"Make the list longer" is the one effort in prospecting that reliably does nothing

A profile is a classifier, and a trait carries information only when its rate differs between the two classes; equally high on both sides means zero discriminating power. Meanwhile people write profiles by recalling the customers who bought — sampling positive cases and never the negative ones is a default cognitive strategy that has been measured over and over, and it grows exactly the kind of profile that is wide, comfortable, and unable to separate anything.

• Hard evidence · people only generate examples that fit their hypothesis: in Wason's (1960, Quarterly Journal of Experimental Psychology) 2-4-6 task, subjects guessing a number-sequence rule almost exclusively proposed sequences consistent with their current guess and rarely tried one that could overturn it, leaving them stuck for long stretches on a rule that was too narrow and wrong. Classic and repeatedly replicated. Your ICP is that guess, and your lost deals are the test you never ran.
• Hard evidence · the positive test strategy: Klayman & Ha (1987, Psychological Review) made that precise — people default to examining cases where the hypothesis holds. Under some structures it's an efficient heuristic, but when the task is telling two classes apart it fails systematically. Prospecting is the second kind of task.
• Hard evidence (meta-analysis) · the scorecard beats the gut: Grove et al. (2000, Psychological Assessment) pooled 136 studies comparing mechanical prediction (formulas, scorecards) against expert clinical judgement. Mechanical was on average equal or superior, and clearly worse only in rare cases. So the ugly 0/1 table will probably outperform a veteran rep's "I have a good feeling about this one."

Boundary: with fewer than five wins, everything this subtraction produces is noise. Skip the proportions and use one qualitative signal instead — what had just happened inside the company the last time someone bought. Profiles also expire: recompute when the product or the market moves, rather than running a table you subtracted two years ago. (Whether one specific name on the list deserves more of your time is the qualification piece.)

Using words the other person already uses reads as "this one gets it" more reliably than your more accurate phrasing does, because familiar words take less work to process — and that feeling of ease gets misread as true, credible, warm. This is processing fluency: people cannot tell "I understood this easily" apart from "this is right," so they treat the first as evidence of the second.

• Hard evidence · easier to process means better liked: Reber, Winkielman & Schwarz (1998, Psychological Science) raised only the processing fluency of a stimulus via contrast and priming, leaving content untouched, and liking rose significantly — with subjects entirely unaware of the cause. Peer-reviewed, solid.
• Hard evidence · fluency even shifts judgements of truth: McGlone & Tofighbakhsh (2000, Psychological Science) found that semantically equivalent aphorisms were rated as more accurate descriptions of human behaviour when they rhymed. Fluency doesn't just move liking; it gets counted as evidence that the statement is correct.
• Hard evidence (meta-analysis) · mere exposure: Bornstein (1989, Psychological Bulletin) pooled more than 200 studies confirming the exposure effect — familiarity alone produces liking, and it holds even without conscious recognition. Translated to outreach: in the first second a stranger isn't judging your logic, they're judging whether the sentence parses.

Boundary: these experiments measured liking and truth judgements, not reply rates, so the crossing into sales deserves a discount — they give you a direction, not an effect size. The reverse risk is real too: fluent isn't correct, and their own phrasing may be repeating a wrong self-diagnosis. So use their words in the opening, where the job is getting them to talk, and never in the diagnosis, which only follow-up questions can do.

This is a standard explore–exploit problem: each candidate channel is a slot machine whose true payout you don't know, and you're allocating between trying more of them and concentrating on the best one so far. The optimal shape is always the same — equal small bets first, then fast concentration — which is neither going all in at the start nor spreading five ways forever. There's a second layer: channels where you have an identity have decreasing unit cost, because what you leave there gets seen repeatedly, while a purely cold channel starts from zero on every single touch.

• Mathematical result (not psychological evidence) · multi-armed bandits: Gittins (1979, JRSS Series B) proved the optimal index policy for the discounted multi-armed bandit. That's a theorem, not an empirical regularity, and the practical "spread small, then concentrate" recipe is a crude approximation of that family of results.
• Industry practice (not experimental, weaker evidence): Ross & Tyler's Predictable Revenue (2011) argues for splitting lead generation and closing between different people, on the grounds that the two require different rhythms and skills. Widely adopted — but what supports it is company case studies rather than controlled trials, so hold it as a hypothesis, not a conclusion.
• Moderate strength · saturate one small circle before expanding: the recurring pattern in Rogers's Diffusion of Innovations survey is that new practices work through a small, homogeneous circle before spreading outward. A synthesis of observational research — direction credible, force discounted.

Boundary: channel economics drift, and a channel that is cheap this year may be crowded next year, so retest what you cut with a small bet every six months. And "real conversations" fails as a counting unit in very low-price, very high-volume businesses — nobody needs a twenty-minute call before buying a coffee. There, count conversion steps instead.

The lower the reply rate, the less information a given sample size can carry — one reply's worth of random fluctuation can double the ratio. Yet people systematically believe small samples represent the population, so they read "this opening doesn't work" off twenty emails and kill an approach that was actually fine. There's a second layer, and it's emotional: every rejection genuinely stings, so people substitute "let me tweak the wording" for "let me send another batch" — because tweaking wording requires facing no new rejection at all.

• Hard evidence · the law of small numbers: Tversky & Kahneman (1971, Psychological Bulletin) showed that even statistically trained researchers overestimate how representative small samples are, and consequently conclude too early and abandon effective approaches. Peer-reviewed and classic. Staring at the reply rate on twenty emails is exactly this.
• A calculation, not a citation · how big a sample: by the standard two-proportion power calculation (α = 0.05, power = 0.8), reliably distinguishing a 2% from a 4% reply rate needs upwards of a thousand sends per arm. That's computed rather than cited — but its implication is hard: your sample size cannot detect any small change.
• Contested — treat as a mechanism hint only: Eisenberger, Lieberman & Williams (2003, Science) found in the Cyberball paradigm that social exclusion raised activity in dorsal anterior cingulate cortex (dACC) and anterior insula, regions overlapping with the distress component of physical pain — the sting of rejection isn't purely metaphorical. But the strong version has been weakened: Woo et al. (2014, Nature Communications) used multivoxel pattern analysis to show that physical and social pain are separably represented within those same regions. Fine as an explanation for why outreach is hard to sustain; not a settled fact.

Boundary: all of this holds only for outreach with low reply rates, modest deal sizes, and enough volume to accumulate a sample. If your total addressable set is forty companies, the statistics collapse entirely — the right move there is to research and design for each one separately, treating outreach as forty projects rather than one batch.