How Many Hands Has This Quote Passed Through?

Somewhere in your evidence folder sit two documents. One is a raw customer interview transcript — forty minutes of a real buyer describing, in their own words, the problem that made them go looking for you. The other is a polished pitch deck with a glowing testimonial on slide nine. Most teams, and almost every AI tool, treat these two documents as equals. Drop them in the folder, point the model at it, generate.

They are not equals. The testimonial on slide nine is a quote that has already survived a marketing filter. Someone selected it, trimmed it, and placed it there because it persuades. It tells you something real — but mostly about the marketer who chose it, not about the buyer who said it.

We call the missing measurement evidentiary directness: how many hands has this quote passed through before it reached your evidence base? It turns out to be one of the most useful questions you can ask about marketing evidence — and one that almost nobody asks systematically.

Three questions, not one

When you evaluate a piece of GTM evidence, there are actually three independent questions in play, and they are easy to blur together:

Where is it in the pipeline? Raw capture, reviewed, distilled into a conviction? That's a stage question.

How good is it? Substantive, confirmed by a human, fresh? That's a quality question.

How directly did the buyer speak? That's the directness question, and it is orthogonal to the other two. A pitch deck can be high quality — well argued, current, accurate — and still be three removes from the buyer's mouth. A messy support ticket can be low polish and still be the buyer speaking with nobody selling.

Inside T2D3 OS these are three separate axes on every document. The third one exists precisely because the first two can't capture it: in our quality grading, a customer interview and an uploaded sales deck that both pass human triage land on the same tier, with the same weight. Directness is the correction that separates them.

The four classes

Every document type falls into one of four directness classes, ordered by remove from the buyer's own mouth.

Direct buyer voice. The buyer speaks, unmediated, captured verbatim. Customer and prospect interviews. Support tickets and support calls — a customer describing their own problem in their own words, with nobody selling. Survey verbatims. This is the raw material messaging should be built from, because nobody has edited it for persuasion yet.

Proximate. The buyer speaks, but the seller is in the room steering the conversation — sales call transcripts, product demos, onboarding calls. Or an insider stands in for the buyer: a team interview with someone who has actually held the persona's job. Real buyer language leaks through, but the setting shapes it. A prospect on a demo asks "can it do X?" — that objection is gold, but it was voiced inside a conversation the seller framed.

Mediated. Words already selected and polished to persuade. Pitch decks, website copy, advertising, competitor collateral, filled-in strategy templates. This is where testimonials live — which surprises people, until they remember how testimonials get made. Mediated material is still evidence; it's just evidence about positioning choices as much as about buyers.

Reported. A third party summarizing other people. Analyst reports, press articles, market research. Useful for the shape of a category; weakest for how any actual buyer talks.

And one deliberate rule: a document of unknown provenance is treated as the weakest class when you ask "is this the buyer speaking?" — but it is not penalized in weighting. Unknown means unknown. We only demote what we positively know is at a remove; we never punish a document just because it hasn't been classified yet. That distinction sounds small. Across a corpus mid-ingest, it's the difference between a fair system and one that quietly buries most of your library.

What this looks like when the software enforces it

Here is where we'll say the immodest part plainly, because it's the part we think more AI tools should copy.

In T2D3 OS, directness isn't a tagging convention someone remembers to apply. Every document type in the system carries a declared directness class, and the map is total: if we add a new document type to the product without classifying its directness, our own build fails. A new type of evidence cannot quietly default to "as good as a customer interview." The compiler — not a reviewer's memory — is what guarantees that a testimonial never impersonates an interview.

The classes then act as weight multipliers when evidence is assembled for any AI generation. Direct buyer voice carries full weight. Proximate carries 0.85. Mediated drops to 0.6, reported to 0.5 — and those defaults are tunable live by an administrator, because a reasonable team might weight analyst reports differently in a category where analysts genuinely move deals. When the system distills several documents into a derived artifact, the artifact inherits the strongest directness among its grounding — so an insight traced to a real interview keeps that provenance as it travels downstream.

The result: when our AI drafts a messaging framework or an ICP, the buyer's own words systematically out-vote the marketing department's paraphrase of them. Not because a prompt asks nicely, but because the arithmetic of evidence selection is built that way.

How to apply this without our software

You don't need T2D3 OS to use the idea. Three practices carry most of the value:

Audit the folder. Take your current "voice of customer" collection and label every item with one of the four classes. Most teams discover the same uncomfortable ratio: the folder is dominated by mediated material — decks, case studies, website copy — with a thin sliver of direct buyer voice. That ratio is your real exposure, because whatever writes your messaging (an agency, an intern, a model) will reproduce the mix it's fed.

Chase directness deliberately. The cheapest direct-voice sources are usually already inside the building: support tickets, sales-call recordings you're allowed to transcribe, survey open-text answers, onboarding calls. A single genuine prospect interview is worth a folder of polished collateral when the question is how buyers actually talk.

Quote at the source, not the summary. When you cite customer language in messaging work, trace it to the least-mediated version you hold. If the same claim exists in a case study and in the raw interview behind that case study, use the interview. The case study version has already been rounded toward what marketing wanted to hear.

Why this matters more in the AI era, not less

Generative AI made producing marketing prose nearly free — which means the scarce input is no longer writing, it's evidence about your specific buyer that your competitors don't have. That evidence is only as valuable as it is direct. Feed a model a corpus of mediated material and it will confidently generate more mediated-sounding material: smooth, plausible, and interchangeable with every competitor doing the same thing.

The teams that win the next few years of B2B marketing will be the ones that treat first-hand buyer language as a managed asset — captured close to the source, graded honestly for how many hands it passed through, and weighted accordingly every time a machine or a human writes on the company's behalf.

If you're curious what a system that takes this seriously says about your own go-to-market, our free GTM diagnostic at t2d3.pro reads your public messaging the way we read evidence: with the filters labeled. And if you want the whole operating system behind it, the founding member program is where to start.