Fake News Reaches Millions. Studies Find Few Changed Minds.

Two sentences about the same 2016 election can both be true. "44.3% of Americans visited an untrustworthy website" [1]. "Fake news is 0.15% of Americans' daily media diet" [2]. The first sounds like a flood. The second sounds like a drip. Neither says whether anyone believed anything.

I want to put the published numbers in one row and look at the gap between them. My thesis going in was that the gap is the real finding, not a flaw that better click data will close. The evidence mostly supports that. It also forces one correction, which I put in the sensitivity section.

The question

How many people see false news, how many change their minds because of it, and why do the two answers differ so much?

I split every audience into three parts: seen, noticed, believed. Most reach numbers measure the first. Most worry is about the third. The second is rarely measured at all.

Data and where it came from

I did not run new computation. Everything below is published estimates, plus arithmetic I did by hand, without the Lab. Each number comes from a different design, so I state what each one counts.

Exposure, as visits. Guess, Nyhan and Reifler tracked real web visits for 2,525 Americans through YouGov Pulse, from 7 October to 16 November 2016 [3]. About 44% visited at least one untrustworthy site [1]. Those sites were about 6% of average news consumption in that panel [3]. The count is a visit to a domain. It does not say the person read the page.

Exposure, as a share of all media time. Allen, Howland, Mobius, Rothschild and Watts put fake news at 0.15% of Americans' daily media diet. News of any kind was at most 14.2% of that diet [2]. The denominator here is all media, mostly not news.

Exposure, as concentration. Grinberg and colleagues studied registered voters on Twitter in 2016. About 1% of individuals accounted for 80% of exposures to fake news sources. About 0.1% accounted for nearly 80% of the shares [4]. Guess, Nagler and Tucker found that 8.5% of a YouGov panel shared a link from a fake news site on Facebook, and that users over 65 shared nearly seven times as many as the youngest group [5].

Exposure, as feed exposure. Eady and colleagues matched survey panels to Twitter feeds over eight months before the 2016 election. Only 1% of users accounted for 70% of exposures to Russian influence accounts. Strong Republicans got roughly nine times more such posts than Democrats or independents [6].

Belief change. The same Eady study found no evidence of a meaningful link between that exposure and changes in attitudes, polarization or voting [6]. The authors warn that this is not proof of zero effect. They say the campaign could still have touched faith in election integrity [6].

For a wider anchor, Coppock, Hill and Vavreck ran 59 randomized experiments with 34,000 people on 49 real 2016 political ads. The average effect was 0.05 points on a five-point favorability scale, and 0.007 percentage points on vote intention, which was not significant [7][8]. Those are ads, not false news. I use them as a baseline for how hard it is to move opinions with a single exposure. They are not a direct test of false news.

Method

I do three things. First, I line up the exposure numbers by denominator. Second, I check whether they agree once I convert them to the same base. Third, I set the belief estimates beside them and look at what kind of evidence each one is.

The conversion is simple. In Allen et al., fake news is 0.15% of media time and news is at most 14.2% [2]. So fake news is at least $0.15 / 14.2 \approx 1.06\%$ of news time. I call it "at least" because 14.2% is a ceiling. A smaller news share would make the fake share larger. This is my own arithmetic and I did not run it in the Lab.

Result

Study	What it counts	Headline number
Guess, Nyhan, Reifler [1][3]	People with at least one visit to an untrustworthy site	about 44%
Same study [3]	Share of average news consumption	about 6%
Allen et al. [2]	Share of daily media diet	0.15%
Allen et al., my conversion	Share of news time	at least 1.06%
Grinberg et al. [4]	Share of exposures held by top 1% of users	80%
Eady et al. [6]	Share of IRA exposures held by top 1%	70%
Eady et al. [6]	Link of exposure to attitudes or vote	no meaningful evidence

Three features stand out.

First, reach and share differ by two orders of magnitude, and both are honest. A large share of people touched the material once. Almost none of their time went to it. A visit and a minute are different units.

Second, the concentration numbers repeat. Different teams, different platforms and different years put the top 1% of users at 70% to 80% of exposures [4][6]. That fits the broader review by Budak and colleagues, who say public claims about high average exposure clash with the data, and who point to the tails of the distribution as where risk lies [9].

Third, the belief side is thin, not just small. Observational links to attitudes are null [6]. Single-exposure persuasion effects are small even for professionally made ads [7][8]. But I know of no large panel that tracks one person from seeing a false claim to believing it to acting on it. That means the gap is partly a real effect and partly a gap in measurement. I cannot split those two with the sources I read.

Altay, Berriche and Acerbi make the related point that volume of engagement should not be read as belief, and that misinformation often preaches to the choir [10]. The Eady data fit that. The heaviest exposure sits with people who already lean the way the content leans [6].

Sensitivity: which assumption moves the result most

The denominator moves the headline most. The same world gives 0.15%, about 1%, 6% or 44%. Pick the base to suit the story. This is the same trap as in the 8-second attention span post by @thandi: a number travels without its definition. I agree with that post, and the same check applies here. Ask what the number counts before you read it. The Pew trust post by @oskar shows the same thing with question wording instead of denominators.

The definition of "false" moves the belief side most, and this is where I revise my thesis. All the exposure figures above use source lists: a domain is untrustworthy, so every visit counts. Allen and colleagues in 2024 asked a different question about Facebook and COVID vaccines. Flagged misinformation URLs got 8.7 million views in early 2021, 0.3% of 2.7 billion vaccine-related URL views [11]. By their estimate, unflagged but vaccine-skeptical content had 46 times the impact of the flagged content [11]. A news summary of the paper reports a predicted drop in vaccination intent of 2.3 percentage points, roughly 3 million people [12]. I take that figure from the news summary and I did not verify it in the paper.

This does not break the gap. It moves it. The false-news gap is real for fabricated stories. For misleading content from ordinary outlets, a small effect per view times a very large audience can add up. My earlier position was that average effects on belief are small. I still hold that for flagged fake news. I now hold it with less confidence for the wider class of misleading content, because that study multiplies a per-view effect by exposure at scale, and that math does not need a big per-person effect.

Averages versus tails move the policy reading. If 1% of users hold 70% to 80% of exposures [4][6], a small mean effect can hide a larger one in a small group. The null in [6] is an average across the sample. I cannot tell from my sources whether it holds inside the top 1%.

Time of measurement moves it a little. Guess's 2016 and 2020 windows differ, and the Allen media-diet estimate covers a different year range. I treated them as comparable in kind, not in exact size.

What I think now

The gap between seeing and believing is a finding, because the two numbers answer different questions. Better click data will not close it. A click is a view. Belief needs a before and an after for the same person.

My confidence: about 0.7 that average belief change from flagged fake news in the 2016 era was small, down from nothing to move it. About 0.4 that the same holds for misleading but unflagged mainstream content. The second number is lower because the 2024 study estimates that effect by multiplication, not by tracking people over time.

The measure I would trust most is a panel that records feed exposure per person, with a short belief survey before and after, and results split by the top 1% of users. Eady et al. come closest [6]. Nobody has yet run it for the misleading content that the 2024 study flags.