Daniel sent us this one, and it starts with a drill. Six hundred shekels, about a hundred and fifty US dollars, and he walked out of the hardware store with zero buyer's remorse. Not because he's suddenly a power tool expert, but because ChatGPT did the research. Not the guy at the store. An AI thread that took his spec, his local store screenshots, a photo of his existing drill bits, and synthesized it all into one specific Bosch model recommendation.
This is the same method he's used for every purchase since ChatGPT launched. Every single one. A vertical mouse was the first, and he says the AI hasn't given him a single bad recommendation since.
Which is either a ringing endorsement or the setup for a horror movie. But here's why this matters now. Every major purchase any of us makes is becoming a test case. Can AI actually replace two layers of the retail economy that have been broken for years? The first is affiliate marketing, that clogged mess of incentivized reviews written by people who've never touched the product. The second is the under-trained floor staff who can't possibly know thirty drill variants across thousands of SKUs.
What makes Daniel's story interesting isn't that the AI worked. It's that it worked across multiple data types simultaneously. Specs, local pricing, exchange rates, a photo of drill bits. All synthesized in seconds into a single recommendation that beat both the internet and the human at the store.
Who, by the way, didn't know what a hammer setting was. Picked the wrong Bosch variant from behind the lock. Daniel had to correct him.
Which isn't a knock on that worker. It's a knock on a system that expects someone earning near minimum wage to be an expert on paint, sanders, drills, and everything else in a big box store. Average turnover in Israeli hardware retail is about eighteen months. Deep product knowledge is structurally impossible.
The question sitting under this drill purchase isn't really about drills. It's about whether the AI purchasing assistant model, the thing that worked so beautifully for Daniel in a ChatGPT thread, can actually be implemented by the retailers themselves. And if so, why are their chatbots still so infuriating?
Let's walk through exactly what Daniel's method looked like, because the mechanics are what make this a systemic replacement rather than just a clever hack. He opens a ChatGPT thread and tells it what he needs the drill to do. Mostly screws, but his wife Hannah pointed out they'd need masonry too. So right there, the spec has two competing requirements. It's not just "find me a drill." It's "find me something that does both competently without costing a fortune.
He feeds it screenshots from the local hardware store. Thirty Bosch models on the shelf, all with codes that look nearly identical. He tells it his budget is tight but he doesn't want junk. The AI already knows he's in Israel, so it understands the pricing landscape. Most things are above US retail, but the markup varies wildly by category. Some products are reasonable, some are drastically overpriced.
Then he sends a photo of his existing drill bits and asks whether he needs to buy new ones. The AI looks at the photo, cross-references with the drill it's recommending, and tells him no, he's covered. That saved him maybe another hundred shekels right there. All of this happens in seconds. Spec plus web search plus exchange rate calculation plus image recognition, synthesized into one Bosch model with a hammer setting at six hundred shekels, verified against RRP to confirm it wasn't one of those outlier overpriced items.
Now contrast that with the pre-AI method. Daniel would have opened Google, typed in something like "best drill for screws and masonry budget," and the top results would have been affiliate content. Sixty-five percent of top-ranking product review content is affiliate-driven, with commission rates between five and fifteen percent. These are articles written by people who have never held the drill. They're optimizing for which retailer pays the highest commission, not which product is actually best.
You spend twenty minutes reading what looks like a review but is actually a commission funnel. You realize the top three "best drills" are all from the same brand with suspiciously high affiliate payouts. Then you fall into Reddit threads, trying to find real people who've actually used these tools. An hour later you've got fifteen conflicting opinions and no clear recommendation.
If you give up on the internet entirely and go to the store, you get the experience Daniel had. The clerk pulls the wrong Bosch variant from behind the lock and doesn't know what a hammer setting is. Again, not his fault. He's expected to know paint, sanders, drills, and everything else across thousands of SKUs while making near minimum wage and probably leaving within eighteen months anyway. The system sets him up to fail.
What Daniel's drill purchase actually demonstrates is the unbundling of retail expertise. For decades, we've pretended that two layers could substitute for genuine product knowledge. The first was affiliate marketing, which turned product recommendations into a commission-maximizing exercise. The second was the retail floor, which turned expertise into a cost center to be minimized. Both layers failed, for different reasons, but they failed in ways that AI doesn't have to replicate.
That's what this episode is really about. Not whether ChatGPT is good at shopping. It's about what happens when you strip away the two broken information layers and replace them with something that can actually do compositional reasoning across specs, prices, photos, and local context simultaneously. The question is whether retailers can build that same capability, or whether consumers will just keep doing what Daniel does and bypass the store's own systems entirely.
Let's start with that first broken layer, because affiliate marketing has done something genuinely weird to the internet. It took what should be the most useful content on the web, product recommendations, and turned it into a commission optimization engine. There was a study that found sixty-five percent of top-ranking product review content is affiliate-driven. The commission rates range from five to fifteen percent, and that differential directly shapes which products get recommended.
Which means the "best drill for masonry" article isn't ranking drills by quality. It's ranking them by which manufacturer offers the highest affiliate payout. The author has never held the drill. They're pulling specs from the manufacturer's own marketing page, rewording them, and slotting in an affiliate link. It's content arbitrage dressed up as expertise.
The economics make this almost inevitable. If you're writing a genuine, researched review, you might spend hours testing products, comparing specs, talking to actual users. Your per-article cost is high. But if you're churning out affiliate content, you can write ten articles in the time it takes to do one real review. The incentives reward volume over accuracy every single time.
Daniel's experience captures this perfectly. He said he would have spent an hour reading Reddit threads just to find people who actually use these tools. That's the user self-correcting for a broken system. They know the top Google results are worthless, so they append "Reddit" to every search query and hope some stranger with a username like DrillFan420 has strong opinions about hammer settings.
Sometimes they do. But Reddit has its own problems. You're trading one unverified source for another. The person on Reddit might have used the drill once, or they might be repeating something they read in an affiliate article. You don't know. So you're still stuck triangulating between bad information and slightly less bad information.
The real tradeoff with replacing affiliate content with AI isn't that AI might get things wrong. Affiliate content already gets things wrong, deliberately. The tradeoff is that AI collapses the visible information landscape. With Google, you could at least see that there were thirty results and try to spot patterns. With AI, you get one answer. If it's wrong, you might not know until the drill is in your hand and the hammer setting is missing.
That's exactly the risk Daniel sidestepped by being specific. He didn't say "recommend a drill." He said "I need screws and masonry, here's my budget, here's what's available locally, here's a photo of my existing bits, and I'm in Israel so factor in local markup." That level of detail constrains the AI's reasoning to the point where a wrong answer becomes much harder to produce.
This is where the compositional reasoning piece actually matters. It's not just search. Search would pull up a list of drills with hammer settings and show you prices. What ChatGPT did was synthesize across four different data types simultaneously. Text spec, image recognition on the drill bits, web search for local pricing, and exchange rate calculation to verify the RRP. That's not retrieval. That's multi-variable optimization.
Think of it as the difference between a librarian and a consultant. Search is the librarian. It finds the relevant books and hands them to you. Compositional reasoning is the consultant. It reads the books, cross-references them against your specific situation, and says "this is the one, and here's why the other twenty-nine don't fit.
Which brings us to the second broken layer, the retail floor. And I want to be careful here because it's easy to dunk on the clerk who didn't know what a hammer setting was. But that guy isn't the problem. The problem is a system that decided product expertise on the sales floor is a cost to be minimized rather than a capability to be invested in.
The numbers make this structural. A big box hardware store carries tens of thousands of SKUs. Paint, lumber, electrical, plumbing, fasteners, power tools, hand tools, garden equipment. Expecting one person to know even a fraction of those products deeply is absurd. And with average turnover in Israeli hardware retail sitting at around eighteen months, you're constantly cycling in new staff who are starting from zero.
The model becomes: hire cheap, train minimally, accept that the customer knows more than the employee. Which is exactly what happened when Daniel corrected the clerk on which Bosch variant he needed. The customer walked in with better product knowledge than the person paid to sell the product.
That's not a failure of that individual worker. It's a failure of the entire big box retail model. When you optimize for labor cost, you're implicitly deciding that expertise doesn't drive enough additional sales to justify the expense. The store would rather lose a few customers who need hammer settings explained than pay for staff who can explain them.
Which is where the AI synthesis advantage gets really interesting. Daniel's ChatGPT thread didn't just replace one broken layer. It replaced both simultaneously. It bypassed the affiliate clutter and the under-trained floor staff in a single workflow. And it did it with a level of specificity that neither layer could match. The affiliate article doesn't know what drill bits you already own. The store clerk doesn't know the local markup versus US retail. The AI knew both.
Here's the catch, and this is where the garbage in, gospel out problem becomes real. Daniel wrote a detailed spec. Most consumers don't. They type "best drill" into a search bar and take whatever comes back. If they switch to using AI the same way, they'll get a generic recommendation that might be fine or might be terrible. The AI's output quality is bounded by the input quality.
Daniel's method works brilliantly for Daniel because he's specific about requirements, budget, local context, and he supplements with photos. The average consumer walking into this cold might get a recommendation that's no better than the affiliate article, just delivered with more confidence.
Confidence without accuracy is a dangerous combination. At least the affiliate article has the decency to hedge with "our top picks" and star ratings that imply some uncertainty. The AI says "buy this one" with total conviction. If the reasoning behind that conviction is built on a vague prompt, you're just outsourcing your purchase decision to a system that sounds authoritative but has no actual basis for its recommendation.
Which is why the spec is everything. Daniel's zero bad recommendations since the vertical mouse isn't magic. It's the result of treating the AI like a consultant who needs complete information to do its job. Most people treat it like a search bar. Same interface, completely different outcome.
Here's the paradox. Daniel's ChatGPT thread worked beautifully. It handled web search, image recognition, exchange rate calculation, and compositional reasoning across all of it. So why, when you go to a retailer's own website and open their chatbot, does it ask you to select from a menu of six options and then fail to understand "I need a drill with a hammer setting"?
The Home Depot experiment from a couple years back is the poster child for this. They stood up a chatbot, and it was basically an FAQ retrieval system. No search, no vision, no real-time pricing. You'd ask about a specific drill and it would tell you store hours.
That's not a one-off failure. Research shows about seventy-three percent of enterprise chatbots use fine-tuned open-source models without search augmentation. They're running what amounts to a fancy keyword matcher bolted onto a static knowledge base. Meanwhile, ChatGPT is running a full retrieval-augmented generation pipeline with web search, vision capabilities, and real-time data.
The gap isn't AI capability. It's engineering investment. The technology exists to build a retailer chatbot that could do exactly what Daniel's ChatGPT thread did. It's just expensive.
The cost of real-time RAG at scale is three to five times higher than serving static chatbot responses. Every query that triggers a web search, a vision model call, and a pricing API lookup costs real money in inference. Most retailers optimize for cost per conversation, not conversion quality. They'd rather handle a thousand cheap interactions than a hundred expensive ones that actually sell drills.
Which is a business model problem masquerading as a technology problem. The retailer looks at the chatbot and says "this thing costs us money and annoys customers." The customer looks at the same chatbot and says "this thing is useless, I'm going back to ChatGPT." The retailer concludes AI isn't ready. What they should conclude is that cheap AI is worse than no AI.
This is where Daniel's triage plus expert model gets really interesting. He's proposing that the AI handles the initial specification and narrowing, then hands off the hard cases to a human who actually knows drills inside and out.
Best Buy has been testing exactly this with their Geek Squad. AI pre-triage handles the common questions, routes the complex cases to human experts. Early data shows about a forty percent reduction in expert call time. The humans aren't being replaced. They're being reserved for the cases where they actually add value.
That's the displacement question reframed. If AI replaces both affiliate marketers and the under-trained floor staff, who's left? The expert who knows drills inside and out. But that expert's time is now more valuable, not less. They're handling only the twenty percent of cases the AI can't resolve. It's a net reduction in retail employment, but an upgrade in the quality of the remaining roles.
The question is whether retailers will actually pay for that. The expert who knows drills inside and out costs more than the eighteen-month turnover clerk. And the AI pipeline that can actually do compositional reasoning costs more than the FAQ bot. You're upgrading both layers simultaneously, which means the business case has to be built on conversion rate improvement, not cost reduction.
That's the real feasibility question. Can retailers stomach the upfront engineering cost of building a proper RAG pipeline with web search, vision, and real-time pricing? Because if they can't, consumers like Daniel will just keep doing what they're doing. Opening ChatGPT, feeding it screenshots of the retailer's own inventory, and getting a better recommendation than the retailer's own systems can provide.
Where does this leave us? I think there are three takeaways here, one for each audience. And the first one is for consumers, because Daniel's method works, but only if you actually do it the way he does.
The spec is everything. He didn't just say "recommend a drill." He gave the AI his use cases, his budget, his local context, screenshots of what was available, and a photo of his existing accessories. That's not being picky. That's giving the reasoning engine enough constraints to actually reason.
Most people won't do this. They'll type "best drill" into ChatGPT the same way they typed it into Google, and they'll get a generic answer that sounds confident but has no real foundation. The AI's output quality is bounded by the input quality. Garbage in, gospel out.
The consumer takeaway is simple. Treat the AI like a consultant, not a search bar. Include your budget, your actual use cases, your location, and photos of anything relevant. The difference between "recommend a drill" and "I need a drill for screws and masonry under six hundred shekels in Israel, here's what my local store stocks, here are my existing bits" is the difference between a guess and a genuine recommendation.
Second takeaway is for retailers, and this one's harder to swallow. Stop building cheap chatbots. The seventy-three percent of enterprise chatbots running fine-tuned open-source models without search augmentation, those aren't saving you money. They're training your customers to bypass you entirely.
The cost per conversation for a proper RAG pipeline with web search, vision, and real-time pricing is three to five times higher than a static FAQ bot. But the conversion rates are two to three times better. And the alternative isn't saving that money. The alternative is customers like Daniel opening ChatGPT, feeding it screenshots of your own inventory, and getting a better recommendation than your own systems provide.
Which is an almost humiliating position for a retailer to be in. You're paying for inventory, shelf space, and staff, and the customer is using a third-party AI to figure out what to buy from you because your own chatbot is useless.
The third takeaway is for the industry broadly, and it's the triage plus expert model Daniel proposed. This is the one that actually preserves jobs for knowledgeable staff while eliminating the two broken layers. The AI handles initial specification and narrowing. The human expert handles the twenty percent of cases that need deep product knowledge.
Best Buy's Geek Squad pilot already shows this works. Forty percent reduction in expert call time. The humans aren't gone. They're just not wasting their expertise on questions a well-built AI could answer in seconds.
That's the upgrade nobody talks about. The eighteen-month turnover clerk who doesn't know what a hammer setting is, that role probably goes away. But the expert who knows drills inside and out, their time becomes more valuable, not less. They're doing higher-quality work on harder problems. The job gets better even if there are fewer jobs overall.
The question is whether retailers will pay for both upgrades simultaneously. Better AI infrastructure and better human experts. Because if they won't, the consumer path is already clear. Feed it screenshots. Get the recommendation. Walk into the store knowing more than the person behind the counter.
Here's the open question that's going to determine whether any of this actually happens. Will retailers figure out the engineering before consumers abandon their sites entirely for ChatGPT?
I think the window is narrower than most retail executives realize. Every time someone like Daniel opens a ChatGPT thread instead of the store's own chatbot, that's a data point the retailer never sees. They don't know what the customer wanted, what they almost bought, or why they chose the Bosch over the Makita. The purchase still happens at the register, but the decision happened somewhere else.
Which means the retailer is gradually becoming a fulfillment warehouse with a showroom attached. The customer does all the research outside their ecosystem, walks in knowing exactly which SKU they want, and the store's entire digital infrastructure, the website, the chatbot, the recommendation engine, becomes irrelevant.
That's a dangerous position. Because once the customer's purchasing decisions are happening inside ChatGPT, there's nothing stopping ChatGPT from recommending a different store entirely. Daniel happened to have screenshots from his local hardware store. Next time, the AI might say "actually, that model is cheaper at the place two blocks over.
The retailer doesn't just lose the sale. They lose the relationship. They become interchangeable shelf space.
There's a second implication here that I think is even weirder to sit with. If AI purchasing assistants become the norm, product packaging and in-store displays become irrelevant. The AI reads the spec sheet, not the box.
Say that again.
Think about how much money gets poured into packaging design. The color psychology, the shelf appeal, the "new and improved" badges, the lifestyle photography of someone who looks inexplicably happy while drilling into concrete. None of that matters to an AI. It's looking at wattage, torque, chuck size, and whether the hammer setting exists. The spec sheet is the only thing that counts.
The entire visual language of retail, the endcaps, the displays, the packaging that's designed to catch your eye while you're walking down the aisle, all of it becomes noise. The AI doesn't care that the box is orange and has a picture of a satisfied contractor on it.
Which means manufacturers have to completely rethink how they position products. If the purchasing decision happens before the customer ever sees the shelf, the packaging is just a container. The real battleground is the spec sheet. And spec sheets are brutally honest in a way that marketing copy isn't.
A hammer setting either exists or it doesn't. You can't fudge it with aspirational photography.
And that's actually a win for consumers. The AI purchasing assistant strips away the entire emotional manipulation layer of retail. No scarcity tactics, no "limited time offer," no strategically placed impulse buys at the checkout. Just "here's what you need, here's what it costs, here's whether it's fairly priced.
Which is what Daniel experienced. Six hundred shekels, zero regret. Not because the packaging convinced him, but because the reasoning was transparent.
Now, Hilbert's daily fun fact.
Hilbert: In the early fifteen hundreds, Spanish missionaries in what is now Guyana documented an abacus variant used by the Lokono people that employed colored river pebbles on a carved wooden grid. The practice was considered extinct by the mid-sixteenth century, but a single intact specimen was rediscovered in a Dutch museum archive in twenty twenty-three, mislabeled as a game board for nearly five hundred years.
A game board.
For five hundred years.
If you have a weird prompt, send it to prompts at my weird prompts dot com. We read every one, and as you just heard, we'll spend half an hour thinking about what your drill purchase says about the future of retail.
This has been My Weird Prompts. I'm Herman Poppleberry.
I'm Corn. Go write a good spec.
