Beyond the click: Juliana Jackson explains perception-led segmentation

The reason I wrote that article is because from an analytics and digital marketing perspective, we have been taught to use specific things like demographic data, correlation, and causation to sell our stuff online. As a company, you build goods, you want to sell them for more than it costs you to make them, and you want to profit. For a long time, the industry has been obsessed with clicks, time spent on page, and everything that happens on the interface. I would say that it did work for some time. But if we think about where we are right now as a society and as a culture, all the perceptions, thoughts, and impressions that people will have about a brand or a company will not happen in this very controlled environment.

I always like to say that the way a company builds its brand perception happens in decentralized places like Reddit, TikTok, YouTube, or whatever social media platform people use. As a company and as a brand, it would be very short-minded and shortsighted to focus only on what happens on your website and use click data or conversion data to come up with hypotheses and create campaigns. You can’t just create experiences on a website because you think that everybody that comes there is ready to buy.

The perception-led segmentation that I cover in my newsletter is about the fact that you still have the interface experience on the landing page or the product page, but before a user gets there, they come with some sort of cognitive dissonance, cognitive bias, and an already made-up perception.

The best and easiest way I can explain this is if I personally want to purchase something that I haven’t purchased before, I’m not going to get my information from the company website. I know that the company website is going to show me the most curated version of themselves, their products, and their reviews. Most people that search on Google right now are not ready to buy. I have data for this from SparkToro; Rand Fishkin’s research shows that the distribution of search intent in the US is 68 % informational, 11% navigational, and only 1% transactional.

What does this mean? It means that nearly 70% of the people that search on Google for something are just seeking information, answers, and validation. Yet, if you go on an e-commerce website, they treat you like you’re ready to convert right now. They push their products in your face, yelling “buy it, buy it, buy it,” with sales promotions and more. Every time you go on an e-commerce website, you’re just being bombarded by so many pop-ups, spin-the-wheels, and consent forms.

That’s why people will not take their information from there. They will look on TikTok, they will look on Instagram, they will go on Reddit. Everybody uses Reddit nowadays; that’s a key source of information. The reason people look in these places is because they are looking for helpful, people-first content. That’s not going to happen on a landing page with a cool carousel. Nobody gives a shit. Consumers don’t give a shit. It’s your job as a company to make them give a shit. The only way you can do that is by telling stories and positioning products and services within the lifestyle they could have.

All of these things that happen outside of the brand’s control are what I call “brand moments.” A TikTok video, a Reddit thread, a competitor’s podcast, a Trustpilot review. All of this stuff is outside of the brand’s control, it’s not linear, and it’s not easy to measure. These are the things that create different profiles of people. Those people that come from these brand moments will eventually come to your website, but they won’t compare you to what you have on the website. They’ll compare you to what somebody said on Reddit or what somebody said on TikTok. It’s very important to have this in mind when you optimize your e-commerce or SaaS website. You have to have in mind the baggage that people come with.

It’s research. There are ways to look at this. This is all unstructured data. It’s reviews, it’s videos, it’s content that is not in your control, but it’s still there. You can still take it, process it, transform it, normalize it, and analyze it. This is a very, very good use case for AI. However, I will say from experience that not all unstructured data types are made equally.

For instance, if you’re a mid-sized e-commerce company, you have Google Search Console activated. You can take all the search queries from there and use natural language processing to analyze them and see the patterns in the searches that get people to your website. Knowing that will give you an idea of where to improve your content. Should you have more lifestyle pictures? More information about your products? Better product pages or categories?

I can give you a great example. In the UK, there was this one query that was at the top when I was looking at this type of data: “Why are car batteries so expensive?” There was not one blog article from all the people that sell car parts in the UK that answered this question. It’s hilarious because car batteries are not even that expensive, but that was the top query. I remember suggesting to a friend that they should probably write an article about it because they would rank very high. This is one way you can look at this stuff.

Another one is to analyze customer reviews. I’ve been known to talk about this. Customer reviews are a very, very important data source that you can use natural language processing and machine learning to analyze at scale. You can get insights into how people perceive you after they’ve consumed your product. And unless they’re fake, reviews are from real people, which makes them a very good data source.

I was initially very excited about social media, but it’s very, very hard data to work with. When I say social media data, I’m not talking about the quantitative part like mentions and tags. I’m talking about the actual content: the comments, the videos. Blogs are okay, PR is okay, but when you step into the Instagram and TikTok of the world, people combine multimodal content like videos and images with text and metadata. There’s so much spam and use of hashtags that it becomes very dirty data.

You might see a post with a picture of a cat, but the caption is about a specific brand, and then you have hashtags about that brand and their mother and their sisters. An LLM will look at that one comment and say it’s about the brand, but when you as a human verify it, it has nothing to do with it. With this type of analysis, LLM models are still falling behind. I’m taking full responsibility for this: LLMs are not great at processing social media textual data for now. They lack your business context; they only have prior training data. They will use that prior training data, look at your comments, and give you very confident-sounding answers, but that doesn’t make them real.

This is why I would always choose to use a small language model (SLM) for this kind of task. The difference, besides the obvious, is that an SLM is trained on a very specialized set of data. It’s leaner and easier to use and fine-tune. You cannot truly fine-tune an LLM. When you fine-tune a small language model and serve it your specific data and vocabulary, it’s very easy to scale reliably. As an example, me and Krasimir Bambalov, a team lead data scientist who taught me so many things, built a pipeline with a small language model for Starbucks that has never crashed in the three years since we built it. We chose an SLM specifically because it’s very trustworthy.

Yes, if you Google “Juliana Jackson CLV lady,” you will find so much. I was working for a company that was selling a product doing retention and CLV analysis. At the time, I was very enamored with CLV, so I didn’t have critical thinking because I was in a product role and I thought our product was the shit. I was desperately talking about CLV. After I left that job, I took a disruptive strategy course at Harvard Business School and some corporate finance courses. I’m weird, okay, make sure you put that in the article (so, yes, we added it in the article).

CLV is a backward-looking metric masquerading as a forward-looking one. I learned this from the Harvard course. When you think about customer lifetime value, it’s based on historical behavior that is projected into an uncertain future. So what teams do is they end up optimizing for what customers did in the past, rather than what they will do. This creates very, very perverse incentives. Teams end up gaming CLV by extending payback periods, basically saying, “Well, our CLV is so high, we can afford a higher acquisition cost.”

The problem is what people mean by “acquisition cost.” It’s not just how much you spend on ads. It’s the salaries, the overhead, the tools, all the money you spend, even the power bill or the janitor. It’s everything. So teams end up cherry-picking different cohorts or time windows, or they manipulate discount rates to make the numbers look better. Or they focus only on retention versus acquisition, which is what I was wrongly promoting to people at the time. You have to be honest.

Customer lifetime value is a measurement hell because every team calculates it differently. Marketing uses different attribution windows than finance, then you have product teams that look at different costs. Everybody has a different definition of “lifetime.” What the fuck is lifetime? You end up with five different CLV numbers in the same company. My issue is that CLV on its own is seen as this North Star metric, which is wrong. It obscures actionable insights because it’s too aggregated. It’s a mess; it’s an average. You cannot optimize for specific things. A good North Star metric should be actionable, impossible to game, and unambiguous. CLV fails on all of those things.

It depends on the business model, but you should look at contribution margins based on cohorts. You look at time to positive unit economics. What are your unit economics? That’s my question. Product engagement, revenue per customer, payback period, these are very important. The real answer is it depends on who’s looking, because that’s the actual problem. CLV means different things to different people and different stakeholders, which makes it useless as an alignment tool. That was my biggest aha moment after being the CLV lady.

It’s very hard to answer what I have done because I’ve been doing a lot of things. I always say I do stuff on the internet. I started in sales, ended up in product, then product marketing, then I decided I was going to do data, but then I got bored. Now I’m doing a mix of marketing and data, and I’ve been doing data science for the last two and a half years. I’m like a mouton à cinq pattes, the five-legged sheep of the industry.

I think life teaches you to do stuff. I’ve met a lot of people in my life that influenced me in a positive way. I ended up in product because I was selling my marketing and sales services to a company in Romania, and I ended up being their first product owner. I was working only with developers back then, and I had finished computer science in high school, so I knew what they were doing in theory. I got along so well with them because they were gamers playing StarCraft, like me. They were analytical and technical, which I like. They didn’t have this product role, and I was reading about product management, so I kind of created this role for myself. Granted, I had no idea what I was doing, but I learned on my own. I was spending time on forums, talking to people, collecting feedback, and creating user stories. I didn’t even know they were called user stories back then.

I have always been good at writing since high school. I liked writing long essays, poems, and stories. But when I started working at CXL, one of my first assignments was to write an article about product marketing. Peep Laja made me rewrite the article maybe 20 times. He said something to me that changed everything: “If it’s not excellent, don’t do it. There are a lot of people that do mediocre shit on this planet. Do you want to be one of those, or do you want to write excellent stuff?”

After that, I spent a lot of time studying Ogilvy, copywriting, and how to structure information. I was inspired by people like Chris Walker and Dave Gerhardt who were constantly writing on LinkedIn. I noticed there weren’t a lot of women, especially minorities like me, doing it. So I said, “You know what? Why not?” In the beginning, I sucked. I cringe when I look at my old articles. But my best writing has happened in the last two years because I’ve spent a lot of time studying linguistics to work with natural language processing. When you understand the structure of sentences from a technical perspective, it sings differently when you write.

The reason I still write after five years is that it’s therapeutic for me. My process is making a big cup of coffee, putting on some music, and before I write something, I think about it for weeks. When I sit down, if I don’t feel it in the first two hours, I’m not going to write about that topic. But the biggest ROI is the relationships and the people I’ve met. My biggest strategy is meeting people, finding what’s best about that person, and seeing if I can add that to my skills. I learn constantly from other people.

AI is so important for research. It makes my life easier because one of my hobbies, which is very lame, is that I like to read science papers a lot. Now, I can use a tool like NotebookLM, upload five or six academic papers on the same topic, ask questions, and create a mind map. It’s easier for me to digest information. I have ADHD, and in the last three years, it’s been very hard for me to read books; I get bored very fast. So I use AI a lot for parsing these papers and getting shorter, more snackable versions. I also use Claude a lot for data visualizations because it’s so good at Python. AI is a big part of my writing, not in the sense that I use it to write, but I use it to research and brainstorm. Oh, and also I might use it to summarize because I take the Mark Twain approach to writing somethings “I didn’t have time to write a short letter, so I wrote a long one instead”, meaning I tend to write VERY VERY long copy and sometimes I get lucky with GPT 4.5 and it helps me summarize…or it makes it a big mess that I end up sending the whole shebang I wrote initially.