Four of the numbers in every ad guide have no source, and the famous one is misread

Nine statistics turn up again and again in writing about ad creative, usually with no link and often in the same paragraph as each other. We went looking for where each one came from. Four of them dead-end: we could find no primary source, and none of the pages repeating them cite one. Four have a real source that says something materially different from the claim it is used to support. One checks out as written, and it is from 2012.

The most quoted of them — "47% of campaign value is delivered in the first three seconds" — is the interesting case. The study is real, it is Meta's, and its conclusion is close to the opposite of the slogan.

A note on method before the numbers, because it decides how much weight any of this carries. "No primary source found" is a statement about the citation trail, not about the truth of the number. We are not saying anyone invented anything. We are saying that we looked, and that the pages asserting these figures do not show their work — which means neither you nor we can check them.

The 47%: real study, wrong subject

The source exists and is easy to find once you know its name: Facebook IQ, The Value of Video for Brands, published 19 March 2015. Here is the sentence, in full:

The data shows that people who watched under three seconds of the video ad created up to 47 percent of the total campaign value, and people who watched for fewer than 10 seconds created up to 74 percent, depending on the metric.

Read that carefully. The 47% is the share of a campaign's total brand-lift value contributed by a group of people — those who watched less than three seconds — not a measure of what the first three seconds of the creative accomplish.

The distinction is not pedantic, and Meta's own description of the method is why:

Our Marketing Science team then looked at both the lift data from Nielsen and the number of people who watched for varying lengths of time to estimate campaign value at different lengths of view.

The estimate is built from lift data and cohort size. A number computed that way rises when more people watch briefly, not only when the opening frames work harder. It is a result about audience composition, quoted as a result about creative craft.

And Meta's conclusion pointed the other way entirely:

That means that while lift continued to increase the longer people watched, people didn't have to watch a whole video to be affected by the ad.

The study was reassuring advertisers that short views are not wasted. It now circulates as a warning that you forfeit half your results if the first three seconds fail. Same number, inverted lesson.

Four qualifiers get dropped in transit, all of them load-bearing: "up to" (a ceiling, not a central estimate), "depending on the metric", "estimate" (this was modelled, not directly measured), and the sample — 173 Nielsen BrandEffect studies of digital video on Facebook, and per the trade coverage, only campaigns that showed positive lift. Plus the date. 2015 is pre-Reels, describing a feed that no longer exists.

One more thing worth knowing if you go looking: the Meta page most often credited with this figure, Capturing Attention in Feed from 2016, does not contain it. It carries different numbers entirely.

The second 47%, which appears to be a mislabel from day one

There is a variant: "views under three seconds produced a 47% lift in ad recall, 32% in brand awareness and 44% in purchase intent."

That sentence traces to Marketing Land's story about the study, published the same day, 19 March 2015.

Here is the checkable claim: that sentence does not appear on Meta's page. Searching the primary source for "32", for "44", and for any sentence pairing "lift" with "47" returns nothing. What Meta published is one figure — "up to 47 percent … depending on the metric" — across exactly three metrics: ad recall, brand awareness and purchase consideration.

The most economical reading is that the 47/32/44 triplet is the by-metric breakdown of value share, relabelled as lift in the first news story. Two tells support it: 47 is the largest of the three, exactly as "up to … depending on the metric" implies, and the same sentence renames Meta's "purchase consideration" to "purchase intent". We cannot prove the mechanism — Meta's page may have carried a chart whose values were not in the text we read — so treat that as inference, and the absence from Meta's page as fact.

If it is right, it is the most useful thing in this audit. The error did not accumulate through a long tail of SEO blogs paraphrasing each other. It entered at the first hop, in a reputable trade outlet, on publication day. Everything downstream copied it faithfully.

The four we could not source at all

Two observations about that table.

The decimal in 92.1% is doing rhetorical work. One decimal place signals a specific measured dataset. Nothing we found supplies one.

The last two are the same claim, told twice, and they disagree with each other. At constant CPM, cost per click is arithmetically the inverse of click-through rate — double the CTR and you halve the CPC. So "4x CTR" and "50% lower CPC" are not two independent proof points. They are one effect stated two ways, and the arithmetic does not close: 4x CTR at constant CPM implies roughly 75% lower CPC, not 50%. A page citing both as separate evidence is double-counting one mechanism and getting the maths wrong in the process.

We looked for attributable figures to put beside these two, and what we found were vendor and platform self-reports an order of magnitude smaller. We are not going to quote them here, because we could not verify them to the standard this post is holding everyone else to, and a counterweight you can't check is just another unsourced number pointing the other way.

The 79%: the citation is dead, the source is not, and it measures something else

"79% of people say user-generated content highly impacts their purchasing decisions." This one is cited — to a Stackla report called Bridging the Gap, 2019. Follow the link every page uses and it redirects to Nosto (who acquired Stackla) and then 404s.

The report survives, gated, at a different Nosto URL, and the press release text is still readable. Nobody citing the figure has updated their link.

But the bigger problem is what it measures. It is a stated-preference survey — 1,590 consumers in three countries, reporting what they believe influences them — commissioned by a company selling user-generated-content software. It contains no click-through rate, no conversion rate, no lift. It is routinely deployed as evidence that UGC-style creative performs better than polished brand creative. It cannot support that. What people say influences them and what measurably moves them are different quantities, and the gap between them is most of what advertising research exists to study.

The 92%: accurate, and from 2012

"92% of consumers trust earned media above all other forms of advertising." This one checks out as written. It is Nielsen, April 2012, from a survey of more than 28,000 internet respondents in 56 countries.

Three drifts to watch. The population is an online panel, not consumers generally. The fieldwork is from 2011 and Nielsen has run later waves with different numbers. And most importantly, it gets restated as "92% of consumers trust UGC more than ads" — which changes the subject. Nielsen asked about earned media: word of mouth and recommendations from friends and family. Branded user-generated content inside a paid ad unit is not what anyone was asked about. In that restated form, the figure no longer matches its source.

The 20% text rule, which is not a rule and may not be a recommendation either

Two separate things get repeated as current Meta guidance: the policy that ads with more than 20% text in the image were rejected or throttled, and the recommendation that images under 20% text perform better.

The policy was removed in September 2020. That is well documented. The recommendation is the interesting half, because it is still quoted as Meta's current advice on a great many marketing sites.

Meta's Best practices for image ads page contains no "20%" anywhere, and says:

Note: There is no longer a limit on the amount of text that can exist in your ad image. The text overlay tool is no longer available.

Its surviving guidance on text is entirely qualitative — don't obstruct the visuals, use a clean font at a readable size with contrast. The help page that historically hosted the rule now renders as a safe-zone page and also contains no "20%". A search restricted to Meta's business domain for the exact recommendation returned no Meta pages at all — only third-party blogs attributing it to Meta.

That negative is scoped to the pages we checked rather than to all of Meta's documentation. But it holds in the form that matters: the pages that carried the rule no longer mention it, and we could not locate the recommendation on Meta's own domain.

(If you want the guidance that did replace it, the safe-zone rules are percentages, not pixels.)

The pattern

Nine numbers, and the failure modes repeat:

• A cohort result read as a creative result. The 47%.
• The conclusion inverted. A reassurance becomes a mandate.
• Stated preference sold as measured performance. The 79%.
• The subject quietly swapped. Earned media becomes UGC.
• A live number pointing at a dead link. The Stackla URL.
• The commissioning party sells the thing the finding recommends — almost never disclosed downstream.
• Quoting a page that no longer says it. The 20% rule.
• One mechanism counted twice, inconsistently. 4x CTR and 50% lower CPC.

None of these requires anyone to be dishonest. Every one of them is what happens when a number travels without its qualifiers, and each hop drops one more.

The practical version is short. If a statistic has no link, treat it as anecdote. If it has a link, open it — a surprising share of the time the source says something adjacent but different, and a surprising share of the time it does not load at all. If it has a decimal place, be more suspicious rather than less. And check the date: the most-quoted number in video advertising describes a feed that stopped existing years ago.

Sources

• The Value of Video for Brands — Facebook IQ, 19 March 2015. The "up to 47 percent of the total campaign value" sentence, the 74% at ten seconds, the Nielsen BrandEffect methodology (173 studies, test-and-control, three brand metrics), and the "people didn't have to watch a whole video" conclusion (accessed 2026-09-20)
• Even Brief Video Views Drive Brand Lift, Facebook/Nielsen Study Finds — Marketing Land (now MarTech), 19 March 2015. The "47% for ad recall, 32% for brand awareness and 44% for purchase intent" sentence, and the note that only campaigns showing positive lift were included (accessed 2026-09-20)
• Capturing Attention in Feed: The Science Behind Effective Video Creative — Facebook IQ, 20 April 2016. Checked because it is frequently miscredited as the home of the 47%; it carries different figures (accessed 2026-09-20)
• Stackla survey reveals disconnect between content consumers want and marketers deliver — MarTech Series, 25 February 2019, carrying the Stackla press release of 20 February 2019. The 79% figure and the sample of 1,590 consumers and 150 B2C marketers across the US, UK and Australia (accessed 2026-09-20)
• Report: Consumer and marketing perspectives on content in the digital age — Nosto. Where the Stackla report survives after the acquisition, gated, at a URL none of the citing pages use (accessed 2026-09-20)
• Global Trust in Advertising and Brand Messages — Nielsen, April 2012. The 92% earned-media figure and the 28,000-respondent, 56-country sample (accessed 2026-09-20)
• Best practices for image ads on Facebook and Instagram — Meta Business Help Center. The "no longer a limit on the amount of text" note and the qualitative text-overlay guidance (accessed 2026-09-20)

---

Every primary source above was opened and read; Meta's help pages are client-rendered, so they were read in a real browser rather than fetched. Where this post says no primary source was found, that means a targeted search did not locate one — it is a claim about the citation trail and not about whether the number is true. "Earliest appearance" means the earliest we located, not the earliest that exists. One reading — why the 47/32/44 triplet appears in the trade story and not on Meta's page — is inference rather than fact, and is marked as such where it appears.

Terrads is built on the same habit: when a spec or a rule matters, we go and read the platform's own page.