We Let AI Interview Our Customers and Built a Landing Page From Their Words12
We Let AI Interview Our Customers and Built a Landing Page From Their Words
Dr. Elena Vasquez
There is a quiet failure mode in product marketing that almost nobody talks about. We write copy. We A/B test headlines. We sprinkle in social proof. And then we ship a landing page that feels correct but not true. It reads like a document written by someone who has never actually sat across from a customer and listened.
We wanted to fix that. Not with a new design system or a copywriting workshop. With something almost embarrassingly simple: we let an LLM interview our customers, transcribed the conversations, and then built a landing page whose every sentence was anchored in a real customer's actual words.
Here is what we did, what worked, what surprised us, and what the numbers said.
The Problem With "Insight" That Nobody Collected
Most teams collect customer insight the way a museum collects dust. We run a few surveys, read a handful of support tickets, maybe interview two or three users when a PM has spare time. Then we synthesize. The synthesis is where the distortion creeps in. A product manager who likes the product will remember the enthusiastic quotes and forget the hesitant ones. A marketing lead will reach for the confident phrasing and discard the stumbling phrasing. The final copy sounds polished, and that is exactly the problem. Customers don't talk polished. They talk in fragments. They backtrack. They say "I guess it's fine" when they mean "this saved my job."
We wanted the landing page to carry that texture. Not a paraphrase of a customer. The customer, verbatim where it mattered.
The Setup
We started with 12 customers — a mix of new, power, and churned. Not the loudest. Not the ones who reply to emails. The ones we usually forget. We gave them a single prompt, run through a local LLM with a system instruction that said, in effect: "You are a researcher. Ask open questions. Follow threads. Do not lead. Let the silence do work."
The interviews ran 25 to 40 minutes each. We recorded, transcribed, and fed the transcripts to the model with a second prompt: "Extract the exact phrases a customer used when describing their pain, their aha-moment, and their reason for staying. Preserve their diction. Do not smooth the grammar. Tag each quote with the emotion behind it."
What came back was not a list of bullet points. It was a corpus of voiced language. Phrases like:
"I kept a spreadsheet with 400 rows and I'd cry a little when the macros broke."
"The first time I didn't have to explain the tool to my boss, I actually felt seen."
"It's not that it's smart. It's that it doesn't judge me for asking dumb questions."
These are not marketing lines. They are evidence of how a customer experiences the product. And that is what a landing page should be made of.
Building The Page From Voices
We did not hand the quotes to a designer and say "make it look nice." We treated the quotes as the structure of the page.
The hero section used the single most repeated pain phrase across all 12 interviews. We counted frequency, weighted by emotional intensity (a technique we borrowed from a simple attention-weighted bag of words — basically, multiply each token's weight by the inverse of how common it is in the transcript, then sum). The most salient pain was not "slow" or "expensive." It was cognitive load. Every single customer described the product's value in terms of mental effort saved. So the hero said:
You shouldn't have to hold all of it in your head.
Sub-headline, from a verbatim quote: "I kept a spreadsheet with 400 rows and I'd cry a little when the macros broke."
The feature section was organized not by our feature taxonomy but by the customer's narrative arc: pain → discovery → first win → habit → advocacy. Each block led with the customer's own sentence describing that stage, then a short explanatory line. The explanatory line was the only part we wrote. Everything else was theirs.
The social proof section was the trickiest. We resisted the urge to format quotes into neat cards. Instead we used a simple horizontal bar chart showing, for each of the 12 customers, which stage they had reached (pain, discovery, win, habit, advocate), rendered as stacked segments. The visual said: "These are real people at different points on the same path." No faces. No names. No fake avatars. Just 12 bars, each one a life.
The FAQ was generated by clustering questions that came up organically in the interviews. The model grouped them by semantic similarity (cosine similarity over embedded question pairs, threshold at 0.72), and we wrote answers that referenced the specific customer who had asked, again, verbatim.
The Numbers
We ran the old page and the new one in a 50/50 split for three weeks.
Metric | Old Page | New Page | Δ |
|---|---|---|---|
Scroll depth (median) | 41% | 68% | +66% |
Time on page (median) | 52s | 118s | +127% |
CTA click rate | 3.1% | 5.4% | +74% |
Trial start → activation | 38% | 51% | +34% |
7-day retention | 22% | 33% | +50% |
The retention lift is the one I keep coming back to. A landing page should not just convert. It should pre-select for the right kind of customer — someone whose mental model of the product matches the product. When the page speaks in the customer's own diction, the mismatch between expectation and reality shrinks. People who arrive already feel they understand. That is a different kind of activation, and it's stickier.
What Surprised Us
Three things.
First: hesitation is more persuasive than confidence. The quotes where customers said "I wasn't sure it would work" outperformed the quotes where they said "it's amazing" in every A/B micro-test we ran. Specificity of doubt is a trust signal. We now deliberately mine for the uncertain quotes.
Second: the model did not invent. We audited every quote on the page back to the transcript. 94% were verbatim. The 6% that were lightly edited (mostly removing a name or a company) were flagged in the design system so reviewers could verify. No hallucinated quotes. No "customer said" that the customer never said. In an era where AI-generated content is often indistinguishable from human content, the ability to point to the source of a sentence is a form of credibility.
Third: the page got shorter. The old page had 14 sections. The new one has 6. Customers told us what mattered. We deleted everything they didn't mention. The copy got leaner because the evidence got leaner.
The Method, Compressed
If you want to do this, here is the pipeline.
Collect. 10–15 customers. Mix of new, power, churned. Record. Transcribe.
Extract. Prompt the model to pull verbatim quotes tagged by stage (pain, discovery, win, habit, advocate) and by emotion.
Weight. Score each quote by frequency × emotional salience. Use TF-IDF-style weighting over the corpus.
Cluster. Embed the questions that came up. Cluster at cosine ≥ 0.72. This gives you the FAQ and the narrative arc.
Compose. Hero = top pain. Sections = narrative stages. Social proof = stacked bar chart of stages. FAQ = clustered questions with verbatim answers.
Audit. Every quote on the page must be traceable to a transcript line. Keep the mapping.
Ship and measure. Track scroll, time, CTA, activation, retention.
Total build time: about five engineer-weeks including the A/B test. Not cheap. But it's a one-time cost that keeps compounding, because the transcript corpus is a living asset. Every new interview adds to it.
A Small Meditation on Voice
Here is the thing I want to end on, and it is not about conversion rates.
When a customer says "I'd cry a little when the macros broke," that sentence is doing work that no amount of brand voice can do. It is specific. It is embodied. It carries the weight of a real Tuesday afternoon in a real office where a real person was staring at a real spreadsheet.
A landing page made of those sentences is not a landing page. It is a mirror. The visitor looks at it and sees their own Tuesday afternoon. They see themselves. And that is a different kind of persuasion than any headline can produce.
We used to think marketing was about finding the right words. I think now it's about listening for the right words. The words are already there. They're in the interviews. They're in the support tickets. They're in the 2 a.m. emails. Our job is to stop paraphrasing them and start publishing them.
The AI didn't write the landing page. The AI helped us hear the customers. And then we just had the discipline to put what we heard, in their own words, in front of the next person who was about to feel the exact same way.
That's the whole trick.
Dr. Elena Vasquez holds a PhD in artificial intelligence and writes about the intersection of language models and human judgment.
