn on a white notebook-page style background.
In case you missed last week’s edition, you can find it 👇:
That's all for this week.
Thoughts, feedback and questions are always welcome and much appreciated. Shoot me a note at avi@joinsavvyavi.com.
Stay curious,
Avi
💙💙💙 P.S. A huge thank you to my paid subscribers and those of you who share this newsletter with curious friends and coworkers. It takes me about 20+ hours each week to research, curate, simplify the complex, and write this newsletter. So, your support means the world to me, as it helps me make this process sustainable (almost 😄).
I could definitely see how it could predict products that you buy, especially with enough data. If you buy the same things month after month, I think the most disturbing thing here is that teenagers are now having romantic interactions with these LLMs. The lines are just getting so blurred 😔