Taste Is Not Evidence: Why Style References Must Be Hard-Excluded From Your ICP Grounding
Why style references and buyer evidence must never share a lane in an AI marketing system.
Stijn Hendrikse · Sep 3, 2026
Last updated 20 August 2026
Style references must be hard-excluded from ICP grounding at the database query, not the prompt. Tag every document with a signal role — buyer evidence, internal artifact, or style exemplar — and block the exemplar role at retrieval, so a blog post you admire is structurally incapable of ever being returned as proof of what your buyers said.
You saved that Stripe blog post because the rhythm was perfect. You clipped the Linear changelog for its restraint. You dropped a competitor's category essay into your workspace because the structure was worth stealing.
Then you asked your AI to draft positioning for a client — and it came back sounding like a category essay written by someone else's marketing team, arguing someone else's point of view, at someone else's stage of maturity.
That is not a prompting failure. That is a retrieval failure. Your taste library and your buyer evidence went into the same folder, and the model cannot tell the difference between what our buyers said and what we wish we sounded like.
The fix in one sentence: give every document a signal role, mark style references as style_exemplar, and exclude that role at the database query — not at the prompt — so a style reference is structurally incapable of ever being retrieved as buyer proof.
Here is how to build a taste library that makes your output better instead of quietly corrupting it.
Mixing the two lanes corrupts both, not just one
Most teams assume the risk runs one direction: admired content pollutes the evidence base. It runs both ways.
Evidence contaminated by taste. Your ICP grounding is supposed to answer what do our buyers actually say. When an admired blog post sits in the same retrieval pool as a customer interview transcript, the blog post usually wins. It is longer, cleaner, better structured, and more recent. The messy, specific, first-party thing — the transcript where a fractional CMO says she is logged into four different CRMs and four different Slack workspaces — gets outranked by a polished essay about GTM strategy. You end up positioning against a competitor's argument instead of your buyer's pain. In our own buyer research, tool sprawl of exactly that kind carries an estimated $120,000 annual impact on a fractional practice — a number that only shows up in transcripts, never in a category essay.
Taste contaminated by evidence. Run it the other way and it is just as bad. If your voice model reads raw customer transcripts as style references, your brand voice becomes an averaged impression of however your buyers happen to talk. That is useful data. It is a terrible target. As one of our own working notes puts it: "the category blog post is somebody else's positioning wearing an educational costume" — and a sales-call transcript is nobody's voice at all.
The two lanes answer different questions. Keeping them separate is not tidiness. It is the only way either answer stays true.
Build the taste library in five steps
Building a taste library that stays out of the evidence lane is the practical part of this. Do it once per engagement, then maintain it. Five steps, and the whole thing rests on step two.
1. Declare the signal role at upload, not at retrieval.
Every document gets a role before it enters the system: buyer evidence, internal artifact, or style exemplar. Interview transcripts, sales calls, support threads, win/loss notes → evidence. Your own decks and one-pagers → internal artifact (they tell you what your team wanted to say, which is not the same as what buyers heard). Blogs you admire, competitor essays, visual references, newsletters with the cadence you want → style_exemplar.
2. Exclude at the query, not in the prompt.
This is the whole design. A prompt instruction like "do not use style references as evidence" is a convention, and conventions decay the moment someone writes a new prompt. In T2D3 OS, style_exemplar documents are excluded at the database query for evidence retrieval, dropped at ranking, and defaulted away in chunk search. Three layers, none of them dependent on anyone remembering. The lanes cannot cross by construction.
3. Annotate what you actually admire. "I like this post" is not a usable style signal. Write one line per exemplar naming the specific attribute: short paragraphs, no adjective stacking, opens with the cost not the context. This is what a voice layer can act on. Vague admiration produces vague imitation.
4. Keep the library small and opinionated. Ten well-annotated exemplars beat sixty saved links. A taste library is a point of view, not an archive. If two exemplars pull in opposite directions, pick one — the contradiction will show up as fractured output.
5. Audit the lane quarterly. Open your evidence pool and scan for anything that is really someone's marketing. Category blog posts, analyst summaries, competitor case studies. Reclassify them as exemplars or delete them. Every one you find has been silently voting in your positioning.
What you get back
A clean split between the evidence lane and the taste lane gives you two things that were previously fighting each other — and it matters most for operators running several engagements at once, where roughly 50 to 60 percent of working time already goes to navigating and stitching together disconnected AI tools instead of shipping billable deliverables.
- Positioning that argues from your buyers. When the evidence lane contains only first-party material, ICP and value-prop outputs cite the transcript, not the essay. The distinctive, specific claim survives instead of being smoothed into the category average.
- A voice that is a deliberate choice. Your style target becomes something you selected, annotated, and can change — rather than an accident of whatever happened to be longest in the folder.
- Corrections that compound. As our own product notes describe it: "when you thumbs-down a stale deck or confirm a great transcript, you aren't tuning one output — you are re-weighting the evidence for every future generation across the whole system." That only works if the pool is actually evidence.
One question decides whether a document is evidence or taste
Before any document enters your system, ask: am I keeping this because of what it proves, or because of how it reads?
If the answer is "how it reads," it is a style exemplar — and it should be structurally unable to ever answer a question about your buyer.
Convention says keep them apart. Construction makes it impossible to fail. Only one of those survives contact with a busy quarter.