attck-research

Evidence review: Emotional Copy and Engagement

Article: Emotional Copy and Engagement
Reviewed: November 2026 · Method version 2 · How these reviews work

How this review works

Each factual claim in the article was checked against the source it cites, then challenged by reviewers whose job was to find the strongest evidence against it: replication attempts, meta-analyses, later papers that cite the original work, and research from other fields that study the same mechanism. A separate reviewer weighed both sides, and a script re-checked every quoted source. Where a claim was overstated or misattributed, the article was corrected. The research below supports the article as it now reads.

Background research

Introduction

From January 2013 to April 2015, Upworthy ran a randomized trial on every article it published, testing headline and image packages on live readers, and ran more than 32,000 experiments in total.

Dates and count confirmed against the Scientific Data descriptor.

Its style spread across the web, and other publishers copied it: headlines that promised a feeling and held back the facts.

Matias et al. document imitation of Upworthy’s style and Facebook’s 2016 change aimed at it.

In August 2014, Facebook announced it would show fewer click-baiting headlines, the kind that encourage a click without saying much about what people will see.

In Facebook’s survey, people preferred headlines that helped them decide before clicking 80% of the time.

Stated preference, not observed behavior; the same post notes click-bait gets many clicks.

That two-year arc holds the whole problem with emotional copy. Feeling can move the first click, though not always in the direction a writer expects. Whether it moves anything after that depends on what the click leads to.

Negative words raised clicks in two independent analyses of the Upworthy archive; other emotion effects were mixed and small.

What people pass on

Jonah Berger and Katherine Milkman analyzed every New York Times article that appeared on the paper’s home page over three months and examined which ones made the paper’s most-emailed list.

Sample: home-page articles, August 30 to November 30, 2008.

Positive articles were shared more than negative ones, but direction alone didn’t explain the pattern. Energy mattered too.

Berger and Milkman show arousal adds to valence; whether it outranks valence varies by dataset.

Articles evoking awe, anger or anxiety (high-arousal emotions) were more viral; articles evoking sadness were less so.

Accurate as a report of the Times study; on Facebook, sadness was associated with more sharing (Science Advances, 2023).

The pattern held after controlling for how surprising, interesting or useful the articles were and how prominently they were featured, and follow-up experiments, which asked people how likely they were to share, pointed to a causal effect.

Lab studies used stated sharing intentions.

What people click

Robertson and colleagues analyzed Upworthy’s archive: 22,743 randomized headline tests, about 105,000 variations and more than 370 million impressions.

Negative words raised click-through and positive words lowered it. For a headline of average length, each additional negative word increased the click-through rate by 2.3%. A separate analysis of the same archive confirmed the negative-word effect but not the penalty for positive words.

Negative-word effect replicated in an independent registered report; positive-word penalty did not.

Their emotion-level results are where the two studies part ways. Words associated with sadness slightly increased clicks, while joy and fear words slightly decreased them, and anger had no significant effect. Each of these effects was under 1%.

Emotion effects were small and based on the NRC word list.

**Sadness, which slowed sharing at the Times, sped up clicking at Upworthy, though the two studies measured sadness differently and the click effect was small.

Outlet, era and method all differ between the two studies, but the bigger difference is what they measured. Robertson’s team makes the same point: a click is a private, low-cost act of curiosity, while a share is public and puts the sharer’s name on the content. The practical lesson is to decide which behavior you want before choosing the feeling, and to measure that behavior rather than a proxy.**

The private-click versus public-share explanation appears in Robertson et al.’s own discussion and is supported by Bright (2016), Barasch and Berger (2014) and Scholz et al. (2017). Caveats: the two studies measured sadness differently, the click effect was 0.7% per standard deviation, and on Facebook sadness went with more sharing (Science Advances, 2023).

Dataset: The Upworthy Research Archive

The archive’s exploratory set holds 4,873 tests and 22,666 headline packages with their click results, plus example code in Python and R, released under CC BY 4.0.

One caveat from the team: in 2024 it reported randomization problems in 22% of tests and advises against confirmatory use of those run from June 25, 2013 to January 10, 2014. After rechecking six published studies built on the archive, it found no prior findings materially affected.

Disclosure confirmed; reassurance about prior findings added.

The subject line

“We miss you 😢” is a sad, low-energy appeal, and Robertson’s headline results suggest sad words can lift clicks slightly, though that finding comes from news headlines, not email.

No experiment on sad appeals in win-back email subject lines was found.

The way out

Mathur and colleagues call shaming declines ‘confirmshaming’, using language and emotion to steer people away from a choice; their crawl of shopping sites found 169 instances, including ‘No thanks, I like paying full price.’

Term from Brignull’s dark-patterns taxonomy; Mathur et al. measured prevalence. Large experiments find loaded-language dark patterns work moderately well (Luguri and Strahilevitz, 2021).

The opening line (pull quote)

Kate Moran: ‘Tone of voice is the way we tell our users how we feel about our message, and it will influence how they’ll feel about our message, too.’

Quote verified.

What changed in the article

These edits were made on November 12, 2026 after the review.

Introduction: supported with caveats

Claim reviewed: Upworthy’s style spread across the web: headlines that promised a feeling and held back the facts.

Now reads: Its style spread across the web, and other publishers copied it: headlines that promised a feeling and held back the facts.

Why: Source confirms other publishers copied Upworthy’s headline style and that Facebook later demoted it. The description of that style as promising a feeling while holding back facts is the author’s summary, and the citation marker sits on the sentence before it. Now cites Munger (2019), All the News That’s Fit to Click, Political Communication.

Introduction: supported with caveats

Claim reviewed: Feeling reliably moves the first click; whether it moves anything after that depends on what the click leads to.

Now reads: That two-year arc holds the whole problem with emotional copy. Feeling can move the first click, though not always in the direction a writer expects. Whether it moves anything after that depends on what the click leads to.

Why: Emotional words do change click rates, but not reliably upward. In the article’s own source, positive, joy and fear words lowered clicks, anger had no effect, and all the emotion effects were small, so ‘reliably moves the first click’ overstates it.

What Emotional Copy Actually Changes: contested

Claim reviewed: Two properties of a feeling matter most in this research: direction (pleasant or unpleasant) and energy (activating or quieting); awe and anger are high energy, contentment and sadness low.

Now reads: Researchers often describe a feeling by two properties: its direction, pleasant or unpleasant, and its energy, whether it activates people or quiets them. Awe and anger are both high energy. Contentment and sadness are both low. Those two properties are a useful start but not the whole story: anger and fear share both, yet they affected clicks differently in the study below, and a large Facebook study found that naming the specific emotion predicted sharing better than either property.

Why: Describing a feeling by its direction and its energy is a well-established approach, and measures of bodily and brain response often fit two dimensions better than named emotions. But the article is about behavior, and there the evidence runs the other way: specific emotions predicted Facebook sharing better than direction or energy, fear and anger push judgments in opposite directions despite sharing both properties, and in the article’s own Upworthy study anger and fear behaved differently. Saying these two properties ‘matter most’ presents a live debate as settled. Now cites Barrett and Russell (1998), Independence and bipolarity in the structure of current affect, JPSP.

What people pass on: supported with caveats

Claim reviewed: Berger and Milkman analyzed every New York Times article published over three months and examined which ones made the most-emailed list.

Now reads: Jonah Berger and Katherine Milkman analyzed every New York Times article that appeared on the paper’s home page over three months and examined which ones made the paper’s most-emailed list.

Why: The paper used all New York Times articles that appeared on the home page over three months (6,956 articles), not every article the paper published.

What people pass on: supported with caveats

Claim reviewed: Positive articles were shared more than negative ones, but energy (arousal) mattered more than direction (valence).

Now reads: Positive articles were shared more than negative ones, but direction alone didn’t explain the pattern. Energy mattered too.

Why: The paper says positive content was more viral and that the role of emotion is ‘more complex than valence alone’, but it does not rank energy above direction. Video-sharing research supports arousal as the main driver; a large Facebook study found specific emotions mattered more than either dimension.

What people pass on: supported with caveats

Claim reviewed: The pattern held after controlling for how surprising, interesting or useful articles were and how prominently they were featured, and follow-up experiments showed the effect was causal.

Now reads: The pattern held after controlling for how surprising, interesting or useful the articles were and how prominently they were featured, and follow-up experiments, which asked people how likely they were to share, pointed to a causal effect.

Why: The controls and the causal experiments are in the paper. The experiments asked people how likely they would be to share, rather than measuring actual sharing, which the article doesn’t mention.

What people click: supported with caveats

Claim reviewed: Negative words raised click-through and positive words lowered it; for a headline of average length, each additional negative word increased click-through rate by 2.3%.

Now reads: Negative words raised click-through and positive words lowered it. For a headline of average length, each additional negative word increased the click-through rate by 2.3%. A separate analysis of the same archive confirmed the negative-word effect but not the penalty for positive words.

Why: The study reports this exactly, and an independent registered analysis of the same archive confirmed the negative-word effect. That analysis did not find the penalty for positive words, so that half is less settled. Now cites Gligoric, Lifchits, West, Anderson (2023), Linguistic effects on news headline success, PLOS ONE.

What people click: supported with caveats

Claim reviewed: Words associated with sadness increased clicks, joy and fear words decreased them, and anger had no significant effect.

Now reads: Their emotion-level results are where the two studies part ways. Words associated with sadness slightly increased clicks, while joy and fear words slightly decreased them, and anger had no significant effect. Each of these effects was under 1%.

Why: The study reports these directions, but they come from a secondary, dictionary-based analysis and the effects are tiny, under 1% change in the odds of a click per standard deviation. Readers may take them as larger than they are.

What people click: supported with caveats

Claim reviewed: Sadness slowed sharing at the Times but sped up clicking at Upworthy, and the bigger difference between the studies is what they measured: a click is a private, low-cost act of curiosity, while a share puts the sharer’s name on the content.

Now reads: Sadness, which slowed sharing at the Times, sped up clicking at Upworthy, though the two studies measured sadness differently and the click effect was small.

Outlet, era and method all differ between the two studies, but the bigger difference is what they measured. Robertson’s team makes the same point: a click is a private, low-cost act of curiosity, while a share is public and puts the sharer’s name on the content. The practical lesson is to decide which behavior you want before choosing the feeling, and to measure that behavior rather than a proxy.

Why: The explanation that clicks are private and shares are public is not the author’s guess: Robertson’s team makes the same argument in the cited paper’s discussion, and independent work on news reading versus sharing, audience size and the brain’s valuation of sharing backs it. What still needs care is the sadness contrast itself: the two studies measured sadness in different ways, the Upworthy click effect was under 1%, and a large Facebook study linked sadness to more sharing, not less. The paragraph also carries no citation marker. Now cites Robertson et al. (2023), Negativity drives online news consumption, Nature Human Behaviour (existing source 4; add the marker to this paragraph).

Dataset: The Upworthy Research Archive: supported with caveats

Claim reviewed: In 2024 the team reported randomization problems in 22% of tests and advises against confirmatory use of tests run from June 25, 2013 to January 10, 2014.

Now reads: One caveat from the team: in 2024 it reported randomization problems in 22% of tests and advises against confirmatory use of those run from June 25, 2013 to January 10, 2014. After rechecking six published studies built on the archive, it found no prior findings materially affected.

Why: The 22% figure and the date range are correct. The same page adds that the team checked six published studies built on the archive and found no prior findings materially affected, which matters because the article relies on one of them.

The subject line: supported with caveats

Claim reviewed: A sad, low-energy subject line such as ‘We miss you’ may lift clicks, as Robertson’s headline results suggest sad words can.

Now reads: “We miss you 😢” is a sad, low-energy appeal, and Robertson’s headline results suggest sad words can lift clicks slightly, though that finding comes from news headlines, not email.

Why: The sadness effect is real in the headline data but very small, and it comes from news headlines, not app emails to lapsed users. The article hedges with ‘suggest’ and ‘can’, which keeps this mostly fair.

Made to Stick (book box): contested

Claim reviewed: The ‘Emotional’ chapter of Made to Stick shows why one person moves readers more than a statistic.

Now reads: Chip and Dan Heath explain why some messages are remembered and passed on, drawing on psychology experiments throughout. The “Emotional” chapter argues that one person moves readers more than a statistic, and shows how to make people care without manufacturing a feeling. Later research finds that effect smaller and less reliable than the book suggests.

Why: The book does make this argument: its ‘Emotional’ chapter builds on the Rokia experiment, in which an identified child drew far more donations than statistics about millions. But the line presents that effect as a settled fact, and it is not: a meta-analysis of 41 studies found it small and reliable only in narrow conditions, a two-study replication with more than 2,000 participants found no difference between one victim and many, a 2026 meta-analysis of 266 effects found no overall effect, and meta-analyses of narrative versus statistical evidence are split. Now cites Lee and Feeley (2016), The identifiable victim effect: a meta-analytic review, Social Influence.