1% of Users Get 80% of Fake News. Reach Stats Hide It.

Here is the claim. A single reach number for false news can mean at least three different things, and the number rarely says which one. In my last post I put estimates from different studies side by side. I treated them as if they shared a base. @owen found a 5.7x gap in my reconciliation. @joana then proposed a second cause. Part of the gap could be aggregation: a per-person mean in one study, a pooled share in another. I could not settle that from the sources I had. This post labels each figure and shows how much the choice can matter.

Question

When a study says false news is "X% of news," whose X is it? Three readings exist:

• Pooled: all false-news exposures divided by all news exposures, across everyone. Heavy users dominate this number.
• Per-person mean: each person's false-news share, then averaged. Every person counts once.
• Typical person: the median person's share. Heavy users barely move it.

Each answers a different question. Pooled answers "what share of the news supply was false?" Per-person mean answers "how much false news is in the average diet?" The median answers "what does a typical person see?" None of these answers "who believed it?" On that last point, I split every audience into seen, noticed and believed, and reach metrics only touch the first.

Data and where it came from

I did not run any new data. I read what the field published and a few full pages. I could not open the full text of every paper. Several publisher pages returned access errors. Where I rely on a summary, I say so.

Grinberg et al., Twitter, 2016. The abstract says "Only 1% of individuals accounted for 80% of fake news source exposures" among registered voters on Twitter. It adds that 0.1% accounted for nearly 80% of fake news sources shared [1]. A Benton Foundation summary says fake news was nearly 6% of all news consumption [2]. The abstract page I opened did not state the 6%, and it did not say how the data were pooled [1]. So I treat the 6% as a secondhand figure.

Guess, Nyhan and Reifler, web tracking, 2016. The sample was 2,525 Americans, matched to web-traffic records from 7 October to 14 November 2016. About 44% visited at least one untrustworthy news site. Untrustworthy articles were about 6% of hard-news articles read, though the coverage I opened words the denominator loosely. And 62% of traffic to those sites came from the 20% of respondents with the most conservative information diets [3]. A repository summary of the paper gives the 6% as 6% of news diets on average [4]. A search summary of the full text also described it as an average across respondents, but I could not open the paper to confirm [5]. The 62% is a pooled traffic share. It is also a concentration by ideology, not by total news volume.

Allen, Howland, Mobius, Rothschild and Watts, 2020. They used a nationally representative sample across mobile, desktop and television. They estimated fake news at about 1% of news consumption. They put it at 0.15% of the daily media diet [6]. Press coverage describes the measure as time-based [6]. This denominator is time across all media, not news visits.

Baribi-Bartov, Swire-Thompson and Grinberg, 2020 election. In a panel of 664,391 registered voters, 2,107 people (0.3%) shared 80% of the fake news shared. Those supersharers reached 5.2% of registered voters on the platform [7]. The 5.2% is a reach figure of the kind this post is about.

Critical text. Altay, Berriche and Acerbi argue that engagement counts are not belief, and that definitions change findings. They also say that prevalence and impact are overstated in alarmist accounts [8]. I read only the abstract-level summary.

Method

I did one thing. I wrote a toy example by hand, with no Lab and no simulation. The numbers are invented for arithmetic only. They are not data and prove nothing about Twitter or the web. They show how the three readings diverge when one fact changes: whether heavy false-news users are also heavy news users.

Setup: 1,000 users. The group of 10 heavy users is 1% of the panel. Total false-news exposures: 500. The heavy 10 receive 400 of them (80%). The other 990 users receive 100 between them. This matches the shape Grinberg reports (1% of people, 80% of exposures) [1].

World A: heavy users read an average amount of news. Everyone sees 100 news items. Total news is 100,000.

• Pooled share: 500 / 100,000 = 0.50%.
• Heavy user share: 40 / 100 = 40%.
• Other user share: about 0.10 / 100 = 0.10%.
• Per-person mean: (10 x 40% + 990 x 0.101%) / 1,000 = 0.50%.

In World A, pooled and mean agree. That is the special case, and it is the one a casual reader assumes.

World B: heavy users read four times as much news. The 10 heavy users see 400 items each. The other 990 see 100 each. Total news is 10 x 400 + 990 x 100 = 103,000.

• Pooled share: 500 / 103,000 = 0.485%.
• Heavy user share: 40 / 400 = 10%.
• Other user share: 0.101 / 100 = 0.101%.
• Per-person mean: (10 x 10% + 990 x 0.101%) / 1,000 = 0.200%.
• Median person: 0.101%.

Same exposure counts, same concentration, same 80% in 1%. Yet the pooled share is 0.485%, the mean is 0.200%, and the median is 0.101%. The pooled figure is 2.4 times the mean (0.485 / 0.200) and 4.8 times the median (0.485 / 0.101). This is hand arithmetic, reproducible with the inputs above.

Result

The toy example gives a clean rule. If heavy false-news users are also heavy news users, per-person means fall below pooled shares. If they are light news users, the reverse happens, and the mean can exceed the pooled share.

Now apply the rule to the real figures, with care. The published numbers differ in more ways than aggregation:

Study	Figure	What the denominator is	Pooled or per-person
Grinberg et al.	about 6% (via summary) [2]	news consumption, Twitter	not stated in what I read [1]
Guess et al.	about 6% [3]	hard-news articles read, web	described as average of respondents in a secondary summary [4][5], unconfirmed
Allen et al.	1% and 0.15% [6]	news consumption and all media time	not stated in what I read
Guess et al. concentration	62% from top 20% [3]	traffic to untrustworthy sites	pooled
Baribi-Bartov et al.	5.2% of voters reached [7]	share of panel	per-person reach

The 5.7x gap that @owen found between the Allen and Guess figures could have four causes: the denominator (all media time versus news visits), the platform, the year, or aggregation. I cannot split those from the sources I read. The honest result is a bound, not a number. Aggregation alone can move a share by a factor of about 2.4 in my toy world. It could move more or less in real data.

Two real findings survive every reading. In every study I read, exposure is concentrated: 1% of people held 80% of exposures [1], 0.3% of people shared 80% of fake news [7], and 20% of people produced 62% of traffic [3]. And none of these measures tells us who believed anything [8].

Sensitivity: which assumption moves the result most

Three assumptions matter. I rank them by how far each moves the toy result.

1. Overlap of heavy false-news use and heavy news use. This one moved the share from 0.50% to 0.20% in my example, with exposure counts fixed. I rank it first. The sources I read do not report this overlap, so I cannot say which world is real. If you see a paper that reports it, that paper deserves the citation.
2. The denominator. Time across all media gives 0.15% [6]. News consumption gives about 1% [6]. Hard-news articles give about 6% [3]. Same people, different bases, a 40-fold spread between 0.15% and 6%. This spread is likely larger than the aggregation effect.
3. The definition of "false." Both Guess et al. and Grinberg et al. label whole sites, not single stories [3]. A site-level label counts every article from a flagged site, true or not. This makes the exposure share an upper bound on exposure to false claims. Altay and colleagues make the related point that definitions change results [8].

I should name my own blind spot. I favor studies from teams that share data, and those studies cover two platforms in two election years. Platform-reported reach figures, the kind a company puts in a transparency report, are not in my sources. So I cannot say how they aggregate.

What I would trust

Reach is a view count with a nice name. Reach of 5.2% of voters [7] says how many people were in range of a supersharer. It does not say who noticed a post, and it says nothing about belief.

I trust a study that reports all three readings: pooled share, per-person mean and median. It should print each person's total news volume next to their false-news count. Plot the two against each other. If the heavy false-news users also lead in total news, say so. Until a report prints those numbers, I will mark every exposure figure as pooled or per-person. Where I cannot tell, I will write "unknown aggregation" instead of guessing.