The most repeated statistic in brand storytelling has no source, and the chain breaks in a place you can point at.

The claim is that stories are 22 times more memorable than facts, usually credited to the psychologist Jerome Bruner. It appears in no publication of his. Three independent people have looked, including a specialist who owns three of his books and one researcher who read Actual Minds, Possible Worlds cover to cover. Nothing.

Here is where it comes from.

2007. A trade book on persuasion states, on page 125, that Bruner “estimates” a fact is “twenty times” more likely to be remembered if it is part of a story. Twenty, not 22. The word “estimates” already signals this is not a measurement. The book contains no bibliography and no notes.

2013. A workshop document produced under the Stanford Graduate School of Business banner opens with the sentence “Stories are up to 22 times more memorable than facts or figures alone.” The full six-page document contains zero references, zero footnotes, no bibliography, and no mention of Bruner anywhere. The number is stated as a bare fact.

Now. Marketing pages have begun attributing the figure to “Stanford neuroscience studies”. The person it was credited to has disappeared, and the institution launders what remains.

What Bruner actually wrote, in the book most often cited for this, is a distinction between two modes of thought, the paradigmatic and the narrative, which are validated differently: one by truth, the other by verisimilitude. It is an epistemological argument. There is no multiplier anywhere in it.

The career of this number is itself a demonstration of the thing the field claims. A well-formed story travels better than a fact, including when the story is false. Which is a good reason to know which parts of this literature hold.

The citation chain behind the claim that stories are twenty two times more memorable than factsThe citation chain behind the widely repeated claim that stories are twenty two times more memorable than facts, traced step by step. The earliest identified printed source is a two thousand seven trade book on persuasion, which states on page one hundred and twenty five that the psychologist Jerome Bruner estimates a fact is twenty times more likely to be remembered if it is part of a story. The figure at this stage is twenty rather than twenty two, the word estimates already indicates that this is not an experimental result, and the book contains no bibliography and no notes of any kind. The second step is a two thousand thirteen workshop document produced under the Stanford Graduate School of Business banner and distributed by an outside organization, which opens with the sentence that stories are up to twenty two times more memorable than facts or figures alone. The full six page document was extracted and checked and contains zero references, zero footnotes, no bibliography, and no mention of Bruner anywhere, so the number is stated as a bare fact. The third step is currently underway, as marketing pages have begun attributing the figure to Stanford neuroscience studies, meaning the person originally credited has disappeared and the institution now launders the claim. The figure appears in no publication by Bruner, and three independent investigations have failed to find it, including one by a specialist who owns three of his books and one by a reader who went through Actual Minds Possible Worlds in full. What Bruner actually wrote in that book is a distinction between a paradigmatic and a narrative mode of thought, validated respectively by truth and by verisimilitude, which is an epistemological argument containing no multiplier.A number with no measurement behind it2007, a trade book on persuasion, page 125Bruner “estimates” a fact is “twenty times” more likely to be remembered in a story.No bibliography. No notes.2013, a workshop document under a university banner”Stories are up to 22 times more memorable than facts or figures alone.”Zero references. No Bruner.Now, in current marketing pagesAttributed to “Stanford neuroscience studies”.The person has vanished.The claim appears in no publication by Bruner. Three independent investigations found nothing.
Three steps, none of them a measurement. The last one replaces the person with the institution. Source : Maxwell and Dickman, The Elements of Persuasion, HarperCollins, 2007, p. 125; Harnessing the Power of Stories, Stanford GSB and LeanIn.org, 2013 (2013)

The number that is actually measured

There is real evidence on narrative memory, and it is worth knowing because it is both smaller and far more solid.

A 2021 meta-analysis compared memory and comprehension of narrative against expository text across 37 unique studies, 150 effect sizes, more than 75 samples and 33,078 participants.

  • Overall advantage for narrative: Hedges g = 0.55, 95 percent confidence interval 0.31 to 0.79.
  • Memory specifically: g = 0.72.
  • Comprehension: g = 0.48.

An effect of 0.72 is medium to large. In plain terms, a randomly chosen reader of a narrative text beats a randomly chosen reader of an expository text roughly three times in four.

The important part is the moderator analysis, because everything came back non-significant. The advantage held whether the test was oral or written, whether participants listened or read, whether they were adults or children, whether the test was immediate or delayed, and, critically, whether or not content and text difficulty were experimentally equalized. The narrative advantage is not an artifact of narratives being easier or more interesting.

The authors report some evidence of publication bias on the memory tests specifically, and none on comprehension. They also warn explicitly against forcing all instructional material into narrative form.

So the honest sentence is: narrative material is meaningfully better remembered, by an amount that has been measured on 33,000 people, and that amount is nothing like a factor of 22.

One related number deserves the same treatment, because it is often offered as the serious version. A 1969 study found 93 percent recall against 13 percent for narrative chaining. But that protocol asks learners to invent a story linking arbitrary nouns. It is a mnemonic technique, not a demonstration that a message told as a story lands better. Different claim, different experiment.

Measured advantage of narrative over expository text for memory and comprehensionThe measured advantage of narrative text over expository text for memory and comprehension, from a two thousand twenty one meta analysis covering thirty seven unique studies, one hundred and fifty effect sizes, more than seventy five samples and thirty three thousand and seventy eight participants. The overall advantage for narrative is a Hedges g of zero point five five with a ninety five percent confidence interval from zero point three one to zero point seven nine. For memory specifically the advantage is a Hedges g of zero point seven two with a confidence interval from zero point four three to one point zero one. For comprehension the advantage is a Hedges g of zero point four eight with a confidence interval from zero point two one to zero point seven five. An effect of zero point seven two is medium to large, meaning that a randomly chosen reader of narrative text outperforms a randomly chosen reader of expository text roughly three times in four. Every moderator tested returned a non significant result, including whether the test was oral or written, whether participants listened or read, whether participants were adults or children, whether the test was immediate or delayed, and, most importantly, whether content and text difficulty were experimentally equalized, which means the narrative advantage is not an artifact of narratives being easier or more interesting. The authors report some evidence of publication bias for memory tests specifically, with an Egger regression coefficient of two point six eight and a probability of point zero one, and none for comprehension. They warn explicitly against forcing all instructional material into narrative form.The honest figureStandardized advantage of narrative over expository text, Hedges gMemory0.72Overall0.55Comprehension0.48Moderators tested, all non-significantOral or written test. Listening or reading. Adults or children. Immediate or delayed.Content controlled. Difficulty controlled.The advantage survives equalizing content and difficulty. It is not an artifact of easier texts.
33,078 participants. Every moderator came back non-significant, including whether content and difficulty were equalized. Source : Mar, Li, Nguyen and Ta, Memory and comprehension of narrative versus expository texts: A meta-analysis, Psychonomic Bulletin and Review 28(3), 2021 (2021)

What the field’s central instrument actually measures

The concept underneath most brand storytelling advice is narrative transportation, from a 2000 article in the Journal of Personality and Social Psychology. Four experiments, with 97, 69, 274 and 258 undergraduates.

The transportation scale has 15 items on a 7-point scale, with a Cronbach alpha of 0.76. The items ask whether you pictured the events, whether you were mentally involved, whether you found it easy to put out of your mind, whether it affected you emotionally, whether your mind wandered.

So the instrument measures self-reported mental imagery, attentional absorption and affect. It does not measure persuasion, memory or behavior. That matters when someone presents transportation research as evidence that stories sell.

Two design facts are worth knowing before anyone leans on this literature.

Transportation was mostly measured, not manipulated. In three of the four founding experiments, participants read the same text and were then split at the median on their reported transportation. Those are correlations. Only the fourth experiment manipulated transportation directly.

The instrument determines the size of the result. The 2014 meta-analysis of 76 articles and 132 effect sizes found a significant moderator: studies using the validated transportation scale report an average effect of 0.34, while studies using other scales report 0.56. When the measurement is standardized, the effect shrinks by roughly a third.

That same meta-analysis gives the cleanest picture of what transportation is associated with. Affective responses, 0.57. Attitudes, 0.44. Intentions, 0.31. Beliefs, 0.26. And reduction of critical thoughts, negative 0.20, which the authors themselves flag as the weakest of the six consequences they measured.

The chain stops at intentions. The authors state it plainly: they could not examine actual behavior “because of a lack of data”.

Strength of association between narrative transportation and each measured outcomeStrength of association between narrative transportation and each measured outcome, from a two thousand fourteen meta analysis published in the Journal of Consumer Research covering seventy six articles and one hundred and thirty two effect sizes, drawn from a search in five languages over the thirteen years following the founding article, with inter coder agreement at a Cohen’s kappa of point seven five. The reported correlations, corrected for reliability and weighted by inverse variance, are as follows. Affective responses show the strongest association at point five seven, which is a large effect. Attitudes follow at point four four, a medium effect. Intentions at point three one, also medium. Beliefs at point two six, a small effect. Narrative thoughts at point two zero. And reduction of critical thoughts at negative point two zero, which the authors explicitly describe as less significant than the effect on the other consequence variables, making it the weakest of the six. A significant moderator was the measurement instrument itself, since studies using the validated transportation scale report an average effect of point three four while studies using other scales report point five six, meaning the effect shrinks by roughly a third when measurement is standardized. The authors state that they could not uncover some potential consequences such as actual behavior because of a lack of data, and that formal mediation analysis was not feasible because of the low number of intercorrelations, so the causal chain terminates at stated intention with no behavioral outcome anywhere in the meta analytic record.What transportation is associated withReliability-corrected correlation, 76 articles and 132 effect sizesAffective responses.57Attitudes.44Intentions.31Beliefs.26Narrative thoughts.20Fewer critical thoughts-.20The authors call the critical-thoughts effect “less significant than the effect on the other consequence variables”.Actual behavior is absent: they could not examine it “because of a lack of data”.
Affect at the top, critical thinking at the bottom, and no behavior at all. The chain terminates at stated intention. Source : van Laer, de Ruyter, Visconti and Wetzels, The Extended Transportation-Imagery Model, Journal of Consumer Research 40(5), 2014 (2014)

The mechanism nobody has ever tested

The story the field tells about itself is a causal chain: a narrative transports you, transportation suppresses your counterarguing, and with your defenses down you are persuaded.

It is a good hypothesis. It has never been tested as a mediation, and each generation of researchers has said so in print.

The founding article, 2000, in a footnote: the low sample size combined with unreliable effects of transportation on beliefs “did not allow us to test mediation.”

The meta-analysis, 2014: formal mediation analysis “was also not feasible because of the low number of the intercorrelations.”

The meta-analysis dedicated to resistance, 2020: “due to the lack of primary data on the interrelations among constructs, we could not investigate narrative engagement as a mediating path.”

Twenty years, three attempts, no test.

The resistance effect itself, incidentally, is small. Across nine studies directly comparing narrative and non-narrative messages, with 4,866 participants, the effect on resistance was d = negative 0.213. The authors note that a trim-and-fill correction for publication bias would cut the estimate by about a quarter. And resistance in almost all of these studies is a self-report scale, not a count of the counterarguments people actually produced.

Story against data, measured head to head

The most directly useful study in this field restricted itself to experiments that put narrative evidence and statistical evidence in the same protocol, on the same question, with random assignment. Fifteen studies met that bar.

The results split by outcome, and the split is the finding:

  • On beliefs, statistical evidence wins: effect of negative 0.16, confidence interval from negative 0.29 to negative 0.03, p = .014. This is the only clearly significant result in the table.
  • On attitude, statistical evidence wins marginally: negative 0.11, p = .058.
  • On intention, narrative wins marginally: positive 0.10, p = .061.

The difference between outcomes is itself significant. In plain terms: numbers are better at changing what people think is true. Narrative is somewhat better at moving what they intend to do. Both effects are small, and neither substitutes for the other.

That is a more useful conclusion than either of the slogans it replaces. The story does not replace the proof, and it is not decoration on top of the proof. They act on different variables, and a case that needs both needs both.

There is one caveat and it complicates the pitch. Nobody has properly tested narrative alone against data alone against the two combined, in a full factorial, on the same question. The closest available evidence points the wrong way, and it comes from the paradigm below.

Relative persuasiveness of statistical and narrative evidence by outcome variableRelative persuasiveness of statistical evidence compared with narrative evidence, by outcome variable, from a meta analysis restricted to studies that compared the two directly within the same protocol on the same question with random assignment, of which fifteen met the criterion. A negative effect indicates that statistical evidence was more persuasive and a positive effect indicates that narrative evidence was more persuasive. On beliefs, across nine studies and one thousand five hundred and forty six participants, the effect was minus zero point one six with a ninety five percent confidence interval from minus zero point two nine to minus zero point zero three and a probability of point zero one four, meaning statistical evidence was significantly more persuasive, and this is the only clearly significant result in the analysis. On attitude, across five studies and one thousand eight hundred and fifty participants, the effect was minus zero point one one with a confidence interval from minus zero point two three to zero point zero zero and a probability of point zero five eight, favoring statistical evidence marginally. On intention, across seven studies and one thousand three hundred and thirty eight participants, the effect was plus zero point one zero with a confidence interval from zero point zero zero to zero point two zero and a probability of point zero six one, favoring narrative evidence marginally. The difference between outcome variables was itself significant with a between groups Q of eleven point nine seven and a probability of point zero zero three. Heterogeneity was non significant throughout and there was no difference between health topics and other topics. The practical conclusion is that numbers are better at changing what people believe to be true while narrative is somewhat better at moving stated intention, that both effects are small, and that neither substitutes for the other. The authors also state that the mediating role of affective responses has never been tested empirically and that the persistence of these effects over time is unknown.Which evidence wins, by what you are trying to movestatistics more persuasivenarrative more persuasiveBeliefs-0.16p = .014, k = 9Attitude-0.11p = .058, k = 5Intention+0.10p = .061, k = 7The difference between outcomes is itself significant. Both effects are small.Nobody has tested narrative alone against data alone against both combined, in a full factorial.
Fifteen studies with direct head-to-head comparison. Numbers move beliefs. Narrative moves intention. Neither does the other's work. Source : Zebregs, van den Putte, Neijens and de Graaf, Health Communication 30(3), 2015 (2015)

The experimental pillar is contested

The single most cited experiment supporting storytelling is the identifiable victim effect: people give more to one named, photographed child than to statistics about many.

The 2007 studies are careful work. Real money, real donations, an experimenter blind to the hypotheses, envelopes sealed in private. Across four studies, the identifiable victim raised giving from $1.21 to $2.12 in one study and from $1.14 to $2.38 in another.

But read all four. The main effect of identifiability was significant in studies 1 and 3 only. It was absent in studies 2 and 4.

And study 3 contains a result that the storytelling literature almost never quotes. Adding statistics to the identifiable victim did not add to the effect. It removed it: giving fell from $2.38 to $1.43, statistically indistinguishable from the purely statistical condition. The authors write that “statistical information dampens the inclination to give to an identifiable victim.”

Then the replication record. A 2016 meta-analysis of 41 studies put the overall effect at r = 0.05, already tiny. A 2023 paper reanalyzed that same corpus with a robust Bayesian method correcting for publication bias and reported moderate evidence of publication bias and strong evidence for the absence of a mean effect, with a model-averaged estimate of r = 0.002. Its own preregistered replication, with 1,004 participants, concluded: “We found no empirical support for the identifiable victim effect.”

Those authors are careful about their own limits, and so should anyone citing them: their replication used hypothetical donations online, not real money in a sealed envelope, and they explicitly caution against over-interpreting it against the original. Other groups have failed to replicate it too, in six of seven well-powered studies in one case. A few well-powered studies do find it.

The honest summary is that the experimental pillar of storytelling is contested and its corrected effect size is close to zero. Not disproved. Contested.

Successive estimates of the identifiable victim effect across the replication recordSuccessive estimates of the identifiable victim effect across its replication record. The original studies published in two thousand seven used real money and real donations, with an experimenter blind to the hypotheses and donations placed privately into sealed envelopes, and found that an identifiable victim raised average giving from one dollar twenty one to two dollars twelve in study one and from one dollar fourteen to two dollars thirty eight in study three. However the main effect of identifiability was significant only in studies one and three and was absent in studies two and four. Study three also found that adding statistical information to the identifiable victim removed the advantage rather than adding to it, with giving falling from two dollars thirty eight to one dollar forty three, statistically indistinguishable from the purely statistical condition, and the authors wrote that statistical information dampens the inclination to give to an identifiable victim. A meta analysis of forty one studies published in two thousand sixteen put the overall effect at a correlation of zero point zero five, described by its authors as significant but modest, and reliable mainly with a single victim, a photographed child in poverty with low responsibility for their situation, and a monetary request. A two thousand twenty three paper reanalyzed that corpus using robust Bayesian meta analysis with publication selection modelling and reported moderate evidence of publication bias and strong evidence for the absence of a mean effect, with a model averaged estimate of a correlation of zero point zero zero two and a confidence interval from zero to zero point zero zero four, noting that the most precise studies showed the smallest effects. The same paper’s preregistered replication with one thousand and four participants found no empirical support for the identifiable victim effect, with an eta squared of zero point zero zero zero. Its authors caution that their replication used hypothetical rather than real donations and was conducted online, and warn against over interpreting it against the original real money studies.The same effect, three times2016 meta-analysisr = .0541 studies. “Significant butmodest.”Same corpus, bias-correctedr = .002”strong evidence for theabsence of a mean effect”Preregistered replicationnullN = 1,004. “no empiricalsupport for the effect”The result the storytelling literature never quotes, from the original study 3Adding statistics to the identifiable victim did not add to the effect. It removed it: $2.38 fell to $1.43,indistinguishable from the purely statistical condition.The replication used hypothetical donations, and its authors caution against over-interpreting it.
From a small meta-analytic effect, to near zero after correcting for publication bias, to a null preregistered replication. Source : Small, Loewenstein and Slovic, OBHDP 102(2), 2007; Lee and Feeley, Social Influence 11(3), 2016; Maier et al., Collabra: Psychology 9(1), 2023 (2023)

Three more numbers worth retiring

“63 percent remember stories, 5 percent remember statistics.” This is a class exercise reported in a 2007 business book, in which students gave one-minute talks and were later asked what they recalled. No control group, no published sample size, no published method, no peer review. It is an anecdote about a classroom, presented as research.

“95 percent of purchase decisions are emotional.” The real source is a 2003 interview with a Harvard Business School professor, who estimated that about 95 percent of cognition occurs below conscious awareness. The marketing version performs three substitutions at once: cognition becomes purchase decision, unconscious becomes emotional, and an estimate becomes a study result.

“Stories release oxytocin.” The researcher this is credited to has published real work, and it contradicts the claim. In his 2013 experiment, administering oxytocin did increase donations to public service ads, but in the second experiment oxytocin levels did not change after exposure, p = .94. In his largest narrative study, with 163 participants, his own methods section states that oxytocin values “were not reliable to be included in the analysis” and that “None of the endocrine measures significantly predicted the decision to donate.” Peripheral oxytocin measurement in humans is contested in its validity, and his own lab hit that wall.

What this means if you sell to businesses

Set the expectations correctly and the technique is still worth having.

There is no evidence that brand storytelling wins B2B deals. The literature lives in charitable giving, public health campaigns, fiction reading and consumer advertising, mostly with undergraduates and online panels. The one B2B study located measures supplier trust and propensity to recommend. Not purchase. Not contract value.

The whole chain terminates at stated intention. Even the meta-analyses cannot report behavior, and they say so.

Use narrative where the evidence puts it. Memory, at a measured advantage of 0.72. Intention, at a small positive effect. Affect, at the largest association in the field.

Use proof where the evidence puts it. Beliefs. That is where statistical evidence beat narrative in the only head-to-head meta-analysis, and it is exactly what a buyer needs settled before anything else can happen.

And keep the two separable. A case that rests on a story cannot be checked. A case that rests only on numbers is harder to carry into a room you are not in. The version that works is a claim someone can verify, attached to an account of one real situation, in which the numbers appear as facts of that situation rather than as a slide. Which story you are allowed to tell twice is a separate question, answered by the mission, the promise and the values a narrative has to stay inside.

That is not a technique that needs a factor of 22 to justify it. It needs the discipline to say which part of your case is doing which job.