What do you say when you don't know what to say?

Someone you know has had bad news. You have opened the thread four times and closed it, because everything you draft sounds either hollow or presumptuous. I co-founded Subtext, and this is the message people most often abandon halfway through writing.

Two findings should change what you do. A clumsy message beats silence, and that has now been tested head to head rather than inferred from adjacent research. And a moderate message performs about as well as a beautifully crafted one, which means the effort you are pouring into finding perfect words is buying you very little. What actually causes damage sits at the bottom of the range, not the top.

The comparison everyone assumed, now tested

Until recently, every article on this topic argued from analogy. Ray and Floyd tested it directly¹.

They ran a survey experiment with 886 US adults, census-matched on age, sex and ethnicity, in the context of supporting someone who had suffered an injury. They compared nonsupport, meaning saying nothing, against a low person-centred support message, meaning the worst category of message rather than a merely awkward one.

Nonsupport generated less emotional improvement and more negative relational consequences than the bad message did.

Read that carefully. Not “an awkward message beats silence”. A message from the dismissive end of the scale still outperformed saying nothing at all.

Caveats worth carrying: it is a survey experiment using scenarios rather than observed real bereavement, and it concerns injury rather than death. But it is a direct test, and it replaces a decade of reasonable guessing.

Silence is also not rare. A study of 205 young adult cancer patients across 22 countries found roughly three in five had experienced nonsupport, meaning support they expected from someone in their network that never came². It gets decoded as abandonment rather than as respectful restraint.

The ceiling is lower than you think

This is the finding that should actually relieve you.

Supportive messages get coded on a nine-level scale of person-centredness, which is the degree to which a message acknowledges, elaborates and legitimises the other person’s feelings rather than challenging or diverting from them. Low messages deny or criticise the feeling. Moderate messages acknowledge it and then move attention away. High messages validate and elaborate.

Tian and Solomon ran two experiments, one lab and one online, on messages about the death of a parent³. Moderately person-centred messages produced less resistance and better support quality than low ones, and were perceived to be as effective as highly person-centred ones.

So “I am so sorry you are going through this” appears to do the work. The eloquent, deeply elaborated message is not measurably better than the plain sincere one. What you are agonising over is the difference between a seven and a nine, and that difference does not show up in the data. The difference between a three and a seven does. Which is the only thing worth checking before you send, and it is roughly what Subtext is looking at when it reads a draft like this one.

There is a meta-analysis of 23 studies confirming that person-centredness relates to effectiveness overall, with a stronger relationship to perceived effectiveness than actual⁴. One detail in it is worth knowing: the advantage shrinks in studies using live interaction rather than written vignettes, where tone, presence and relationship history carry weight that a rated message strips out. I would report the direction of that finding and not a number, because the effect sizes conflict across the sources I worked from and none was verifiable without the paywalled full text.

Why the bad message is bad

A dismissive message does not merely fail to comfort. It provokes resistance.

Tian, Solomon and St.Cyr Brisini had 325 married adults evaluate a hypothetical support message after a marital disagreement⁵. Low person-centred messages were perceived as conveying dominance and as having weak argument quality, and that predicted psychological reactance through a perceived threat to the recipient’s freedom. In the summary of this work, participants reported feeling angry and mentally arguing back as they read.

That reframes the whole problem. You are not choosing between helping and not helping. You are choosing between helping and generating an argument the other person has silently, with you, while reading.

The four things that reliably go wrong

Positive reframing, for some people. Marigold, Cavallo, Holmes and Wood ran six studies including a confederate interaction⁶. People with low self-esteem were less receptive than others to support that positively reframed their experience, meaning cheering up, optimistic spin, minimising. They were equally receptive to support that validated their negative feelings.

The trap in that study is worth the whole paper. In real interactions, friends offered positive reframing to everyone equally, and offered validation less to the people with low self-esteem. So the people who needed validation most got it least. And when the reframing was rejected, the providers felt worse about the interaction, about themselves and about the friendship.

Note the moderator. This is specifically about low-self-esteem recipients, and it should not be inflated into a universal law about silver linings.

Advice given first. Two literatures converge. Unasked-for support is rated unpleasant, and the mechanism is the recipient inferring they are being seen as incompetent⁷. Advice Response Theory holds that advice threatens both the desire to be approved of and the desire for autonomy⁸. Advice offered before emotional validation generates that threat and gets rejected. The same advice after validation is received completely differently.

So the actionable rule is not “never advise”. It is “never advise first.”

Minimising. Documented mainly through its consequences and through the reactance mechanism above. It invalidates the severity of the experience and signals distance.

Social comparison. The weakest evidence base of the four. What exists traces to an unpublished thesis rather than a peer-reviewed article, so treat it as suggestive. Conceptually it violates person-centredness by moving the focus off the person.

And one piece of folk wisdom that does not hold up. Rimé’s programme established that people overwhelmingly do share emotional experiences, and that the review evidence was generally not consistent with the view that sharing alleviates the memory of what was shared. Venting reliably happens and meets social needs. It does not by itself reduce distress.

This is roughly the shape of what Subtext checks in a draft: whether the message validates before it advises, whether it reframes something the person has not yet finished feeling, and whether the version you wrote at speed sits at the bottom of that scale without your noticing. It is not writing the message for you. It is telling you which end of the range you landed on.

Free to start, on your phone or in the browser.

What bereaved people actually say helps

Lehman, Ellard and Wortman interviewed 94 people who had lost a spouse or a child in a road accident four to seven years earlier, alongside 100 controls describing what they would say⁹.

Most frequently mentioned as helpful: contact with someone who had been through something similar, and the opportunity to express feelings. Most frequently mentioned as unhelpful: giving advice, and encouraging recovery.

The paper’s sharpest conclusion is aimed at the excuse this article is named after. The authors judged the “people don’t know what to say” explanation implausible, because the control respondents, who had never faced the situation, produced perfectly sensible answers about what would help. Bystanders describe in the abstract exactly what works. They just do not do it when it is real.

A separate study of 55 cancer patients found what counts as helpful is source-specific¹⁰. Emotional and esteem support was valued most from close ties, information from medical experts, and similar-other contact from fellow patients. Friends and acquaintances were disproportionately cited for social avoidance, meaning reduced contact and redirected conversation.

The one rule worth knowing

“Comfort in, dump out” comes from a 2013 newspaper op-ed by a clinical psychologist and a mediator¹¹. Draw concentric circles with the person in crisis at the centre and each wider ring holding less-close people. Comfort flows inward. Venting flows outward, to rings wider than your own.

It came from the psychologist’s own experience as a breast cancer patient, when peripheral colleagues expressed their distress to her and she ended up comforting them about her own illness.

It has never been empirically tested. No trial, no peer-reviewed evaluation. It circulates in hospice and grief education because it is intuitive and memorable. That does not make it wrong; it makes it a hypothesis with unusually good face validity. And it is one of very few pieces of popular advice designed to discipline the sender rather than optimise the message.

Does “I don’t know what to say” work?

There is no direct study of the phrase. I looked, and the sources that assert it is received well cite a handbook chapter that does not test it.

What can be said defensibly: bereaved people rate acknowledgement of the loss and permission to express feelings as helpful⁹, and a moderate message performs as well as an elaborate one³. So a plain, honest, unelaborate message is not a compromise.

The phrase works as honesty paired with acknowledgement and presence. “I don’t know what to say, but I’m thinking of you and I’m not going anywhere.” It does not work as a substitute for engaging, and the Lehman finding specifically undercuts using it as an excuse.

So, practically

Send something. That is the finding with the most direct evidence behind it, and it beats the alternative even when what you send is poor.

Say you are sorry, name the specific thing, and stop. Do not reframe it as a blessing in disguise. Do not offer advice before you have acknowledged the feeling. Do not compare it to something that happened to you.

Say what you will actually do, and make it concrete and small. And then do the unglamorous thing that the research on nonsupport keeps pointing at, which is send another one next week, when everyone else has stopped.

Sources

Numbered in the order they appear above. Where a figure could not be verified against the primary source, the text says so.

1. Ray, C. D., and Floyd, K. (2025). Saying the Wrong Thing or Saying Nothing at All: Comparing Outcomes of Nonsupport and Low Person-Centered Support Messages. Communication Studies, advance online publication. N = 886 US adults, census-matched on age, sex and ethnicity. Survey experiment using scenarios. https://doi.org/10.1080/10510974.2025.2532497
2. Ray, C. D. (2024). Nonsupport Experiences of Young Adult Cancer Patients: Prevalence, Acceptability, and Comparisons to Supportive Experiences. Health Communication, 39(6), 1127 to 1139. N = 205 young adult cancer patients across 22 countries. Cross-sectional. https://doi.org/10.1080/10410236.2023.2206178
3. Tian, X., and Solomon, D. H. (2025). The Role of Person-Centered Messages, Parallel Disclosures, and Reactance When Communicating Support for Parental Death. Communication Research, 52(8), 1116 to 1144. Two experiments on messages about the death of a parent. https://doi.org/10.1177/00936502231200554
4. High, A. C., and Dillard, J. P. (2012). A review and meta-analysis of person-centered messages and social support outcomes. Communication Studies, 63(1), 99 to 118. Twenty-three studies. Effect sizes conflict across secondary sources and are not reported here. https://www.tandfonline.com/doi/abs/10.1080/10510974.2011.598208
5. Tian, X., Solomon, D. H., and St.Cyr Brisini, K. (2020). How the Comforting Process Fails. Journal of Communication, 70(1), 13 to 34. N = 325 married adults. Experimental vignette study. https://academic.oup.com/joc/article/70/1/13/5698286
6. Marigold, D. C., Cavallo, J. V., Holmes, J. G., and Wood, J. V. (2014). You can’t always give what you want. Journal of Personality and Social Psychology, 107(1), 56 to 80. Six studies including a confederate interaction. https://psycnet.apa.org/record/2014-22042-001
7. Smith, J., and Goodnow, J. J. (1999). Unasked-for support and unsolicited advice. Psychology and Aging, 14(1), 108 to 121. https://psycnet.apa.org/record/1999-13473-009
8. Feng, B., and MacGeorge, E. L. (2010). The influences of message and source factors on advice outcomes. Communication Research, 37(4), 553 to 575. See also MacGeorge, Feng, Butler and Budarz (2004), Human Communication Research, N = 280. https://journals.sagepub.com/doi/10.1177/0093650210368258
9. Lehman, D. R., Ellard, J. H., and Wortman, C. B. (1986). Social support for the bereaved. Journal of Consulting and Clinical Psychology, 54(4), 438 to 446. N = 94 bereaved (40 spouse, 54 child) plus 100 controls. Retrospective structured interviews. Percentage figures circulating for this study could not be verified and are not used. https://psycnet.apa.org/record/1986-27950-001
10. Dakof, G. A., and Taylor, S. E. (1990). Victims’ perceptions of social support. Journal of Personality and Social Psychology, 58(1), 80 to 89. N = 55 cancer patients. https://psycnet.apa.org/record/1990-13511-001
11. Silk, S., and Goldman, B. (2013). How not to say the wrong thing. Los Angeles Times, 7 April 2013. An op-ed, never empirically tested. https://www.latimes.com/opinion/op-ed/la-xpm-2013-apr-07-la-oe-0407-silk-ring-theory-20130407-story.html