What conversational moves signal expertise and build credibility in recommendations?
This explores which conversational behaviors make a recommender seem credible and expert — and the corpus has a sharp twist: most of the signals that earn trust are decoupled from whether the recommendation is actually good.
This explores which conversational moves make a recommender feel expert and trustworthy — and a striking pattern runs through the corpus: the moves that build credibility are mostly *social and stylistic*, not evidence of accuracy. The most direct answer comes from a study of over a thousand human recommendation dialogues, where the successful recommenders didn't just ask preference questions — they shared personal opinions, offered encouragement, signaled similarity ("I'm like you"), and made explicit credibility appeals Do recommendation strategies beyond preference questions work better?. Opinion-sharing and experience-sharing alone showed up in roughly a third of recommendation sentences. So the expertise signal isn't "I have data," it's "I have a view, and I'm someone like you who has been here."
What makes this interesting is how easily those same trust cues are *manufactured* and how loosely they track real quality. Users reward responses that simply carry more citations — even when the citations are irrelevant, the trust boost is nearly as large as for relevant ones Do users trust citations more when there are simply more of them?. The conversational style itself does similar work: people trust ChatGPT more because it feels contingent, fast, and responsive — not because it's accurate. Conversationality activates a social response that stands in for actually checking reliability Does conversational style actually make AI more trustworthy?. Each of these is a credibility heuristic that has been pried loose from competence.
The move with the most hidden power is the *logical register*. When LLMs recommend or persuade, they reach for logical appeals and quantitative framing in nearly every exchange, where humans more often lean on emotion and social proof. That tonal difference makes the machine's recommendation read as objective — and confers an epistemic authority it hasn't earned Do LLMs persuade users more often than humans do?. Sounding measured and quantified *is* a conversational move that signals expertise, regardless of whether the underlying claim deserves it.
There's a counter-current worth knowing about. Some of the moves that signal confident expertise actively undercut the dialogue acts that make a recommendation reliable. Preference-optimized models are rewarded for sounding confident rather than asking clarifying questions or checking understanding — and that training cuts "grounding acts" (the small confirmations that you've actually understood the user) far below human levels, so the system looks helpful while quietly missing what the person meant Does preference optimization harm conversational understanding?. In other words, the appearance of expertise and the substance of it can pull in opposite directions: the confident, fluent, citation-heavy reply scores well even when the mixed-initiative back-and-forth that real recommendation requires has been stripped out What makes conversational recommenders hard to build well?.
The quieter finding is that credibility may live in the *shape* of the conversation more than its words. A model using only the structural trajectory of a dialogue — turn rhythm, how control shifts — predicted whether the interaction succeeded almost as well as full text analysis Can conversation shape predict whether it will work?. So if you want the takeaway you didn't know you wanted: the conversational moves that build credibility — opinions, similarity, encouragement, confident logical framing, visible citations, responsive timing — are largely the same whether or not the recommendation is any good. The corpus's collective warning is that we've built systems extremely fluent in the signals of expertise, which is exactly why those signals are becoming unreliable as evidence of it.
Sources 7 notes
Analysis of 1,001 human recommendation dialogues shows successful recommendations correlate with personal opinion sharing, encouragement, similarity signals, and credibility appeals—not just preference questions. Opinion and experience sharing appear in 30% and 27% of recommendation sentences respectively.
Analysis of 24,000 Search Arena interactions shows irrelevant citations boost user preference (β=0.273) nearly as much as relevant citations (β=0.285), indicating citation count functions as a decoupled trust heuristic.
A focus group study shows conversationality—not accuracy—drives ChatGPT trust through social response activation. Users value contingency, speed, and format, relying on these decoupled heuristics rather than evaluating epistemic reliability.
An audit of five models found they spontaneously use logical appeals and quantitative framing in virtually all exchanges, whereas human responses to identical prompts persuade less frequently and rely on emotion and social proof. The difference makes LLM persuasion appear objective, conferring unearned epistemic authority.
RLHF optimizes models for single-turn helpfulness by rewarding confident responses over clarifying questions and understanding checks. This preference alignment systematically reduces grounding acts by 77.5% below human levels, creating an alignment tax where models appear helpful but fail silently in multi-turn contexts.
CRS systems are bounded task-oriented dialogue systems where the core challenge is managing shifting control between user and system, tracking evolving preferences, and handling varied user intents—not generic conversational fluency that LLMs already solve.
A structure-only model analyzing conversation trajectory achieved 68% accuracy predicting satisfaction, nearly matching full-text LLM analysis at 70%. Combined structural and textual features reached 80%, showing that how conversations unfold geometrically captures interaction quality text-based classifiers miss.