필사 모드: The Structure of Persuasion — It's Sequence, Not Eloquence, That Moves People

Introduction — The Meeting Where the Better Argument Loses

One scene shows up constantly in tech organizations. Two proposals go up for review, and by the numbers, Proposal A is clearly better. The person who prepared A also worked longer and more diligently than the other presenter. And yet, by the end of the meeting, Proposal B wins.

When this keeps happening, people usually land on one of two conclusions: "the organization is irrational," or "I'm bad at talking." Both are usually wrong. What actually happened is much drier than that. The presenter of Proposal A built the argument on a premise the other side never agreed to, put the strongest piece of evidence on slide 12 where nobody was still looking, and asked for something that, once approved, couldn't be undone. There was nothing wrong with the sentences themselves.

This post isn't about phrasing technique. Phrasing is the last five percent, and in most cases persuasion has already succeeded or failed before you get there. One more thing up front: persuasion is a field with an unusually large amount of unverified pop-psychology floating around. Wherever the evidence here is thin, or was scaled back under replication, this post says so.

What Collapses Before You Fix a Sentence

A large share of persuasion advice stays at the vocabulary level: add "because," use people's names, use odd numbers.

The most widely cited example is Ellen Langer's 1978 copy-machine experiment. Researchers asked to cut in line at a copier, either with a reason attached or without one, and the well-known result is that a reason carrying essentially zero information — "because I need to make copies" — pushed compliance up to 93 percent. From here came the widely repeated conclusion that the word "because" itself is magic.

But the same paper contains another condition. When the request got bigger — asking to go first for 20 copies instead of a few — compliance with that same empty reason collapsed to around 24 percent. What the paper itself actually showed wasn't the magic of a reason — it was that the magic only works for small requests. The condition simply got dropped somewhere in the process of popularization.

The practical takeaway here is this: in front of a request where the other person stands to actually lose something, vocabulary-level techniques barely work at all. Past that point, it becomes a matter of structure. This post covers four axes: premise, the ordering of evidence, the size of the request, and autonomy.

Shared Premise — What Has to Be Established Before the Claim

The enthymeme Aristotle described in Rhetoric is a form where you build your conclusion on top of a premise the audience already holds, without bothering to state it. Ninety percent of real-world persuasion takes this form. The problem shows up when the audience doesn't actually hold that premise. When that happens, the argument doesn't look wrong — it doesn't register at all.

The claim "this approach cuts response time by 30 percent" rests on the premise that response time matters right now. If the decision-maker's priority this quarter is acquiring new customers, that sentence is true for them but irrelevant. It doesn't even get refuted. It just gets dismissed.

There's interesting evidence from political communication about what to do when premises differ: the research on moral reframing by Matthew Feinberg and Robb Willer. In a 2015 study, liberal-leaning respondents supported a larger defense budget more when it was framed in the language of reducing inequality and discrimination than when it was framed in the language of loyalty or authority. Conservative-leaning respondents supported environmental regulation more when it was framed as protecting the purity of the land rather than reducing harm. The content was the same; only the underlying premise changed.

A more important observation comes from the same line of research. When asked to pick which argument would land better with the other side, 64 percent of liberal respondents and 85 percent of conservative respondents correctly picked the reframed version. But when asked to actually write a persuasive message themselves, fewer than 10 percent appealed to the other side's values. People know what they're supposed to do, and by default, still speak from their own premises.

The limits are worth noting too. The effect sizes in the moral-reframing literature are generally not large, and most measurements are attitude scales, not behavior. The samples are also concentrated in the American political landscape. So reading this as "change the value-language and people will come around" overreaches. The accurate reading is this: speaking without checking the other person's premise is almost certainly a losing move, and checking costs one line of a question.

In practice, this looks like confirming, before you propose anything, "is X what we agreed to prioritize this quarter?" If the answer is no, you don't drop the proposal on the spot — you change how it connects. That one sentence often saves a thirty-minute presentation.

The Order of Evidence — What Goes Where Attention Is Highest

"Strongest evidence first" versus "strongest evidence last." This debate is old, and honestly, unresolved. Persuasion research comparing primacy and recency effects flips direction depending on delay, involvement, and the spacing between messages. Writing that gives you a confident rule in this territory has usually gotten ahead of the evidence.

Better to anchor on one thing that is certain instead: attention is finite, and it's the medium, not the order, that determines how attention is distributed.

In something people read, attention piles up at the front. The Nielsen Norman Group's eye-tracking research reports that users read only 20 to 28 percent of the words on an average page, and the F-shaped scan pattern first described in 2006 has now been observed repeatedly for close to twenty years. The twelfth paragraph of a document is, for practical purposes, a paragraph that doesn't exist.

By contrast, when a decision gets made on the spot at the end of a meeting, what's still standing at the end weighs into the decision. So the practical rule splits by medium. For an asynchronous document someone reads and judges, put the strongest evidence in the first paragraph. For a synchronous meeting, put it at both the start and the end. Either way, the middle is where the weak evidence goes.

The habit of tacking on a few more pieces of weaker evidence is also worth doubting. Weaver, Garcia, and Schwarz's presenter's paradox reports that because people average information rather than sum it, adding weak evidence on top of strong evidence can actually lower the overall evaluation. It's counterintuitive and matches practical experience well. But the individual experiments had fairly small samples, and larger-scale reverification is still being worked out, so it's premature to cite it as an established rule. Still, it's worth the question when you're editing a document: "would removing this piece of evidence make the proposal weaker?"

The Size of the Request — Small Enough to Get a Yes

Foot-in-the-door gets cited constantly here: get a small favor approved first, and the approval rate for a bigger favor that follows goes up. The original experiment is Freedman and Fraser's from 1966.

The replication record tells a different story. Jerry Burger's 1999 review synthesizing more than 100 studies concludes: the effect replicates, but it's weak, and nowhere near as robust as commonly assumed. Close to half the studies found no effect, or an effect in the opposite direction. An effect the size the original paper reported has never been observed again since. In short, the grounds for using this as a technique are thin.

Strip the technique away and what's left is a much simpler, much sturdier fact. The cost of approval doesn't come from the size of the request — it comes from how hard it is to undo. Someone approving a three-month full rollout isn't approving three months; they're approving the risk that their own judgment turns out to be wrong three months from now.

Amazon's internal distinction between one-way doors and two-way doors maps onto this exactly. Reversible decisions move fast; irreversible decisions move slow. The proposer's job is to design their own request as a two-way door, which lowers the other person's deliberation burden by itself.

Translated into practice: change "let's switch to the new queue system" into "four weeks in the payments domain only; if the metric doesn't stay below X, we roll back and drop this proposal." The key is that the proposer writes the failure criteria first. Without that one line, the other person has to approve not just the plan but the risk that "even if this fails, this person is going to keep pushing it anyway."

The Most Important Asymmetry — People Refuse to Be Persuaded

If everything up to here has been about designing the message, the last axis is a different kind of thing. People are open to their own judgment changing; they are closed to being persuaded by someone else. These are different experiences that can produce the same outcome.

This is what psychological reactance theory is about. Stephen Rains's 2013 meta-analysis concludes that reactance fits the data best when modeled not as a single emotion but as a combination of anger and negative cognition. A more recent meta-analysis in health communication reports that strong language threatening freedom raised anger (28 studies, r = .21), negative cognition (25 studies, r = .17), and reactance (53 studies, r = .20), and that these in turn lowered persuasive outcomes. The sizes are small, but the direction is consistent. Phrases like "you absolutely must" or "there's no room for debate here" carry a measurable cost.

So does explicitly guaranteeing autonomy fix this in reverse? Here you have to be careful. The so-called BYAF technique — adding "but you are free to refuse" to raise compliance — was reported effective in Carpenter's 2013 meta-analysis, but a preregistered meta-analysis that re-synthesized 52 experiments (roughly 19,500 people) found that the overall effect of g = 0.44 dropped to g = 0.11 once you restrict to low-bias-risk studies, with a confidence interval that crossed zero. Replication indicators were also very low. Claims of the form "add one sentence and compliance goes up" follow this exact path in this field almost every time.

The approach that has held up better is one where you get the other person to generate the argument themselves. Elliot Aronson's line of research on self-persuasion has reported repeatedly that arguments people construct themselves last longer and provoke less resistance than arguments handed to them from outside. The mechanism is obvious enough: an outside message gets discounted, but your own thinking doesn't.

The field evidence comes from deep canvassing. David Broockman and Joshua Kalla's 2016 Science paper reported that roughly ten minutes of non-judgmental conversation produced attitude change that persisted for more than three months, and a 2023 follow-up concluded that the core mechanism wasn't imagining the other person's perspective but hearing and confirming it directly. This line of research is, if anything, unusually credible, because it was redesigned after a 2014 fabrication scandal with strong preregistration and data-sharing requirements built in.

There are three practical translations here. Give people the material to reach a conclusion, instead of handing them the conclusion. Don't leave only one option on the table. And use what the other person has already said as your material. That last one is especially powerful in negotiation, which is covered separately in the salary negotiation post.

The Elaboration Likelihood Model and Inoculation Theory — Not the Same Weight of Evidence

These are the two theories you inevitably run into once you study persuasion even a little, and how much you can trust them in practice differs quite a bit.

The elaboration likelihood model (ELM) is the framework Richard Petty and John Cacioppo laid out in 1986. People process a message through one of two routes: given motivation and ability, through a central route that scrutinizes the argument itself; otherwise, through a peripheral route that leans on cues like the speaker's charisma or the sheer volume of material.

This framework has had an enormous influence for forty years, but it has real problems if you cite it as a validated law. The sharpest criticism is circularity. The model's central variable, argument strength, is often not defined in advance but judged after the fact as "the argument that generated a lot of favorable thoughts," which turns it into a post-hoc explanation rather than a prediction. Arie Kruglanski pointed out that early studies conflated the type of argument with message length. That the two routes aren't mutually exclusive, and that the elaboration continuum itself hasn't been sufficiently validated, have also been raised repeatedly.

So the accurate way to put it is this: ELM is a useful checklist, not a predictive model. There's one thing worth taking from it in practice: the same message works differently depending on how much the audience is elaborating. If the audience is in a state where they're spending time and attention on this issue, raise the quality of your argument. If they're not, no amount of additional good argument gets read, so it's better to invest in form and trust signals instead.

Inoculation theory is a different situation. Proposed by William McGuire in the 1960s, it holds that pre-exposing someone to a weak version of a counterargument and refuting it in advance makes their attitude less shaken later by a strong attack. In John Banas and Stephen Rains's 2010 meta-analysis synthesizing 54 cases, inoculation messages outperformed both supportive messages and no-treatment controls, and the average effect size widely cited in the literature is around d = 0.43. By social psychology standards, that's a moderate effect, which counts as fairly large in this field.

The scale of the evidence has grown recently. Roozenbeek and van der Linden's team's 2022 Science Advances paper ran five preregistered randomized controlled trials (n = 5,416), a replication study (n = 1,068), and a YouTube field experiment (n = 22,632), and reported that manipulation-technique recognition rose by an average of 5 percent even in the real YouTube environment.

The same paragraph needs its counterpoint too. Modirrousta-Galian and Higham pointed out that gamified inoculation interventions might not be raising discernment at all, but simply shifting response criteria to be more conservative — making people more suspicious of everything — and proposed using signal-detection-theory metrics instead. A subsequent reverification in a different sample found no confirmed improvement in the ability to tell true from false. Inoculation's decay has also already been reported in the original meta-analysis. Two weeks out, the effect fades.

Brought down to practice, the most solid finding is a cousin of inoculation. Per Daniel O'Keefe's 1999 meta-analysis, a refutational two-sided message — one that names the counterargument and then answers it — beat a one-sided message on both credibility and persuasiveness (42 studies, d = 0.16). By contrast, a two-sided message that named the counterargument but never answered it did worse than a one-sided message (65 studies, d = -0.10). Raise the objection, but always answer it — that's the single most practical line in this literature. The specific mechanics of handling objections continue in how to handle objections.

Here's a table summarizing the evidentiary state of the claims covered in this post. Get in the habit of checking this column before you cite any of them.

For ego depletion, see the 23-lab preregistered replication; for anchoring, see the Many Labs project.

Rewriting an Actual Message

Say you're posting a proposal to improve the CI pipeline in a team channel.

Before:

CI has been really slow lately. Each build takes 40 minutes, and everyone's frustrated. I think switching to running tests in parallel would help a lot. Most other companies are doing this now too, and there are several tools available. Please take a look.

Count what's missing: no connection to a goal the team has already agreed on. The only usable piece of evidence is "40 minutes," and even that has no source. No answer to any anticipated objection. No visible alternatives considered. And critically, no request for what should actually be approved. "Please take a look" isn't a request — it's an avoidance of one.

After:

At last quarter's goal-setting meeting, we agreed that shortening deploy lead time was our top priority. Right now, the single biggest segment in that lead time is CI. Based on the last four weeks of pipeline logs, the median is 38 minutes and the 90th percentile is 62 minutes, and 71 percent of that comes from the integration test stage. At an average of 9 reruns a day, that works out to roughly 11 hours of team-wide waiting per week.

The proposal is to split integration tests into 4 shards. Let me raise the two objections I'd expect first. First, sharding surfaces shared state between tests, so failures will likely go up initially. For that reason, I'd like to treat shard failures as warnings rather than blockers for the first two weeks. Second, runner costs go up by roughly 400,000 KRW a month. Converted to labor cost, the waiting time above is worth more than that.

As alternatives, I also looked at change-impact-based selective test execution and stronger caching. Selective execution would miss cases because our monorepo's dependency graph still isn't accurate enough. Caching is already in place, and there wasn't much headroom left.

There's one request: apply this to payments-domain tests only, as a two-person, five-day task next sprint. If the median doesn't drop below 25 minutes, we roll back and I'll drop this proposal. If there's no objection by Thursday, I'll go ahead on that basis.

All four axes are in there. The first paragraph connects to a premise the team has already agreed on, and puts the strongest evidence — the actual measurements — where attention is highest. The second paragraph has the proposer raising the two strongest objections themselves, and answering them. The third paragraph shows the alternatives that were considered. The request in the last paragraph is reversible, the failure criteria are written by the proposer's own hand, and the other person still has time left to object.

It's worth noting that this got longer. Something short gets read but doesn't get decided. Persuading through documents is covered separately in persuasive writing for engineers.

Closing — Persuasion Isn't Changing Minds

Being good at persuasion isn't the ability to flip someone's thinking. It's much closer to removing, one at a time, the friction that shows up when the other person is about to move. If the premise is misaligned, fix that first. If your evidence sits somewhere nobody reads, move it. If saying yes looks risky, make it reversible. If a phrase reads as coercion, strip it out.

And there's one habit worth keeping when you read advice in this field: ask about effect size and whether it replicated. A large share of sentences that start with "research shows" have been substantially scaled back since the 2010s, and writing that doesn't say so isn't help. It's noise.