Glints: Where Original POV Comes From When AI Can Derive Everything Else

Ask an AI to write your next thought-leadership piece and it will do something specific: it will derive. It will take everything already written in your category, find the center of mass, and express it fluently. The result reads fine. It is also, by construction, the average of what already exists — and the average of what already exists is precisely the thing nobody needs another article about.

This is the quiet crisis in B2B content right now. Derivation has become free. Any team with a subscription can produce unlimited competent-sounding prose derived from the public record. Which means the value of derivable content is collapsing toward zero, and the value of the one thing AI cannot do — originate — is climbing. The scarce input is no longer writing capacity. It is the insight that cannot be derived from what exists.

In the T2D3 methodology we have a name for that insight: a glint. And in T2D3 OS we made a deliberate, slightly unusual architectural decision about where glints live. This piece is about both — what a glint is and how to harvest yours, and why our system keeps them in a separate lane that buyer evidence can never touch.

What a glint is

A glint is the interesting thing that doesn't fit. The contrarian take a founder dropped in a meeting that made everyone pause. The customer's offhand remark that contradicts your category's conventional wisdom. The internal memo arguing something your industry considers settled. The surprising pattern one of your salespeople noticed that no analyst has written up.

Glints share three properties. They are original — you cannot get them by summarizing what's already published. They are uncomfortable — they usually cut against consensus, including sometimes your own positioning. And they are human — they come from lived contact with customers, markets, and products, not from recombination of text.

We've written before about why your customers' unwritten knowledge is the last durable moat in B2B SaaS. Glints are the internal counterpart of that idea: not what your customers know that AI can't read, but what your team believes that nobody else has said. Together they are the only two raw materials for content that a competitor with the same AI tools cannot replicate.

The averaging problem

Here is the trap most teams fall into: they treat all their knowledge as one pile. Interview transcripts, strategy docs, hot takes, meeting notes — everything goes into the same folder, or the same RAG index, and gets used to ground everything else.

That feels efficient. It is actually destructive, in both directions.

In one direction, opinions contaminate evidence. Your ICP definition, your personas, your value propositions should be grounded in what buyers actually said and did — verifiable, first-party evidence. Feed a contrarian internal manifesto into that grounding and you get strategy that confirms your own beliefs back to you with borrowed confidence. The AI will happily cite your opinion as if it were market data.

In the other direction — and this is the one almost nobody notices — evidence processing destroys opinions. Every serious evidence pipeline is built to dedupe, cluster, reconcile, and weight toward consensus. That is exactly what you want for buyer evidence, and exactly what kills a glint. Run "the interesting thing that doesn't fit" through machinery designed to find what fits, and it gets averaged into the middle or filtered out as an outlier. The most valuable sentence in your company quietly disappears into a theme cluster.

The fourth lane

T2D3 OS resolves this with a structure we call the fourth lane. When a document enters your signal library, a human triages it: signal (buyer evidence), noise (out, forever), or — the interesting case — glint: voted out of evidence, but kept.

The distinction matters. Waste is filtered out of everything permanently. A glint is filtered out of evidence and preserved as raw material, with the reviewer's one-line reason for the vote attached — which is often the most interesting sentence in the record, because it names exactly why this take is unusual.

Then comes the part we consider genuinely different: the wall. Glints are excluded everywhere buyer evidence is assembled — ICP grounding, persona synthesis, value-prop validation, quality scoring — and they are consumed at exactly three seams in the entire system:

1. Content surfaces, where original point of view is the whole point.
2. Contrarian ideation, the posture our content engine uses when it hunts for novel-POV pieces rather than how-to material.
3. Brand voice, where your distinctive stances shape how you sound.

Three doors. Nowhere else. And the wall is structural, not procedural: glints come back from the system as their own lightweight type that cannot be cited, weighted, or anchored the way evidence can. It is not that we ask the AI politely not to use opinions as market data — the plumbing makes the leak impossible. Even the excerpt is capped short by design, because a glint is provocation for a writer, not source material for a strategist.

The payoff runs both directions at once. Your strategy stays honest, grounded only in what buyers actually demonstrated. And your content stays sharp, because the system hands its writers a curated shelf of your company's most interesting unpublished thoughts — pre-sorted, with the reason each one was flagged as remarkable.

How to harvest glints — with or without our software

The lane is ours; the discipline is portable. Four practices:

Create the third bucket. Most triage is binary: useful or junk. Add the third verdict — "not evidence, but interesting" — to whatever review process you run. The moment reviewers have somewhere to put the outlier, they stop discarding it. This is the highest-leverage change and it costs nothing.

Record why it doesn't fit. When someone flags a glint, make them write one sentence: what consensus does this contradict? That sentence is usually the thesis of a future article. A glint without the reason is a bookmark; with it, it's a brief.

Protect glints from your own summarization. If you run AI summaries over meeting notes or interview archives, understand that summarization is consensus-seeking — it will erase your outliers first. Pull glints out before the pile gets summarized, not after.

Spend glints deliberately. A glint is only worth something published. Route the shelf to whoever writes your thought leadership, and make "which glint is this anchored to?" a standing question for any piece that claims a point of view. In our own pipeline, an idea claiming novelty must be anchored to specific flagged insights before it can ship — an unanchored hot take is exactly the derivable content the market is drowning in.

The economics underneath

Step back and the structure reflects a bet about where B2B marketing is going. As answer engines become the discovery layer, two kinds of content will matter: content machines can verify (grounded, specific, citable) and content machines could not have generated (original, contrarian, human). The mushy middle — fluent derivation — is already worthless, and getting more worthless every quarter as the cost of producing it approaches zero.

Both surviving kinds need infrastructure. Verifiable content needs an evidence base with real provenance. Original content needs a place where the underivable thing is caught, named, and kept safe from the averaging machinery. Most companies have neither; they have a folder.

A glint is small — a sentence in a meeting, a heresy in a memo. But it is the one input in your entire content operation that no competitor can buy, no model can derive, and no amount of tooling can substitute for. The least a system can do is notice when one appears, and refuse to let it be averaged away.

If you're curious what a signal library with lanes looks like in practice — evidence graded and grounded, glints caught and kept — you can see it running at t2d3.pro, where we use T2D3 OS to run T2D3's own marketing, this article included.