What the evidence says actually separates a good conversationalist from a great one. Every claim here traces to an experiment, not to folklore.
The striking thing about this literature is how consistently miscalibrated people are. Not randomly wrong — wrong in the same direction, on almost every dimension. You underestimate how much people liked you. You overestimate how awkward depth will be. You pick the wrong story. You can't tell when someone wants to leave. Nearly every correction below is a correction in the same direction: toward the other person, and further than feels natural.
People ask a mean of 6.7 questions in a 15-minute get-to-know-you conversation. Which kind matters far more than how many. Huang et al. 2017, JPSP 113(3)
| Type | What it is | % of all | Effect on responsiveness |
|---|---|---|---|
| Follow-up | Draws on what they just said | 40.5% | +1.92 p<.001 — and the only type that also raises liking |
| Full-switch | New topic, unconnected | 27.6% | −2.83 p<.001 — actively harmful |
| Mirror | Bounced straight back ("You?") | 19.0% | +0.16 n.s. — empirically inert |
| Introductory | Opening enquiry | 5.5% | +0.98 n.s. |
| Partial-switch | Adjacent topic | 5.5% | too rare to measure |
| Rhetorical | Not seeking an answer | 1.9% | too rare to measure |
Nearly half the questions people ask are the two types that do nothing or actively hurt. "Ask more questions" is bad advice. "Make your next question depend on their last answer" is the whole skill.
Mechanism: the effect runs entirely through responsiveness — the sense of being listened to, understood, and cared about. A follow-up question is the cheapest available proof that you processed what they said.
Askers in these studies did not predict any of this. Predicted liking was 5.16 for high-question-askers versus 5.22 for low — no difference at all. People do not know that asking questions makes them more likeable. Which means the internal signal telling you you're asking plenty is not measuring anything.
Ask a question. Let them answer. Then immediately pivot to answering it yourself. Three flavours: ask-bragging, ask-complaining, ask-sharing. Brooks & Yeomans 2025, JEP: General 154(3)
If you notice you want to share your own answer — just share it. Direct self-disclosure beats the question-shaped version by 15 percentage points. The question wrapper isn't politeness; it's a tax.
Four possible responses when someone tells you something went well. Only one predicts relationship quality — and the polite one is not it. Gable et al. 2004, JPSP 87(2)
| Constructive | Destructive | |
|---|---|---|
| Active | Enthusiastic, elaborating, asks for detail.
"That's brilliant — how did you find out?" The only one that works. |
Quashes the event. "Won't that mean more hours?" |
| Passive | Mild, understated. "Oh, nice." Patterns with the destructive column, not the constructive one. |
Ignores it. Changes the subject. |
This is the single easiest upgrade available to someone already competent, because the failure mode is politeness, not rudeness. "Oh, that's great" feels like a supportive response. It measures like a destructive one.
Two findings sit in tension, and resolving them tells you what to actually train.
High-quality listening measurably alters what the speaker understands about their own position, with effects up to d=1.45. The operationalized behaviours include eye contact, leaning in, non-judgemental stance — and crucially verbal reflections that check understanding: "I want to make sure I understood — did I get that right?" Itzchakov et al. 2024, JPSP 126(2)
Five studies, N=1,225. Minds wander 24% of the time. Participants could listen attentively, inattentively, or with divided attention on command — and partners could not detect the difference. Even partners who literally could not hear the words, but were paid to act like they could, went undetected. Collins et al. 2024, JEP: General 153(2)
The verbal layer of listening carries signal; the nonverbal performance layer largely doesn't, because it's identical whether you're listening or not. Optimize for saying something that proves you understood — not for looking attentive. A follow-up question is the cheapest such proof, which is why §1 and §4 are really the same instruction.
Nodding and "mhm" are generic — they fit any story. Wincing at the right moment, exclaiming at the turn, laughing at the specific line: these are specific, and they are the ones distracted listeners stop producing. When they disappear, the narrator measurably tells the story worse, especially at the dramatic ending. Bavelas, Coates & Johnson 2000, JPSP 79(6)
You are not a passive audience. Your reactions are part of how well the story gets told.
Given a choice between telling a novel story and a familiar one, speakers expect the novel one to land better (4.49 vs 3.19) and 66.7% choose it. Listeners react worse to the novel story — 4.56 vs 5.61, d=1.10. A complete reversal, and speakers never notice. Cooney, Gilbert & Wilson 2017, Psych Science 28(3)
The cause: novel stories require background the listener doesn't have. Familiar stories let the listener fill in the gaps themselves, which is the part they enjoy.
Corollary, and it stings: people who'd just had a superior experience felt no better going into a group conversation, and measurably worse coming out — d=0.57 on mood, d=1.41 on feeling excluded. Cooney et al. 2014 "Have great stories" is not the advice it appears to be.
People overestimate how awkward a deep conversation will be by d=2.23. The gap between expected and actual awkwardness was roughly four times larger for deep conversation than for shallow. Mechanism: people underestimate how much strangers care about their disclosures. Kardas, Kumar & Epley 2022, JPSP 122(3)
So: go deeper than feels comfortable, because the discomfort signal is a known forecasting error — but calibrate to setting. Depth works in a conversation that has already established itself. It does not work as an opener.
932 conversations analysed. They almost never ended when both people wanted them to. Mastroianni et al. 2021, PNAS 118(10)
Leaving a conversation early is not a rejection and is not read as one. Statistically, the other person probably wanted out before you did. This makes exits emotionally cheap — which is what makes high-volume practice possible at all. If every conversation has to be run to its natural end, you can't afford many. If you can leave cleanly at 60 seconds, you can afford dozens.
Successful humour raises status (5.03 vs 4.43), mediated by perceived competence and confidence. The asymmetry is the useful part: an inappropriate joke raises perceived confidence whether or not it lands, but lowers perceived competence — and that combination nets out worse than simply being serious. Bitterly, Brooks & Schweitzer 2017, JPSP 112(3)
Status ranking, all pairwise p<.001: Serious > Successful-inappropriate > Failed-inappropriate. The risk that matters is not will it be funny but is it appropriate. A joke that flops harmlessly costs you little; one that lands but shouldn't have costs you more than silence.
One hidden cost: pairing a negative self-disclosure with a joke makes the disclosure less likely to be believed. Self-deprecation buys likeability at the price of being taken seriously. Bitterly & Schweitzer 2019
The most directly trainable recipe in this literature, because it operates at word level. Teaching it raised perceived receptiveness by β=0.57 and persuasiveness with ideological opponents by β=0.24. Yeomans et al. 2020, OBHDP 160
| Do | Don't |
|---|---|
| Positive statements | Negations — "no," "wrong," "not" |
| Acknowledge understanding — "I see your point" | Pile on justification — "because," "therefore" |
| Name actual agreement — "I agree that…" | Restate your position louder |
| Hedge — "somewhat," "might," "in some cases" | Absolute claims |
Field validation worth noting: across 585 matched pairs of Wikipedia editor threads, low receptiveness predicted becoming the target of a personal attack 60.3% of the time — more strongly than it predicted becoming the attacker. Unreceptive language doesn't just fail to persuade; it recruits hostility.
After five-minute conversations with strangers, actual liking averaged 5.82; participants believed it was 5.17. Tracked across an academic year among Yale suitemates, the gap persisted at every timepoint but the last. Boothby et al. 2018, Psych Science 29(11)
It is larger for shyer people. Which means: your post-conversation read on how it went is a biased instrument, biased downward, and biased most when you're least confident. Don't act on it.
Huang, Yeomans, Brooks, Minson & Gino (2017), "It Doesn't Hurt to Ask: Question-Asking Increases Liking," Journal of Personality and Social Psychology 113(3), 430–452. Free PDF from HBS
Read Study 1 and the taxonomy table. It is the shortest path from "I know I should ask questions" to knowing which ones. If you read one paper from this document, read this.