Five video statistics, chased to the document under them

The 95 per cent retention claim was 200 B2B buyers, and Insivia says so under the headline

Insivia is a marketing agency, not a laboratory, and now the most honest source on this number. The disclaimer sits under the headline that spread the claim: a small survey, recall of short video against written text, and "Those exact percentages represented a pattern in our survey, not universal laws of cognition." No randomisation stated, no instrument named, nothing published, and recall the participants reported themselves. Insivia dates the survey to "over 16 years ago" and says what happened next: "Over time, that claim took on a life of its own."

What it turned into is a neuroscience finding. The 95 versus 10 pair travels with the word retention, no denominator and a brain graphic, into decks that would never cite a 200 person agency survey if the sentence said so. It was not invented. It was promoted.

A randomised comparison points the other way. In a 2026 paper, Ela Bakmaz and Nina Pavlin-Bernardic randomly assigned 194 university students to a video or a text condition and found the text learners "achieved higher scores on the knowledge test", with neither cognitive load nor interest explaining the gap. One study, one population of students, so the reading that survives is that the direction is contested. A duller sentence than 95 against 10, and the whole reason nobody puts it in a deck.

Forrester's post says 50, not 53, and Nate Elliott wrote it

Four things are wrong with the sentence everybody quotes, and the primary shows all four. Start with the byline: "Posted by Nate Elliott on January 8, 2009". The claim is credited almost everywhere to James McQuivey, who did not write it.

The figure in the text is 50, not 53, and it arrives with a clause attached: "on the keywords for which Google offers video results, any given video in the index stands about a 50 times better chance of appearing on the first page of results than any given text page in the index." That clause is the third error: the popular version drops it, and the finding only existed on keywords where Google already returned video.

The method was "a list of 40 of the most-searched keywords", run once on Google US, Google UK, MSN UK and Yahoo UK. The arithmetic is in the post: 16,000 videos on average competing for 1.5 video results, about 11,000 to 1, against 4.7 million text pages competing for 9.4 text results, about 500,000 to 1. That compares one video in Google's index with one text page in it. It says nothing about what embedding a video does to your own page, the fourth error and the reason anybody quotes it.

Elliott flagged it himself on 8 August 2012, in a bracketed note at the top of his own post: "these findings are now almost 4 years old, and are almost certainly no longer accurate." Fifteen months later Search Engine People published that "you are 53 times more likely to appear on Google page one if you have a video on your website", and linked its source, the post carrying that note. 50 became 53, "any given video in the index" became "you", and the qualifier disappeared.

The claim is not a lie somebody told, it is a measurement that lost its qualifiers one republication at a time.

A minute of video is worth 1.8 million words because a 1921 advertisement said a thousand

The report exists. "How Video Will Take Over The World", by James McQuivey, PhD, published 17 June 2008, is a real Forrester document, and the string 1.8 million appears nowhere on its landing page. The body sits behind Forrester's paywall, so the true statement is narrow: the figure is credited to that report, and the one Forrester page anybody can read does not contain it. A Marketwired release of 17 April 2014 passes it on as a bare quotation, no method and no page number.

The frame rate is the interesting omission, because nobody publishing the figure shows a sum and only one sum produces it. A thousand words per picture, thirty frames a second, sixty seconds. That makes the thousand words the load bearing input, and Quote Investigator traces that to Arthur Brisbane in The Post-Standard of 28 March 1911 and to Fred R. Barnard's advertisement in Printers' Ink of 8 December 1921, headed "One Look Is Worth a Thousand Words" and credited to a famous Japanese philosopher. Barnard reran it in 1927 at ten thousand words, as a Chinese proverb. Quote Investigator on both: "The evidence suggesting a Japanese or Chinese origin occurred rather late and is weak."

Recompute at 24 frames a second, the cinema rate, and the same minute is worth 1.44 million words. The number moves with the delivery format, which no measurement of how people take in information would do.

The 85 per cent came from three publishers and an agency, and Facebook supplied no figure

The article is Sahil Patel in Digiday, 17 May 2016, and it attributes its own figure in plain words: "As much as 85 percent of video views happen with the sound off, according to multiple publishers." Publishers, named in the piece. LittleThings said "Eighty-five percent of its viewership is occurring without users turning the sound on." Mic said "85 percent of its 30-second views are without sound." Facebook supplied no figure, and the article never says it did. The fourth figure is not even a publisher: MEC North America is a media agency, and "branded videos from its clients average 85-90 percent silent video views." So the platform statistic is three content sites and one agency describing inventory they sell, and none says what counts as a silent view or how it was counted.

The disagreement inside the article is the part to keep. PopSugar told Digiday its silent video views ranged between 50 and 80 per cent, which does not fit the number the piece is built around, and a spread that wide across four self reports is the finding. The muted feed is real, and this journal has covered it on surveys and controlled tests. The 85 per cent is not one of those.

The one number that behaves like evidence is the one whose URL stopped serving it

Cisco's Visual Networking Index forecast for 2017 to 2022, document C11-741490-00, dated 02/19, does what the other four do not. It states the claim in one quotable sentence: "Globally, IP video traffic will be 82 percent of all IP traffic (both business and consumer) by 2022, up from 75 percent in 2017." It defines what it counts, "the sum of all forms of IP video", from IP VoD to file sharing, streamed gaming and video conferencing. It names its inputs, from SNL Kagan and Ovum to comScore, Nielsen and the ITU. And it names the reason it does not expect the share to climb higher: "gaming is one of the forms of traffic that will limit the likelihood that video traffic will exceed the projected 82 percent by 2022."

It is also a forecast, a different object from a measurement, and its methodology appendix says what went into it: analyst projections, in-house estimates and direct data collection. Saying video is 82 per cent of internet traffic today is not what the document claims.

Then the citation rotted. The Internet Archive captured that cisco.com URL on 17 February 2020 serving the paper with the 82 per cent sentence intact, so the reference was once correct. It does not return the paper now. The document number now addresses the Annual Internet Report for 2018 to 2023, which is live and does not contain it. The primary text is readable today from a CERN mirror of the PDF and from the Archive.

The number outlived its document, and the citation under it now resolves to a different Cisco report.

Two of the five came from somebody running something, and both authors have since warned about it

Sort the five by what sits underneath. One has a stated sample and a warning from its author above it. One has 40 keywords in 2009 and a note from the man who ran it saying the result is out of date. One has four interested parties reporting on inventory they sell. One has no sample at all: an advertisement from 1921 and a multiplication. One has a methodology appendix, a document number and a citation that no longer resolves.

So the question is not whether a number is true. It is whether anybody ran anything, whether they published what they ran it on, and whether they would still defend the reading going into your deck. Two of the five pass the first two and fail the third, the commonest shape a bad statistic has: a real measurement, still circulating, past the date its author would defend it.

What to do about it

1. Before a statistic goes in a deck, find the first document that carried it and quote its sample and its year in the same sentence as the number.
2. Check that the cited URL still returns the document, and keep the copy you read, because two of the five primaries here survive only on a mirror or in the Internet Archive.
3. Drop any figure whose value would change if the frame rate or the format changed, because that is arithmetic on a proverb rather than a measurement.