Latest / The DTC Marketing Podcast with Fexingo: Direct-to-Consumer Brands, Shopify, and E-Commerce Growth / How DTC Brands Use AI for Customer Reviews Analysis
Transcript
- Lucas: So you've just launched a new product — a mattress, a sneaker, a face cream — and the first hundred reviews come in. You read maybe ten of them, see that the average rating is 4.3 stars, and call it a day. Luna: Yeah, that's basically the default. And you'd be missing a ton. Lucas: Right. Because those reviews — the full text, not just the star count — are probably the richest dataset a DTC brand has access to. And most brands treat them like a report card rather than a R&D department. Luna: So what's the alternative? Actually reading every single one? Lucas: I mean, you could, but that doesn't scale. The alternative is AI — specifically natural language processing — to analyze review text at volume and surface patterns that a human would miss. And there's a really good case study I want to walk through. Luna: Which brand? Lucas: Brooklyn Bedding. They're a DTC mattress company — they make the Brooklyn Bedding Signature Hybrid, among others. A few years ago they were getting solid reviews, around 4.5 stars average. But returns were still running at about 15 percent, which is fairly typical in the mattress space but costly. Luna: Right, because mattresses are bulky and expensive to ship back. Lucas: Exactly. So they decided to take their last 10,000 reviews — the actual text — and run them through a sentiment analysis tool called Thematic. What they found was that the word 'firm' appeared in 23 percent of reviews, but the sentiment around it was split. About half the people who mentioned 'firm' loved it. The other half said it was 'too firm' or 'uncomfortably firm'. Luna: So it wasn't that the mattress was bad — it was that one specific attribute was polarizing. Lucas: Right. And they could have missed that if they just looked at the overall rating. Because the people who found it too firm still gave it three or four stars — they didn't hate it, they just didn't love it. But the return rate among that subgroup was much higher. Luna: So what did they do? Lucas: They reformulated the comfort layer — the top two inches of foam — to be slightly softer. And then they tracked the reviews of the new version. Within six months, the mentions of 'too firm' dropped by almost 40 percent, and returns overall fell from about 15 percent to just over 12 percent. Luna: That's a 20 percent reduction in returns for one product change. That's huge for margin. Lucas: Exactly. And the cost? They already owned the data. The AI analysis was maybe a few thousand dollars using a tool like Thematic or Sift. Compare that to the cost of a focus group or a survey panel — which might take weeks and still not give you that kind of granularity. Luna: But isn't there a risk that the AI misinterprets things? Like, sarcasm or slang? Lucas: It's a real issue. Most NLP models are trained on general text — news articles, social media. Customer reviews have their own quirks. 'This pillow is the bomb' could be positive or sarcastic depending on context. That's why you need to train or fine-tune the model on your own dataset. Thematic, for example, lets you create custom categories. Luna: So it's not just plug and play. Lucas: No, but it's not that hard either. You essentially feed it a sample of your reviews, tag a few hundred manually — like 'this is about comfort', 'this mentions durability', 'this is a complaint about shipping' — and the model learns. Then you run it on the rest. Luna: What about smaller brands? If you're a DTC brand doing maybe a thousand orders a month, do you have enough data for this to work? Lucas: It depends. You need at least a few hundred reviews to get meaningful patterns. But a thousand orders a month — if even 10 percent leave a review, that's a hundred reviews a month. Over a year, you've got over a thousand. That's enough to start. Luna: And if you don't have the budget for a tool like Thematic? Lucas: Then you can do a lighter version with ChatGPT. Export your reviews to a spreadsheet, give it to ChatGPT with a prompt like 'Analyze these reviews and tell me the top three recurring themes about comfort.' It's not as structured, but it works. I've seen brands get real value from that. Luna: Interesting. So the barrier to entry is actually pretty low. Lucas: Yeah. And the payoff can be enormous. Because reviews are direct customer voice — unsolicited, unfiltered. Surveys suffer from recency bias and social desirability bias. People say what they think you want to hear. Reviews — especially bad ones — are brutally honest. Luna: Which brings up another angle: using AI to detect fake reviews. That's a whole other use case. Lucas: Absolutely. Fake reviews are a plague, especially on Amazon. But even on your own site, competitors or disgruntled ex-employees can leave bogus negative reviews. There are AI tools — Sift has a fraud detection module — that flag patterns like multiple reviews from the same IP address, or reviews that use very similar phrasing. Luna: So it's not just about understanding your customers, it's about protecting your data integrity. Lucas: Exactly. Once you start treating reviews as a strategic asset, you need to make sure the asset isn't polluted. One mattress brand — I think it was Saatva — found that about 6 percent of their negative reviews were actually from people who had never bought the product. They removed those, and their average rating went up by a tenth of a point. Luna: Every tenth of a point matters when you're advertising that rating. Lucas: It does. And it also helps you make better decisions. If you're reformulating a product based on review analysis, you want to be sure the data is real. Luna: Let's talk about the limitations. What about brands that get very few reviews — like a luxury, low-volume product? Lucas: That's a real challenge. If you're selling a $5,000 watch and you only get 50 reviews a year, the sample is too small for meaningful NLP. In that case, you might need to supplement with other data — customer service transcripts, return reason codes, maybe even social media mentions. Luna: But for most DTC brands, especially in categories like apparel, beauty, and home goods, the volume is there. Lucas: It is. And I'd argue that even if you're small, the cost of not doing it is higher than you think. Every return you could have prevented is margin you left on the table. Luna: Right, because returns aren't just the refund — they're shipping, restocking, and the lost sale. Lucas: Exactly. So let me give you a framework for getting started. Step one: export your last 500 to 1,000 reviews. Step two: pick a tool — either a dedicated platform like Thematic or Sift, or a DIY approach with ChatGPT. Step three: identify your top three themes, both positive and negative. Step four: prioritize the negative themes that correlate with high return rates. Step five: test a product change and measure the impact. Luna: That's actionable. And it doesn't require a data science team. Lucas: Not at all. And that's the thing — this is one of those rare marketing moves that can directly affect the product itself. It's not just about messaging or ad targeting. It's about making what you sell better. Luna: Speaking of making things better — a quick note. We keep this podcast ad-free, and the only reason we can is because of listeners who support us directly. If you've gotten something out of these conversations — maybe you've even used one of these frameworks — a couple of dollars a month genuinely makes a difference. It's buy me a coffee dot com slash fexingo. And it really does keep us going. Lucas: Yeah, small amounts add up. And it keeps us independent — no sponsors to please. Luna: Exactly. So back to review analysis — one thing I want to ask: do you think this will become table stakes for DTC brands in the next few years? Lucas: I think it already is for the smart ones. The brands that treat reviews as just a vanity metric are going to get outcompeted by the ones that mine them for product intelligence. It's like A/B testing — ten years ago it was a competitive advantage, now it's just expected. Luna: So the question isn't whether you should do it, it's how fast you can start. Lucas: Right. And the data is already sitting there in your Shopify or WooCommerce backend. You just have to unlock it.