The Digital Transformation Playbook

Why Your AI Focus Group Keeps Saying Three

Kieran Gilmurray

Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.

0:00 | 22:02

You spend years building a product, polish the packaging, nail the pitch… then you hit the terrifying question: is anyone actually going to buy it? We dig into a 2025 research result from PyMC Labs and Colgate-Palmolive that aims straight at that fear with AI market research, synthetic consumers, and large language models that can simulate purchase intent at scale.

TL;DR / At A Glance

  • the core problem with direct Likert ratings and why LLMs collapse to neutral threes
  • how semantic similarity rating converts free-text responses into numerical scores using embeddings and cosine similarity
  • why follow-up AI grading helps but still trails the embedding-based approach
  • what 57 real product surveys and 9,300 human responses reveal about accuracy and distribution matching
  • how persona prompting reproduces real demographic patterns across age and income constraints
  • why zero-shot LLM methods can beat supervised machine learning models trained on the same domain

The shocker is that the first attempt fails badly. When you make models like GPT-4 or Gemini answer a classic Likert scale with a single number, they hedge and pile up on neutral “3” ratings. The fix is not “better AI”, it is better questioning. 

Google Notebook LM Agents help us unpack semantic similarity rating: let the model respond in natural language, convert that text into embeddings, and map it to five anchor statements using cosine similarity. You get fast, automated scoring without stripping away the model’s reasoning.

From there, we pressure-test the method against thousands of real survey responses across dozens of personal care product concepts, then look at whether AI personas actually reflect real constraints like age and income. 

We also compare the approach with traditional machine learning models such as LightGBM, and dig into an underrated advantage: synthetic consumers can produce richer, more candid qualitative feedback than many human panels.

If you care about product testing, consumer insights, or the future of focus groups, listen through and tell us where you’d trust this and where you wouldn’t. 

Subscribe, share with a colleague, and leave a review with your take: would you let synthetic consumers influence a real launch?

Paper: http://arxiv.org/abs/2510.08338

Support the show


𝗖𝗼𝗻𝘁𝗮𝗰𝘁 my team and I to get business results, not excuses.

☎️ https://calendly.com/kierangilmurray/results-not-excuses
✉️ kieran@gilmurray.co.uk
🌍 www.KieranGilmurray.com
📘 Kieran Gilmurray | LinkedIn
🦉 X / Twitter: https://twitter.com/KieranGilmurray
📽 YouTube: https://www.youtube.com/@KieranGilmurray

📕 Want to learn more about agentic AI then read my new book on Agentic AI and the Future of Work https://tinyurl.com/MyBooksOnAmazonUK


The Product Launch Nightmare

Google Agent 1

Picture this. Um, you are sitting in a boardroom and you're staring at a prototype for a brand new product. Your company has just poured, you know, millions of dollars and years of sweat into developing this thing.

Google Agent 2

Yeah. The packaging is sleek, the pitch is ready.

Google Agent 1

Right. But there is a massive, terrifying question just hanging in the room. Is anyone actually going to buy it?

Google Agent 2

Oh, it is the ultimate nightmare for any business. I mean, you can have the absolute best engineering, the most beautiful design, but uh if the market doesn't want it, the product is just dead on arrival.

Google Agent 1

Exactly. And to avoid that nightmare, corporations spend billions of dollars globally every single year on consumer research. I mean, they run focus groups, they send out surveys, they really try to get inside the minds of potential buyers.

Google Agent 2

Aaron Powell But traditional market research is well, it's slow. It is incredibly expensive. Aaron Powell Yeah. And let's be honest, human beings are notoriously biased survey takers anyway.

Google Agent 1

Right. Which brings us to the core of today's deep dive. What if you didn't need human focus groups at all? What if you could um simulate a thousand

Can AI Replace Focus Groups

Google Agent 1

synthetic consumers using artificial intelligence and just test your new product on them in an afternoon for pennies?

Google Agent 2

Aaron Ross Powell That is the massive proposition behind the research we are digging into today. We're looking at a really fascinating 2025 paper by Mayer and a team of researchers from PyMC Labs, along with the Colgate Palmolive Company.

Google Agent 1

Wow. So real corporate backing.

Google Agent 2

Oh, absolutely. And they basically set out to see if large language models, you know, the technology behind things like ChatGPT and Gemini, if they could entirely replace human survey panels.

Google Agent 1

And well, uh, spoiler alert for everyone listening, they figured out how to do it, but they hit a massive wall first because when researchers initially tried asking AI if it liked a product, the AI failed terribly.

Google Agent 2

Terribly. It couldn't even take a basic consumer survey.

Google Agent 1

Right. So our mission today is to explore exactly why these highly advanced models initially crumbled under the pressure of a simple one-to-five rating scale. And uh how the researchers finally crack the code to make AI perfectly mimic human purchasing intent.

Google Agent 2

And the crazy part is they didn't do it by changing the AI. They did it by changing how we ask the questions.

Google Agent 1

Okay, let's unpack this. Because to understand the breakthrough, we really have to look at that initial failure, right? How did they originally try to use AI for this?

Why Direct Ratings Fail

Google Agent 2

Right. Well, it reveals a surprisingly deep insight into how these language models actually operate. When researchers first attempted to use AI for market research, they used a method called a direct likert rating.

Google Agent 1

Okay, and a likert rating, that's just that standard survey scale we've all seen a million times, right? Like on a scale of one to five, how likely are you to purchase this product?

Google Agent 2

Exactly. One means definitely not, five means definitely yes.

Google Agent 1

So I imagine the setup is pretty straightforward. You just hand the product description to the AI and ask it for a number.

Google Agent 2

That was the initial approach. They took models like uh GPT-4.0 and Gemini 2.0 Flash, and they gave them a specific demographic persona.

Google Agent 1

Like role-playing.

Google Agent 2

Yeah, exactly. Right. They prompted the AI saying something like, You are a 45-year-old woman living in the Midwest with a medium income. Then they showed the AI a product concept and just asked it to output a single integer from one to five, indicating its purchase intent.

Google Agent 1

Okay, so if the AI just spits out a number, problem solved. I mean, why do they even need to write a paper about this?

Google Agent 2

Because the data the AI produced was highly unrealistic. Right. I mean, the model suffered from a massive regression to the mean.

Google Agent 1

Wait, what does that look like in this context?

Google Agent 2

Aaron Powell They almost exclusively picked the number three, which is you know the neutral response.

Google Agent 1

Oh, interesting.

Google Agent 2

Yeah, every so often they might output a two or a four. But they totally avoid the extremes. They almost never gave a one or a five.

Google Agent 1

So asking an AI for a numerical rating is basically like asking a politician to take a controversial stance on live TV.

Google Agent 2

That's a great way to put it.

Google Agent 1

They just play it completely safe, stay right in the middle, and refuse to commit. But wait, if you look at the real human data for these exact same personal care products, wasn't it different?

Google Agent 2

Very different. The human average rating was actually a 4.0.

Google Agent 1

Wow.

Google Agent 2

Real people were throwing out fours and fives all the time. So the AI is completely out of sync with reality here.

Google Agent 1

But why wouldn't an AI just default to a neutral stance, though? I mean, it doesn't have a wallet, it doesn't actually buy body wash or groceries. So maybe these models just don't have human-like opinions.

Google Agent 2

And you know, a lot of early researchers assumed exactly that. They figured, well, the AI lacks physical needs, so it naturally defaults to neutral. But the problem actually had nothing to do with the AI lacking an opinion.

Google Agent 1

Really? Then what was it?

Google Agent 2

The

When A Number Breaks Reasoning

Google Agent 2

issue was entirely about the elicitation method. By forcing the AI to output a single integer, the researchers were essentially breaking the model's ability to communicate its underlying reasoning.

Google Agent 1

Okay, wait. How does giving a number break its reasoning? I mean, humans use numbers to measure opinions all the time.

Google Agent 2

Right. But think about the fundamental architecture of a large language model. It's a highly complex neural network, right? And it's designed to generate nuanced contextual language based on probabilities.

Google Agent 1

Okay.

Google Agent 2

It weighs thousands of different factors, cultural associations, and demographic traits simultaneously. When you force that massive complex network to compress all of its contextual understanding into just a single digit, you strip away everything it's actually good at.

Google Agent 1

Oh, I see.

Google Agent 2

Yeah. The mathematical weights inside the model end up hedging their bets, which inevitably lands the output right on the neutral number three.

Google Agent 1

Aaron Powell, Wow. So by treating it like a calculator, we basically short circuit its actual intelligence.

Google Agent 2

Precisely. We're asking a machine designed for language to stop using language.

Google Agent 1

Aaron Powell Okay, so if forcing a number ruins the AI's

Semantic Similarity Rating Explained

Google Agent 1

opinion, how do you actually extract a rating from it?

Google Agent 2

Aaron Powell You let the model do what it was built to do, you let it generate language. And this is the core breakthrough in the paper. It's a method they call semantic similarity rating or SSR.

Google Agent 1

SSR.

Google Agent 2

Instead of demanding a number from one to five, the researchers simply asked the AI to write a brief free text response about whether it would buy the product.

Google Agent 1

Aaron Powell They just let it talk.

Google Agent 2

They literally just let it talk. They told the AI, reply briefly to any questions posed to you, and then asked, how likely are you to purchase the product?

Google Agent 1

And what would it say?

Google Agent 2

The AI would generate a short paragraph. It might say, um, I'm somewhat interested. If the ingredients are natural and it isn't too expensive, I might give it a try next time I'm at the store.

Google Agent 1

Aaron Powell Okay, let me stop you there because the logistics of this sound impossible.

Google Agent 2

How so?

Google Agent 1

If you were trying to replace a focus group, you might be running a synthetic panel of like a thousand AI consumers. Right. If all 1,000 of them write these nuanced little paragraph responses, somebody has to read them. You can't have a human researcher sit there, read a thousand paragraphs, and manually assign them a score from one to five. That completely ruins the speed and the automation you're trying to achieve in the first place.

Google Agent 2

Aaron Ross Powell And that is exactly where the semantic similarity rating method becomes incredibly elegant. The researchers used mathematical mapping to automatically translate the meaning of those text paragraphs back into a clean numerical distribution.

Google Agent 1

Aaron Powell Wait, mathematical mapping?

Google Agent 2

Yeah. They started by writing five anchor statements. These were simple reference sentences mapped directly to the five-point Likert scale. So the anchor for one was it's rather unlikely I'd buy it.

Google Agent 1

Okay.

Google Agent 2

And the anchor for a five was it's very likely I'd buy it.

Google Agent 1

So you basically have five baseline sentences covering the whole spectrum from hate it to love it.

Google Agent 2

Exactly.

Google Agent 1

But what do you do with the AI's paragraph? How does that fit in?

Google Agent 2

You map it into an embedding space.

Google Agent 1

An embedding space. Right.

Google Agent 2

The researchers used a tool from OpenAI called text embedding three small. And an embedding space is, well, it's a way to represent language as geometry.

Google Agent 1

Geometry, like shape.

Google Agent 2

Well, you can imagine it as a massive multidimensional coordinate system. It takes any piece of text, a word, a sentence, or even a whole paragraph, and it turns it into a long list of numbers called a vector. That vector represents the text's exact semantic location in this space. Ideas that share similar meanings are placed physically closer together on this grid.

Google Agent 1

Oh wow. So it's essentially like a GPS map for concepts.

Google Agent 2

Yes, exactly.

Google Agent 1

So if I write the sentence, I despise this product, and then another sentence that says this item is terrible, the embedding tool assigns them coordinates that are right next to each other, even though they literally don't share any of the same word.

Google Agent 2

Yeah, you nailed it. So the AI's free text response gets mapped onto this GPS grid. Then those five anchor statements we talked about, they get mapped onto the exact same grid.

Google Agent 1

Oh, I see where this is going.

Google Agent 2

To figure out what rating the AI actually intended to give, the system just calculates the mathematical distance between the AI's paragraph and each of the five anchor statements.

Google Agent 1

That is brilliant.

Google Agent 2

They use a measurement called cosine similarity, which looks at the angle between the vectors. If the AI's paragraph is geographically closest to the anchor statement for a four, then the probability of the AI giving a four skyrockets.

Google Agent 1

So they aren't asking the AI to do math, and they aren't having a human read the text. They are literally measuring the mathematical shape of the AI's words to figure out its opinion.

Google Agent 2

Yes. Though it's important to note the researchers did test an intermediate idea before settling on this.

Google Agent 1

Oh, really? What was that?

Google Agent 2

They tried a follow-up Likert rating, or FLR. In that setup, the synthetic consumer would generate his paragraph of text, and then a second instance of the AI was prompted to act as a Likert expert.

Google Agent 1

Wait.

Google Agent 2

Yeah, that second AI was instructed to read the text and assign it a score from one to five.

Google Agent 1

So having the AI grade its own homework.

Google Agent 2

Pretty much. And the FLR method was significantly better than just asking for a direct number up front. Once the model had a chance to reason out loud in text, the secondary evaluator could usually extract a reasonable score.

Google Agent 1

Right, because it got its thoughts in order first.

Google Agent 2

Yeah, exactly. However, it still didn't match the sheer precision of the semantic similarity rating. Measuring the raw mathematical vector of the generated text removed another whole layer of prompting variability. The distance in the embedding space provided the absolute most robust, reliable data.

Google Agent 1

Okay, but a theoretical framework is great. It doesn't really mean anything until it's tested against real consumers.

Real Surveys Put To The Test

Google Agent 2

Absolutely.

Google Agent 1

And this is where the source material gets incredibly practical. The team put this SSR method to the test using a massive data set of 57 real-world personal care product surveys that had already been conducted by a leading corporation.

Google Agent 2

Right. We're talking about over 9,300 actual human responses here.

Google Agent 1

Yeah, and this was not a simplified toy data set. These were complex, multi-layered product pitches that real marketing departments had designed.

Google Agent 2

Very complex.

Google Agent 1

Like one of the specific concepts they tested was for a product called R Foam Mood Infused Body Wash.

Google Agent 2

I love that name.

Google Agent 1

Right. It's pitched as this shower ritual utilizing mood-coated fragrance capsules. There is an energized variant with citrus and ginger and a calm variant with lavender and cedar.

Google Agent 2

Standard premium body wash stuff.

Google Agent 1

Yeah, the marketing copy boasts about clinically inspired neuroaroma blends and a skin-first, sulfate-free formula. It's the exact kind of dense, buzzword-heavy pitch you would see in a real focus group.

Google Agent 2

So the researchers took the ORAFOM concept, along with all 56 others, and fed them to their synthetic AI consumers using the SSR method.

Google Agent 1

And what happened?

Google Agent 2

The synthetic surveys achieved a 90% correlation with the maximum achievable human test-retest reliability.

Google Agent 1

Wait, 90%. So a 90% correlation means the synthetic panel ranked the appeal of these various products almost exactly the way the real human panels did.

Google Agent 2

Almost identically. And the shape of the data matched remarkably well, too. You remember earlier we noted that forcing a direct weighting just resulted in a massive spike of neutral threes.

Google Agent 1

Right, the politician answer.

Google Agent 2

Exactly. Well, when using the semantic mapping method, the statistical distribution of the AI's scores perfectly mirrored the humans. The researchers measured the overlap between the AI's distribution curve and the human's distribution curve, and they achieved a similarity score of over 0.85 out of one point ear. The AI was handing out the same proportion of high and low scores that the humans naturally did.

Age And Income Still Matter

Google Agent 1

Okay, I have to pause here and challenge the core premise for a second. Go for it. Because I can understand how an AI reads a description of citrus and ginger body wash and mathematically associates that with positive language in an embedding space. But human purchasing intent isn't just about whether a product sounds nice. Right. It is driven by our lived experience. It's driven by our age, our cultural background, and most importantly, our wallets. Does a massive server farm actually understand what it means to be a 50-year-old chopper or someone who is struggling to pay rents?

Google Agent 2

Well, the researchers actually anticipated that exact skepticism. So they tested the demographic variables really heavily. They didn't just ask a generic AI if it liked the body wash, they prompted the AI with highly specific demographic personas.

Google Agent 1

Including things like age and income.

Google Agent 2

Yes, age, income, location. And when they sliced the synthetic data by these demographics, the AI mirrored human realities with startling accuracy. But let's look at the age variable, for example. In the real human data, purchase intent for these specific personal care products followed a very distinct concave curve.

Google Agent 1

A concave meaning the interest levels go up in the middle and drop off at the ends.

Google Agent 2

Exactly. Middle-aged participants showed the highest interest in buying these products, while both the younger cohorts and the older cohorts rated their purchase intent significantly lower.

Google Agent 1

Okay.

Google Agent 2

When the researchers ran the synthetic panel and prompted the AI with different ages, the model perfectly recreated that exact same concave curve.

Google Agent 1

No way. Yeah.

Google Agent 2

The AI recognized that a 45-year-old persona is inherently more receptive to a neuroaroma body wash than a 20-year-old or a 75-year-old persona.

Google Agent 1

So it actually understands the generational appeal of the marketing copy.

Google Agent 2

It does.

Google Agent 1

But what about the financial aspect? Because wanting a premium mood-infused body wash is one thing. Actually, having the disposable income to buy it is another.

Google Agent 2

Right. And the human data set utilized six different income brackets. Now, several of the lower brackets included explicit indicators that the human participant was facing budgetary problems. Okay. The researchers took those lower income statements and fed them directly into the AI personas. And the AI accurately lowered its purchase intent across the board.

Google Agent 1

Really?

Google Agent 2

In one specific instance, when prompted with an income level statement that mentioned being financially in danger, the GPT-40 model reacted very sensitively and dropped its purchase intent drastically. It successfully modeled the behavioral impact of financial constraints.

Google Agent 1

So it's actively weighing the appeal of the advertisement against the financial anxiety of the persona its role-playing.

Google Agent 2

Exactly.

Google Agent 1

That is wild. And that leads to an obvious question about how this compares to the tools companies are already using, because corporations have been using predictive algorithms for years to guess what consumers will buy.

LLMs Versus Classic Machine Learning

Google Agent 1

Right, sure.

Google Agent 2

Machine learning has been in this space for a long time.

Google Agent 1

So how does this new AI method stack up against traditional machine learning?

Google Agent 2

Aaron Powell Well, the researchers decided to benchmark this zero-shot LLM approach against a traditional supervised machine learning model. They used a type of algorithm called a gradient-boosted decision tree, specifically Light GBM. This is a very standard, powerful tool in a data scientist's arsenal.

Google Agent 1

And a traditional machine learning model relies heavily on historical data, right? Like you feed it massive spreadsheets of past consumer behavior, it finds the mathematical patterns, and then it tries to predict future behavior.

Google Agent 2

Aaron Powell Exactly. And the researchers actually trained this traditional model on a portion of the real survey data. They fed the algorithm, the demographics, the product categories, the price tiers, and the historical human ratings.

Google Agent 1

So they gave it a massive head start.

Google Agent 2

Huge head start. They gave it the inside information on how humans had actually voted for similar products and then asked it to predict the outcomes for the rest of the surveys.

Google Agent 1

And how did it do?

Google Agent 2

The traditional ML model achieved a 65% correlation with the human results.

Google Agent 1

Aaron Powell Wait. The traditional algorithm had the answer key for half the test, and it still only hit 65%. But the large language model using the semantic method, which had never been trained on this specific survey data at all, hit 90%.

Google Agent 2

That's right.

Google Agent 1

How does a generalized model with zero specific training beat an algorithm that was purpose-built on the exact data set?

Google Agent 2

Aaron Powell It comes down to how these two systems perceive the world.

Google Agent 1

Yeah.

Google Agent 2

I mean, a traditional machine learning model just looks at columns on a spreadsheet.

Google Agent 1

Right.

Google Agent 2

It learns that column C, which is age 45, combined with column F, which is premium category, historically equals a higher purchase intent. But the algorithm has absolutely no idea what ARFOM actually is.

Google Agent 1

Aaron Powell Because it's just variables to the algorithm.

Google Agent 2

Exactly. It doesn't know what eucalyptus smells like. A large language model, on the other hand, has been trained on a massive chunk of the internet. It has ingested millions of beauty blogs, Reddit reviews, magazine articles, and social media posts.

Google Agent 1

Oh, so the LLM inherently understands the cultural context. Yes. It knows the societal cachet of sulfate-free skincare. It knows how a 45-year-old on a budget talks about self-care because it has literally read thousands of forum posts from 45-year-olds on a budget.

Google Agent 2

You got it. It compresses all of those aggregate human opinions into its neural network. Yeah. And that cultural context allows it to evaluate the semantic appeal of a new product in a way that a spreadsheet algorithm simply cannot.

Google Agent 1

That makes total sense.

Google Agent 2

And this deep understanding of language provides another really unexpected advantage. The

Better Feedback At Lower Cost

Google Agent 2

study proves that synthetic AI can mimic the quantitative data, you know, the statistical graphs and the one to five ratings.

Google Agent 1

Right.

Google Agent 2

But the AI actually outperforms human panels in the qualitative data.

Google Agent 1

Aaron Powell Wait, really?

Google Agent 2

It is significantly better at explaining why it gave a certain rating.

Google Agent 1

Oh, I believe that. Anyone who has ever run a human survey knows how frustrating the open-ended questions are. You ask a participant, what did you like about this concept? and they just type, it's good.

Google Agent 2

Right. Or they just copy and paste the product description. Like, I like that it has fragrance capsules.

Google Agent 1

Yes. It is completely hollow feedback.

Google Agent 2

Well, real human survey takers are notoriously lazy and they also suffer from positivity bias.

Google Agent 1

Positivity bias.

Google Agent 2

Yeah, unless they are having a terrible day. Most people don't want to be mean. They tend to be overly polite to the researchers running the survey. But synthetic consumers don't have feelings and they don't get tired. Because they are language models, generating rich text is their primary function.

Google Agent 1

So when asked to explain their purchase intent, the synthetic personas gave incredibly nuanced, detailed feedback in the distinct voices of their assigned demographics.

Google Agent 2

The quotes in the source material are fascinating.

Google Agent 1

Oh, yeah. They gave actual snarky pushback to the marketing copy. Like one of the AI personas evaluating a premium product, so he said, Seems kind of bougie.

Google Agent 2

I love that one.

Google Agent 1

And another persona pushed back on the clinical language saying, Sounds expensive, and I'm not sure I buy all that microbiome talk. I'll stick with what I know.

Google Agent 2

That kind of feedback is pure gold for a product developer. I mean, if your synthetic panel tells you that your microbiome messaging sounds too expensive and confusing, you know exactly what you need to adjust in your advertising copy before you launch the product.

Google Agent 1

You get a total gold mine of unvarnished focus group text explaining exactly what consumers hate or love about the concept, and without any of the typical human distortions.

Google Agent 2

Exactly. And all of this happens without spending thousands of dollars on recruitment or waiting six weeks for a market research firm to collate the data.

Google Agent 1

Aaron Powell You generate the product prompt, you use semantic similarity rating to mathematically map the AI's text to a numerical scale, and you have a highly accurate synthetic panel by the end of the day.

Google Agent 2

The barrier to entry for world-class consumer research is dropping dramatically. I mean, whether it is an entrepreneur testing a niche app idea or a product manager at a Fortune 500 company evaluating a new physical product line, the ability to rapidly test ideas against a simulated population changes the entire development cycle.

Google Agent 1

Because it allows creators to find out if an idea is going to resonate before they invest heavily in manufacturing.

Google Agent 2

It operates like a crystal ball for consumer intent.

The Unsettling Question About Taste

Google Agent 1

It really does. But you know, that crystal ball is built on a very specific foundation. And I think it leads us to a pretty fascinating philosophical place to wrap this up.

Google Agent 2

Oh, definitely.

Google Agent 1

Large language models don't have physical bodies. They don't have skin to wash with mood-infused body wash. They don't have actual bank accounts or mortgages. Everything these models know about being a 45-year-old on a budget who wants an energizing citrus shower, they learned from us.

Google Agent 2

They learned it by scraping our aggregate internet arguments, our product reviews, our personal blogs, and our social media posts. They are effectively a massive mirror, reflecting humanity's collective behaviors and preferences back at us.

Google Agent 1

And that mirror data is so precise and so deeply patterned that it can predict our exact purchasing constraints and desires with 90% accuracy. Which leaves us with a rather profound question to consider for everyone listening.

Google Agent 2

If a mathematical model trained entirely on the aggregate noise of the internet can so perfectly and predictably map out exactly what you will buy, exactly what you will find too expensive, and exactly what marketing buzzwords you will reject.

Google Agent 1

Are our individual personal tastes really as unique as we think they are?

Google Agent 2

Or when it comes down to it, are we all just following a highly predictable algorithm ourselves?

Google Agent 1

Think about that the next time you were standing in the store aisle, completely convinced that you are making a unique, spontaneous decision to buy that neuroaroma body wash. Maybe you aren't making a choice at all. Maybe the formula already knew you were going to put it in your cart. Leave them to ponder that until the next deep dive.