Skip to main content

Should a virtual service agent make small talk, or copy the way each customer writes? What a Psychology & Marketing study found

Many chatbots and avatars are scripted to sound warm in every line. Two Dutch experiments with an animated adviser called Charlotte set friendly, businesslike and mirrored replies against each other, once for holiday advice and once for a tax return.

van Pinxteren et al. show an animated agent’s social replies raised rapport in holiday and tax advice chats, and mirroring raised engagement for holidays.
Date

14 November 2026

Marketing in 3 · 2023 in research · No. 3 · Paper published 2 February 2023

A polished chrome service bell on a glossy black desk beside a brushed steel sign reading Informacije / Information, both mirrored in the desk under deep blue light
When the help desk is a screen, the agent's wording does the welcoming. Photo: Milos Lopusina on Unsplash.

The short answer

Small talk helps; copying the customer is optional. People who took holiday or tax advice from an animated virtual agent felt more rapport when its replies were warm and personal rather than strictly businesslike. An agent that matched each customer's style, turn by turn, did as well on rapport as one that was friendly throughout, and made the holiday chat more engaging. Any effect of style on satisfaction and return intentions came through that sense of connection.

Evidence level
  1. One study
  2. Several studies
  3. Meta-analysis

Key takeaways

  • Across 264 participants and two services, an agent that used social lines, all the time or when mirroring, built more rapport than a strictly businesslike one.
  • Over 10 agent replies, mirroring always used fewer friendly lines than the agent that was social throughout, yet rapport came out no different in either service.
  • Mirroring raised engagement in 1 of the 2 services, the holiday chat, and left the tax chat unchanged on this measure.
  • The agent's style accounted for only about 6% of the variation in rapport and in engagement after the holiday chat, and under 4% after the tax chat.

What the researchers did

Michelle M. E. van Pinxteren of Radboud University in Nijmegen, Mark Pluymaekers and Anna Krispin of Zuyd University of Applied Sciences in Maastricht, and Jos Lemmink of Maastricht University took up a design question every team building a service chatbot or avatar has to settle. Should the agent stick to the task, or add the greetings, small talk and words of support a good human employee would use? Earlier studies disagreed, partly because people differ in how much chat they want and because that want changes within a single conversation.

Their candidate answer was mirroring. If the customer writes in a friendly, personal way, the agent replies in kind; if the customer is brief and to the point, so is the agent. They tested this against two fixed scripts in two settings: advice on where to go on holiday, a service people use for enjoyment, and advice on claiming study costs in a tax return, a purely practical one.

The agent was Charlotte, an animated character who typed her answers and showed facial expressions such as a polite smile. Participants read a short scenario, then picked one of two possible replies at every turn, one businesslike and one social, so the conversation followed their own style. After Charlotte's tenth reply the chat ended and they filled in a questionnaire.

Participants
142 people in the holiday experiment and 122 in the tax experiment, Dutch-speaking volunteers recruited in the southern Netherlands at a university of applied sciences
Agent
Charlotte, a virtual adviser with a female name and face, built on an online platform for conversations with avatars
Three versions
Task-oriented in every turn, social-oriented in every turn, or mirroring the style the participant had just used
Measured
Rapport, engagement, satisfaction with the encounter and intention to use the service again, each on seven-step scales of agreement
Checks
Four communication experts sorted every scripted line into task or social without a single error

Businesslike, friendly or mirrored

The difference between the scripts was in wording only. The businesslike Charlotte opened with "Welcome to the virtual travel agent." and closed with "The end of the conversation." The friendly one introduced herself by name, said she was curious about what else the traveller liked, and wished them a nice day. The facts and questions were the same.

The mirroring version drew on communication accommodation theory. It holds that people adjust to how others speak to show they understand them. Earlier work on chats between people and with virtual assistants found that social remarks cluster at certain moments, mostly greetings, goodbyes and moments of socialising. A customer's own choice of reply seemed a practical cue for when such remarks are welcome.

Two outcomes stood for a sense of connection. Rapport is the feeling of clicking, of having some chemistry with the other party, and was measured with ten statements. Engagement is how much someone wants to take part in the conversation, measured with five.

Holiday advice: social lines raised rapport, mirroring raised engagement

The results come as unstandardised regression coefficients, written B; a larger B means a bigger effect of the script on the outcome.

Set against the businesslike script, the two versions with social lines raised rapport (B = 0.55, 95% confidence interval 0.16 to 0.95). Their effect on engagement did not reach significance.

The comparison between mirroring and constant friendliness showed the reverse pattern. Rapport did not differ, but engagement was higher with mirroring (B = 0.56, interval 0.072 to 1.05). In this service, social wording was linked to rapport, and replying in the customer's own style was linked to engagement.

Holiday advice: effect of the agent's style on rapport and engagement (unstandardised B)
Holiday advice: effect of the agent's style on rapport and engagement (unstandardised B)00.10.20.30.40.50.6Social vs task: rapport: 0.55 Social vs task: rapport0.55Social vs task: engagement: 0.42 Social vs task: engagement0.42Mirror vs always social: rapport: 0.30 Mirror vs always social: rapport0.30Mirror vs always social: engagement: 0.56 Mirror vs always social: engagement0.56
Holiday advice: effect of the agent's style on rapport and engagement (unstandardised B)
Group Value ()
Social vs task: rapport 0.55
Social vs task: engagement 0.42
Mirror vs always social: rapport 0.3
Mirror vs always social: engagement 0.56

Experiment 1 (travel advice, 142 people), Table 4. Social vs task compares the task-only agent with the social and mirroring versions together. Amber bars are significant (p < 0.01).

Tax advice: only the social lines made a difference

The second experiment repeated everything with a tax adviser. Social lines again raised rapport compared with the businesslike script (B = 0.44, interval 0.035 to 0.86). In the authors' words this calls for caution, because the model predicting rapport did not explain a significant share of the differences overall.

Nothing moved engagement. Mirroring did no better than steady friendliness on either outcome, with coefficients close to zero.

One descriptive detail the authors do not analyse: on the questionnaire about how much social talk people want from a service, the tax group averaged 4.60 and the holiday group 5.28. Whether that helps explain the weaker results for tax is our speculation.

Tax advice: effect of the agent's style on rapport and engagement (unstandardised B)
Tax advice: effect of the agent's style on rapport and engagement (unstandardised B)00.10.20.30.40.50.6Social vs task: rapport: 0.440 Social vs task: rapport0.440Social vs task: engagement: 0.120 Social vs task: engagement0.120Mirror vs always social: rapport: 0.065 Mirror vs always social: rapport0.065Mirror vs always social: engagement: 0.066 Mirror vs always social: engagement0.066
Tax advice: effect of the agent's style on rapport and engagement (unstandardised B)
Group Value ()
Social vs task: rapport 0.44
Social vs task: engagement 0.12
Mirror vs always social: rapport 0.065
Mirror vs always social: engagement 0.066

Experiment 2 (tax advice, 122 people), Table 6. The amber bar is significant (p < 0.05), though the rapport model as a whole was not.

From connection to coming back

Neither experiment showed the style of the replies directly changing satisfaction with the encounter or plans to return. It worked through the sense of connection instead. Rapport and engagement both predicted satisfaction, and satisfaction in turn predicted that intention.

Together with the style contrasts, rapport and engagement accounted for 63% of the differences in satisfaction for holiday advice and 40% for tax advice. In the holiday data, engagement and satisfaction were correlated at 0.76. Taken as a whole, the social lines' total effect on intention to return was significant for holiday advice (B = 0.65) and not for tax (B = 0.32).

More friendly lines are not always better

The finding with the widest reach concerns quantity. People talking to the mirroring Charlotte always saw fewer social lines than people talking to the agent that was friendly throughout, yet they felt just as much rapport. More human-like remarks did not produce a stronger social response, contrary to an assumption common in earlier work on people treating computers as social actors.

The authors are careful not to oversell mirroring. Its boost to engagement appeared in one service only, and how closely the agent's style matched the participant's did not predict rapport or engagement. They suspect the benefit may come from variety in the agent's wording rather than from copying as such, and call for tests that separate the two. Because rapport and engagement responded differently, they also advise anyone studying such agents to measure both.

The authors' advice

For practitioners their recommendation is plain. A social style used throughout looks like a safe choice for practical services and a beneficial one for enjoyable services. In the holiday chat it raised rapport, and in the tax chat the authors saw no negative effects.

Mirroring is harder to justify. In the experiments customers chose from prepared replies, which made copying their style easy. With free typing, the agent first has to judge the style of every message, and the authors doubt the gain seen here would repay a considerable investment.

Marketing in 3 steps

Ready. Set. Market.

ReadyDecide and prepare
  • Sort your chats into treats and chores. List every job your chatbot or avatar handles. Label each as a treat, such as trips, gifts or dining, or a chore, such as bills, forms or returns. Start the friendly rewrite with the treats.
  • Write each reply twice. For every scripted answer, draft a plain version and a warm version that carry the same facts. Keep the two within about five words of each other so the warm one does not slow anyone down.
SetBuild it
  • Put warmth at the edges. Give the agent a name and a greeting, one line of interest in the customer's plans, and a friendly goodbye. Keep the middle of the chat on the answer itself.
  • Follow the customer's lead where it is cheap. If your platform can tell a chatty message from a terse one, use one simple rule. A hello or a thank you gets one back. A one-word reply gets a short, direct answer. If it cannot, stay friendly throughout.
MarketLaunch and measure
  • Run warm against plain. Split your chat visitors randomly into two groups of equal size for 14 days or more. One group gets the warm script, the other the plain one. Change nothing else, not the agent's look, its answers or the page it sits on.
  • Ask two questions after every chat. End each conversation with two quick ratings from 1 to 7: did you feel a connection with the assistant, and did you enjoy the conversation? Compare both scores by script, and check how many people come back within 30 days.

Get it by email

New issues every Monday and Thursday, in your inbox. Free, no ads, unsubscribe anytime.

Your address is used only to send Marketing in 3. You will get one email to confirm.

What to expect

  • Rapport first. Friendly wording lifted the feeling of a connection, so that is the score to watch before satisfaction or repeat use.
  • Small shifts. Expect modest movement in connection scores rather than a transformation, and look for changes in satisfaction or intentions to return only once connection moves.
  • More payoff for enjoyable services. Holiday advice gained more from friendly lines than tax advice did.
  • Mirroring costs more to build. With open typing the agent must first classify each incoming message, an extra build step that, in the authors' judgement, may not repay the effort.

Where it does not work

  • Prepared replies, not free typing. Participants chose between two scripted lines at each turn, which kept mirroring simple and nothing like an open chat.
  • One agent, one look. Charlotte was always presented as a woman. A male or gender-neutral agent might be judged differently, and her appearance and expressions were never varied.
  • One place, one language. Volunteers were Dutch speakers from one Dutch university of applied sciences, aged 22 to 23 on average, and culture shapes how people react to human-like machines.
  • Two services. Holiday and tax advice may not represent every enjoyable or every practical service.
  • Small groups and imaginary tasks. Between 39 and 50 people used each version of the agent, and they planned an imaginary trip and an imaginary tax return, so nothing real was at stake.

Questions marketers ask

Should a chatbot make small talk with customers?

On this evidence, a little social talk helps. A virtual agent that greeted people by name, showed interest and wished them well built more rapport than one that stuck to business, both for holiday advice and for tax advice (van Pinxteren et al., 2023).

How does mirroring, or linguistic mimicry, work in a chatbot?

The agent copies the style of the customer's last message. A personal, friendly message gets a personal reply, and a brief, factual one gets a brief answer. Here it matched the friendly script on rapport and made the holiday conversation more engaging, but it did nothing extra for tax advice.

How do rapport and engagement with a virtual agent differ?

Rapport means feeling you click with the agent. Engagement is a person's motivation to join the conversation. The two were closely related, yet the agent's wording affected them differently, so the authors recommend measuring both.

Does chatbot tone matter more for some services than others?

It appeared to. Friendly replies helped most when people were choosing a holiday, a service used for enjoyment. For a tax question the benefit was smaller and less certain, though friendly lines did no harm.

Will a friendlier chatbot bring customers back?

Only indirectly. The wording did not change satisfaction or plans to return on its own. It raised rapport and engagement, and people who felt more of both were more satisfied and keener to come back.

Also new at the time

Journal of Consumer Research · 2023

How Social Media Influencers Impact Consumer Collectives: An Embeddedness Perspective

Mardon, Cocker, & Daunt

A long-running qualitative study of the YouTube beauty community traced what happened as ordinary members became social media influencers. Exchanges in the community had been governed by shared norms tied to familiar roles. When influencers tried to loosen those norms to make money, the roles clashed and members who were not selling anything got less from taking part, until influencers and their followers tried to bring the business back within community norms by renegotiating roles. Brands paying creators are also dealing with the community around them, and that community has its own sense of what is fair.

Journal of Marketing Research · 2023

Virtual Fitting Room Effect: Moderating Role of Body Mass Index

Yang, Xiong, Mao, & Ma

Virtual fitting rooms let online shoppers try clothes on an avatar shaped like them. A large field experiment with real transactions and five lab experiments found that the tool lifted evaluations and purchases among shoppers with a lower body mass index but lowered them among those with a higher one, whose avatar threatened their self-image and led them to blame the garment. Promoting varied beauty standards, giving avatars mannequin faces, offering a chance to do good and showing high-status items all eased that reaction. A try-on tool can help some shoppers and put off others, so read its results group by group.

Source

van Pinxteren, M. M. E., Pluymaekers, M., Lemmink, J., & Krispin, A. (2023). Effects of communication style on relational outcomes in interactions between customers and embodied conversational agents. Psychology & Marketing, 40(5), 938-953. https://doi.org/10.1002/mar.21792

This brief is written for practitioners from the published article. Open access (CC BY-NC-ND 4.0): free to read at the publisher. Numbers are as reported by the authors; read the article for full methods.

Keep reading

More issues on AI · Service

Bruno Schivinski

Bruno Schivinski is Associate Professor of Marketing at Gdańsk University of Technology. He studies how consumers engage with brands in digital and social media.

About Bruno · All Marketing in 3 issues · Subscribe