The citation floor: why two quotes beat one in customer interview synthesis

Run twenty customer calls through any modern tool and you will get a clean list of themes, each one footnoted with a quote. NotebookLM does it. Your coding agent does it. Klyr does it too. So if cited synthesis is everywhere, why does Klyr refuse to show you a theme backed by only one quote? Because customer interview synthesis doesn't fail at the citation step. It fails at the moment a single vivid complaint gets promoted to a pattern, and nobody notices the sample size was one.

Citations are table stakes now

Let's be honest about the landscape. Pasting transcripts into NotebookLM and asking for themes-with-sources is genuinely good, and it's free. Coding agents will happily summarize a research folder and link back to the lines they used. The link-to-source pattern is no longer a differentiator. If a tool's whole pitch is "every claim is cited," it's selling you 2023.

The problem citations were supposed to solve is fabrication: did the model invent this, or is it really in the transcript? That's worth solving. But it's the easy half. A citation proves a quote exists. It says nothing about whether one quote is enough to act on.

The citation floor: why two, not one

Klyr enforces a hard rule. Any theme supported by fewer than two verified source quotes is dropped before you ever see it. We call it the citation floor, and it is deliberately blunt. One customer saying a thing is an anecdote. Two independent customers saying the same thing, in their own words, is the faint beginning of a pattern. The floor is where we draw the line between "someone felt this" and "this might be real."

Two is not magic. It won't make a finding statistically significant, and we don't pretend it does. What it does is kill the single most common failure mode in customer interview synthesis: the n-of-1 theme: the articulate user who gives you a perfect, quotable rant about onboarding, and suddenly "onboarding friction" is a top theme, built entirely on one person who happened to be good with words.

The discipline matters more than the citations

Here is the part that's easy to miss. The value isn't the two quotes sitting under the theme. It's the themes that never made it onto the page. The floor's real output is a deletion. Every weak theme it removes is one fewer thing your roadmap has to fight off later.

A citation is a receipt. It tells you a claim is traceable. A floor is a filter. It tells you a claim cleared a bar before it earned your attention. Receipts are cheap. Bars are the discipline. Anyone can attach a quote to a sentence. Refusing to surface the sentence until it has two is the harder, more useful move, because it costs you findings. Good filters always cost you something.

What a vivid quote hides

Single-source themes are dangerous precisely because they're persuasive. The best quote in your dataset is usually an outlier. It is the most extreme, most fluent version of a feeling. It reads like signal. It feels like the thing to fix. And it pulls a disproportionate share of attention in the meeting, because nobody argues with a customer's exact words on a slide.

The floor is a structural defense against your own confirmation bias. You can't lean on a quote that quietly got filtered out for lack of corroboration. That's the point. The discipline protects you from the finding you most wanted to be true.

Where most synthesis tools stop, and the loop begins

Cited themes are an endpoint for most tools. You get the report, you screenshot it, the evidence goes stale in a doc. The floor only earns its keep because of what happens after it. In Klyr, themes that survive feed a ranked Feature Bet, which becomes a Build Pack, a PRD, acceptance criteria, and a prompt you can hand to Cursor or Claude Code. Then the outcome of that bet gets written back into Product Memory.

So the floor isn't a one-time gate on a static report. It's the entry condition for a graph that compounds. A theme that clears two citations today becomes a node you can revisit when the next round of interviews either reinforces it or quietly contradicts it. That's the loop, and it's where the moat actually lives. You can see a live cited report or compare it against NotebookLM if you want the contrast. Your coding agent forgets every session. The evidence-to-outcome graph is the memory it loses.

Honest limits of a blunt rule

A floor of two is a heuristic, and heuristics have edges. Sometimes one customer is right and the herd is wrong: a sharp early signal from a single power user can be exactly the thing worth chasing. The floor will drop that, and you should keep a way to override it when your judgment says so. We'd rather lose the occasional true single-source insight than let the dozens of false ones through, because in practice the false positives vastly outnumber the prophets.

Two is also a floor, not a ceiling. Two quotes from two people in the same five-person team isn't independence; it's one opinion wearing two hats. The number is a starting discipline, not the end of judgment. Treat it as the bar a theme must clear to deserve a conversation, not as proof the conversation is settled.

The takeaway

Citations tell you a claim is real. A floor tells you a claim is worth your roadmap's attention. The first is table stakes; every decent tool does it. The second is a choice to delete your weakest findings before they cost you a quarter, and then to remember what survived. If you want customer interview synthesis that argues with you instead of flattering you, start with the floor and follow it into the loop. Start free and watch what gets dropped.