The 7-38-55 rule: the figure that keeps being misused in presentation advice

On CNN's Larry King Live on 26 May 2002, Dr Phil McGraw, talking about going for a job interview, said that 93 per cent of communication is nonverbal (Lapakko, 2007, p. 7). On 31 October 2024 ELLE Vietnam, a lifestyle magazine, defined the same rule in an interview guide: 7 per cent of communicated information through words, 38 per cent through tone and 55 per cent through posture and facial expression (ELLE Vietnam, 2024). Taken literally, 20 minutes of preparing a talk would give the words 1.4 minutes. What does the figure stand on? This post re-reads the two original experiments, Mehrabian's own limits, and a survey of how the figure spread.

What did the two 1967 experiments actually do?

They were two short laboratory studies in which listeners guessed whether a speaker liked or disliked someone. By the authors' own description, neither contained a talk or a full sentence.

The Wiener column rests on the description on page 249 of the Ferris paper; the original paper (Mehrabian and Wiener, 1967) I could not read. Its own example, as Ferris reports it: the word "scram", said in a tone independently judged to show a positive attitude towards the listener, is taken overall as positive (Mehrabian and Ferris, 1967, p. 249).

The rest of this section comes from the Mehrabian and Ferris paper itself. Three female speakers were asked to vary their tone while saying "maybe" so as to communicate like, neutrality or dislike towards an imagined listener, and three female models were photographed trying to show those attitudes on their faces: voice and face were both performed (Mehrabian and Ferris, 1967, p. 250). The channels were chosen so that the attitude read from the voice matched the attitude read from the face when each stood alone (Mehrabian and Ferris, 1967, p. 248). From the three speakers and three models the authors kept two of each for the combined stimuli (Mehrabian and Ferris, 1967, p. 250). Twenty people each rated all 36 voice-and-face combinations, that is 3 voices × 3 faces × 2 speakers × 2 models (Mehrabian and Ferris, 1967, pp. 250–251). By my count, only 12 of the 36 had voice and face at the same level of attitude: two thirds of the conditions were built so that the channels disagreed.

Where do 7, 38 and 55 come from?

From a proposal on the last page of the Ferris paper, not from an experiment with all three channels. Lapakko notes that the numbers are drawn from two studies, and that neither compared words, voice and face at the same time (Lapakko, 2007, p. 8).

When the channels clash, the face wins but the voice still pulls. On a scale of −3 to +3, a positive face with a negative voice averaged +0.20, and a negative face with a positive voice −0.91 (Mehrabian and Ferris, 1967, p. 251). The paper's regression is AT = 1.50 AF + 1.03 AV, where AF and AV are coded +1, 0 and −1 for positive, neutral and negative attitude; the face factor accounted for 41.4 per cent of the total sum of squares and the voice factor for 19.3 per cent (Mehrabian and Ferris, 1967, p. 251). In my reading, 1.50 and 1.03 are slopes in rating-scale points, not shares of a message.

My own arithmetic: 55 divided by 38 is 1.45, close to 1.50 divided by 1.03, which is 1.46, so the face-to-voice ratio seems to come from here. The Ferris paper does not show the calculation, and I could not find in it how the 0.07 for words was obtained.

The decisive sentence is on page 252: "It is suggested that the combined effect of simultaneous verbal, vocal, and facial attitude communications is a weighted sum of their independent effects", with coefficients .07, .38 and .55. The paper then says that Anderson's procedures could be used to test "this proposed weighted-sum model for any single decoder" (Mehrabian and Ferris, 1967, p. 252). The three coefficients are a proposed model, with testing left for later. The Wiener paper is cited as the source of the words-versus-voice result; I could not read it (see the last section).

Mehrabian's book Silent Messages (1971) prints the equation as "Total liking = 7% verbal liking + 38% vocal liking + 55% facial liking" (Mehrabian, 1971, p. 43), then rewrites it for any feeling, adding "Numerical values in this equation are only approximate" (Mehrabian, 1971, p. 44). It also says the scoring procedure "is of course artificial and could not readily be used outside the laboratory" (Mehrabian, 1971, p. 45).

Where does Mehrabian himself limit the figure?

To feelings and attitudes, and to cases where the words conflict with the expression. On the page describing Silent Messages (undated, with a 1995 to 2016 copyright line; I read it on 3 November 2026) he writes that the equations "were derived from experiments dealing with communications of feelings and attitudes (i.e., like-dislike)", and that "unless a communicator is talking about their feelings or attitudes, these equations are not applicable" (Mehrabian, n.d.).

In the book, the equation can be used for only one dimension of feeling at a time, such as like-dislike, or specific feelings such as joy or anxiety (Mehrabian, 1971, p. 44). In private correspondence of 1995 quoted by Lapakko he wrote: "Please remember that all my findings on inconsistent or redundant communications deal with communications of FEELINGS AND ATTITUDES" (Lapakko, 2007, p. 11). Also in 1995 he offered an example: you tell someone the eraser they are looking for is in the second right-hand drawer of your desk, and who could claim that the verbal part of that message is only 7 per cent (Lapakko, 2007, p. 8).

The book also extrapolates: when any nonverbal behaviour contradicts speech it is more likely to determine the total impact, including touch, distance, posture and gestures, which the author calls a "reasonably safe" generalisation (Mehrabian, 1971, p. 45). The two experiments contained no touch, distance or gestures.

How is the figure misused, according to Lapakko's survey?

Lapakko searched Google for pages saying communication is 93 per cent nonverbal on 5 January 2007, got more than 263,000 hits, and examined the first 100; after removing 21 pages, 79 remained (Lapakko, 2007, p. 9). Only four questioned the numbers and 75 passed them on as correct; 16 named Mehrabian as the source, 34 gave vague sources such as "one study at UCLA", 10 named another source and 19 gave none (Lapakko, 2007, pp. 10–11).

On substance, 63 pages used the figure as a general statement about communication, nine attached it to "communication effectiveness" and two to "total impact"; only one described it correctly as "emotional meaning" (Lapakko, 2007, pp. 11–12). Two sources Lapakko quotes stretch the 93 per cent to posture and appearance, although the Mehrabian and Ferris photographs showed the head only (Lapakko, 2007, p. 12; Mehrabian and Ferris, 1967, p. 250). One page turned it into audience memory, which the experiments did not measure (Lapakko, 2007, p. 12).

Limits: one search engine, a convenience sample, data from 2007, pages in English. I did not survey Vietnamese-language material; the ELLE page is simply an example I read in enough detail to judge. Its definition does not mention the condition of feelings and attitudes, and adds posture to the 55 per cent; a later piece of advice does match the original condition: when tone and body language contradict the words, listeners tend to believe tone and body language (ELLE Vietnam, 2024).

Why should the 93 per cent stay out of presentation advice?

Because it measures something else under other conditions. First, the question was attitude, not content: participants judged whether a speaker liked or disliked the listener, from a single word (Mehrabian and Ferris, 1967, p. 250), and Lapakko writes that with "maybe" as the only word the role of language was "largely irrelevant by design" (Lapakko, 2007, p. 8). Second, the conflict was deliberate, as counted above. My assumption is that in an ordinary talk words, voice and face mostly agree; none of the sources I read measures how often. Third, nobody was asked what they understood or remembered, and there was no talk, gesture or posture.

Lapakko offers a test question: if 93 per cent of meaning is nonverbal, what would be the point of learning a language (Lapakko, 2007, p. 14)? How to lay out slides for the audience is a different question, which I take up in the post on cognitive load.

What can a presenter take from this research?

Use it for the sentences in which you talk about feelings or attitudes, where your words could clash with your voice and face. The rest of the talk lies outside its scope.

Hypothetical example. A department head in Can Tho (the person and all details are my invention) reports on the third quarter to the board in 20 minutes. Three sentences in those 20 minutes concern attitude: thanks to the team, an apology for a large order delivered late, and "I believe in this plan". The rest is figures, reasons and a proposal.

The middle column is what the sources say and the right-hand column is my reasoning. None of the sources I read measures how much practising voice or practising wording makes a presentation more persuasive. This use suits you if your talk has a few sentences expressing an attitude towards the audience; it does not help when you need to split preparation time between content and delivery.

Going deeper: where did the figure come from, and where does this post stop?

The two experiments did not grow out of research on presenting. The Ferris paper rests on the "double bind" theory of schizophrenia of Bateson and colleagues: people who often receive contradictory attitude messages, the theory says, learn to reply with contradictory messages of their own (Mehrabian and Ferris, 1967, pp. 248–249). It ends by asking for further experiments with schizophrenics or children (Mehrabian and Ferris, 1967, p. 252). The sample was all female; the authors write that the model is "likely" to apply more broadly, which is a conjecture not tested in the paper (Mehrabian and Ferris, 1967, pp. 251–252).

The main critic is Lapakko (2007): the original research has serious methodological limitations, yet the numbers are still in many textbooks (Lapakko, 2007, p. 7). He quotes Burgoon (1985), who calls the 93 per cent estimate based on faulty analysis, and Hegstrom (1979), who writes that a formula like this "gives the impression that more is known" about the contribution of the channels than is actually known (Lapakko, 2007, p. 8).

Limits of this post. I did not read the Mehrabian and Wiener paper, so I cannot verify the number of participants, the word list, or how 0.07 was derived. I did not read the 1981 edition of Silent Messages, pages 75 to 80, which Mehrabian points to. I read no study after 1967 that measures the weight of words against voice, so I say nothing about whether the results replicate. This post does not say that voice and facial expression do not matter in presenting; it says these two experiments did not measure it. The library's glossary entry on the 7-38-55 rule and its misuse uses the same meaning as this post.

What did I read? Mehrabian and Ferris (1967) in full, from a PDF on a university course page, with page numbers as printed, 248 to 252. Lapakko (2007) in full, from the Minnesota State University, Mankato repository, with the journal's printed page numbers, 7 to 19. Mehrabian (1971) from a scan of the first edition: pages 42 to 45, the reference list and the preface closely, the remaining chapters only by keyword search. Mehrabian's web page and the ELLE article on 3 November 2026; for the ELLE article, the copy I retrieved began at the sentence "Tuy nhiên, bạn có thể…", and I did not read the Everygirl source the page credits. Mehrabian and Wiener (1967): not read; I checked only the bibliographic record in Crossref and PubMed, and everything about that paper is as described on page 249 of the Ferris paper.