The digital comeback

Did the shift of Lula's campaign toward social media, content creators and fan communities change votes? We test five predictions with 700 reels read on Instagram, the vote in 442,000 polling stations, the connectivity of each municipality and the failures Meta admitted. Lula's reach exploded, but on the day after the first round, before the new strategy; the strategy brought volume. Polling stations with more young voters swung more, especially in better-connected cities, but the polls do not show young voters shifting more than others. There is no sign that reach preceded the polls, nor that Meta's failures brought down Lula's reach.

Daniel Pessini Sobreira · October 27, 2026 · Instagram reels read from the profiles, vote by polling station and electorate profile (TSE), access and coverage (Anatel), social media monitoring series

1. ×3.8reach per Lula reel, relative to Flávio's, on the day after the first round. The new strategy (Oct 7) added nothing to that jump.
2. +0.28 pointof swing per standard deviation more voters aged 16 to 24 at the polling station, within the same municipality; +0.20 in the least connected cities and +0.33 in the most connected.
3. |z| ≤ 1.3on the days of the failures admitted by Meta: no abnormal drop in Lula's relative reach.

Five predictions about the swing on social media

The analysis of what the polls and markets missed showed that, after the first round, Lula came to dominate the open networks, and it proposed the hypothesis that the campaign's shift to social media, creators and communities such as fan groups helped the comeback. The same analysis recorded the technical failures that Meta admitted. This article turns those ideas into predictions that can fail. They were written before the tests were run.

Digital campaigning, persuasion and algorithms

The campaign in the hybrid system. Chadwick (2017) describes a media system in which the logic of television and that of social networks blend: what goes viral online sets the press agenda, and vice versa. Kreiss (2016) and Stromer-Galley (2019) show how campaigns came to organize communication around data and networks, and Bossetta (2018) notes that each platform has its own architecture, which favors certain formats (short vertical video, in the case of Instagram and TikTok). Riedl, Lukito and Woolley (2023) add the role of political influencers, who lend a campaign the trust of audiences it could not reach alone.

Fans as citizens. The fandom literature shows that fan communities have organizational infrastructure (accounts, calendars, engagement targets) that can be converted into political action (BROUGH; SHRESTHOVA, 2012; HINCK, 2019), and that content circulates through the choices of those who share it, not only through the distribution by those who publish it (JENKINS; FORD; GREEN, 2013).

How much this changes votes. Here the literature is skeptical. In 49 field experiments, the average persuasive effect of campaign contact in US general elections was practically zero (KALLA; BROOCKMAN, 2018). Coppock (2023) shows that people change their minds in the direction of new information, but only slightly and similarly across all groups. The experiments run with Meta in the 2020 US election changed the feed algorithm of thousands of users and barely changed political attitudes (GUESS et al., 2023; NYHAN et al., 2023), although exposure is highly segregated by ideology (GONZÁLEZ-BAILÓN et al., 2023). Meta-analyses find a positive but modest relationship between social media use and participation (BOULIANNE, 2015; 2020).

Asymmetries. Freelon, Marwick and Kreiss (2020) show that the right and the left use social media in different ways: the right bets on its own media and platform manipulation, the left on street activism and hashtags. In Brazil, studies of WhatsApp in 2018 showed the strength of circulation in closed groups, beyond the reach of public metrics (RESENDE et al., 2019; EVANGELISTA; BRUNO, 2019). Toret and Guerrero (2026) describe the 2026 case as a "pedagogía digital" that sought to "hackear el algoritmo" without fake news or bots.

Reach exploded before the new strategy

We read on Instagram, on Oct 27, all the reels published by Lula and Flávio since the end of August (503 from Lula and 173 from Flávio between Sep 1 and Oct 24), with the views of each. The chart shows the daily median of views per reel, which does not depend on how many videos were published, and below it the reels per day.

In September, Lula's typical reel had between 1 and 3 million views; Flávio's rose over the month and reached 8 to 20 million in the last week before the first round. On Oct 5, Lula's typical reel jumped to 12 million and stayed at that level. A segmented regression of the log of Lula's median minus Flávio's, with a trend and two breaks, one at the first-round shock (Oct 5) and another at the new strategy (Oct 7), gives a jump of 1.33 at the shock (×3.8; Newey-West standard error of 0.21) and 0.00 at the strategy (standard error of 0.18).

The new strategy changed something else: volume. Lula published 7.9 reels per day before the first round and went to 12.1; Flávio, from 3.0 to 3.6. Since the reach per video of the two was similar in the runoff (median of 8.5 million for Lula and 11.8 for Flávio), Lula's advantage in the total, 2.47 billion views against 0.90, came from publishing more, not from reaching more per video. Prediction H2a is partly confirmed: there was a break, but at the first-round shock, not at the change of strategy, which increased volume.

The content

We classified the captions of Lula's reels by topic, using keywords. After the first round, all topics came to yield three to six times more per video, which indicates a change of audience, not of subject. The clearest change of subject was the contrast with the opponent: reels mentioning Flávio or the Bolsonaro family went from 25 to 74.

Reach did not precede the polls

If reach on social media moved votes, Lula's share of views should rise before the polling average. That is not what appears: the correlation between Lula's share of views (3-day average) and the polling average a few days later ranges from −0.56 to +0.22, is negative when the poll comes 0 to 3 days later and stays near zero at longer lags.

With less than twenty days of data and polls that arrive in waves, the test has little power, and a negative correlation may come from calendar coincidences (reach peaks at debates and weekends). What it allows us to state is more modest: there is no sign that the polls followed reach. Prediction H2b is not confirmed.

Young places swung more; in the polls, young voters did not

If the digital strategy moved votes, the effect should appear among those most exposed to social media. In the 442,458 polling stations that existed in both 2022 and 2026, comparing polling stations in the same municipality and controlling for age, education, blank and null votes and first-round margin and the 2022 swing at the same polling station, one standard deviation more voters aged 16 to 24 corresponds to a +0.28-point swing (standard error of 0.009).

This effect depends on the municipality's connectivity, measured by fixed broadband subscriptions per 100 inhabitants (Anatel). It goes from +0.20 in the quarter of least connected municipalities to +0.33 in the most connected. The interaction withstands controls for the municipality's size and education, so as not to confuse connection with big city: +0.102 point per standard deviation of connectivity (standard error of 0.014).

Young people did not go to the polls more on average, but they did more in connected cities (interaction of +0.117 point on turnout). The pattern is what one would expect if part of the swing had passed through social media: larger among those who use it most, and where access is easiest. It is also a small effect compared with that of the first-round blank and null votes, and connectivity may still carry income differences that the education control does not capture.

The polls, however, do not show the same thing. In 12 pollsters that published age breakdowns in more than one runoff wave (Datafolha among them), we compared the change in Lula's margin among the youngest voters (16 to 24, or 16 to 34 where that is the only cut) and among those aged 60 or older, from each pollster's first to its last wave. On average, young voters moved 4.0 points less toward Lula than older voters (standard error 3.6), and young voters moved more toward Lula in only 3 of the 12 pollsters. The large shift among young voters in Datafolha is the exception, not the rule. Age subgroups have 90 to 230 interviews, and the margin of error of each exceeds 6 points; even so, taken together they do not confirm that young voters were the group that changed the most.

The two pieces of evidence measure different things. The polling-station evidence measures places: stations with more young voters may also be stations in the outskirts and with recent residents, and the larger swing may come from anyone who votes there, of any age. The poll evidence measures people, but with small samples and in a runoff in which the polls themselves got the result wrong. Together, they support that the swing was larger in young, connected places, and do not support that young voters, as a group, changed their vote more than others.

Prediction H2c is only partly confirmed: in places, yes; in people, the polls do not show it.

Fandoms: spontaneous adherence, no measure of effect

There is no series of its own on fandoms, and Instagram does not show the volume of the #FandomsComLula hashtag. What the sources record is that the approach came from the fans: on the night of the first round, fan profiles offered their channels to Lula, with coordinated messages, and the team replied the next day; the organizers speak of about 500 fan clubs and 20 million followers combined, a number that could not be verified. There was prior contact with the government, at a ceremony at the Planalto on Aug 31. The theory of fandom as civic infrastructure (BROUGH; SHRESTHOVA, 2012; HINCK, 2019) describes the phenomenon well, but the available data do not allow separating the effect of the fandoms from the rest of the digital wave. Prediction H2d cannot be tested with public data; as a deliberate campaign strategy, it does not hold up.

Meta's failures do not show up in reach

Meta admitted technical failures that hit Lula's account: the ad account deactivated "by mistake" (Sep 23), an outage of about five hours (Oct 7), the view count interrupted (Oct 8) and boosting blocked (Oct 9 and 10). If they had reduced organic reach, Lula's relative reach should fall on those days below what the series model predicts. The chart shows each day's deviation, in standard deviations.

None of the five days leaves the normal band (the largest deviation is 1.3 standard deviations), and two of them are positive. The test measures the organic reach of reels, not ad delivery, in-app search or the effect of the boosting block, which may have been affected without showing up in views. It is consistent with the conclusion of the TSE rapporteur, who ordered equal treatment without seeing deliberate favoritism, and with the 2020 experiments, in which changing the algorithm changed behavior little (GUESS et al., 2023). Prediction H2e is not confirmed for organic reach.

What remains of the hypothesis

The comeback on social media was real, but its timeline tells a different story from the hypothesis. Lula's audience changed on the day after the first round, together with the mobilization described in the article O susto que mobilizou: the shock led millions of people to seek out, share and comment on Lula's videos before the campaign changed its format. The campaign knew how to ride the wave, posting much more and confronting its opponent, and the effect on the vote shows up in the places one would expect, the young polling stations of connected cities, without confirmation in the age breakdowns of the polls and on a small scale compared with what decided the election: the vote of those who had chosen no one in the first round.

How the numbers were calculated

• Reels. Read from the Lula and Flávio profiles on Instagram on Oct 27 (views, likes and comments shown that day; older videos keep accumulating). The date of each reel comes from the post code, converted to Brasília time. Topics by caption keywords (list in the code); the classification is approximate.
• Interrupted series. OLS of the log of the daily median (Lula − Flávio) with a linear trend and steps on Oct 5 and 7, Sep 1 to Oct 24; Newey-West standard error with 3 lags (BERNAL; CUMMINS; GASPARRINI, 2017).
• Polling stations. Panel of 2022 and 2026 polling stations from the article on mobilization (same code and same location), with the electorate profile by polling station (TSE, Jul 2026); variables as deviation from the municipality mean; WLS weighted by electorate; errors clustered by municipality. Connectivity: fixed broadband density (subscriptions per 100 inhabitants, Nov 2025) and 4G/5G coverage (Anatel, Meu Município).
• Limits. Reach is not a vote, and public metrics do not cover messaging apps, where Palver saw Flávio ahead or tied. The effects by polling station describe places, not people. Connectivity correlates with income and urbanization; the controls reduce, but do not eliminate, this confounding.
• Files. hipotese2.json; data in analise-hipotese2/dados/ and analise-previsoes/dados/; code in hipotese2_analise.py.

References

BENNETT, W. Lance; SEGERBERG, Alexandra. The logic of connective action: digital media and the personalization of contentious politics. Cambridge: Cambridge University Press, 2013.

BERNAL, James Lopez; CUMMINS, Steven; GASPARRINI, Antonio. Interrupted time series regression for the evaluation of public health interventions: a tutorial. International Journal of Epidemiology, v. 46, n. 1, p. 348-355, 2017.

BOSSETTA, Michael. The digital architectures of social media: comparing political campaigning on Facebook, Twitter, Instagram, and Snapchat in the 2016 U.S. election. Journalism & Mass Communication Quarterly, v. 95, n. 2, p. 471-496, 2018.

BOULIANNE, Shelley. Social media use and participation: a meta-analysis of current research. Information, Communication & Society, v. 18, n. 5, p. 524-538, 2015.

BOULIANNE, Shelley. Twenty years of digital media effects on civic and political participation. Communication Research, v. 47, n. 7, p. 947-966, 2020.

BROUGH, Melissa M.; SHRESTHOVA, Sangita. Fandom meets activism: rethinking civic and political participation. Transformative Works and Cultures, v. 10, 2012.

CHADWICK, Andrew. The hybrid media system: politics and power. 2. ed. Oxford: Oxford University Press, 2017.

COPPOCK, Alexander. Persuasion in parallel: how information changes minds about politics. Chicago: University of Chicago Press, 2023.

EVANGELISTA, Rafael; BRUNO, Fernanda. WhatsApp and political instability in Brazil: targeted messages and political radicalisation. Internet Policy Review, v. 8, n. 4, 2019.

FREELON, Deen; MARWICK, Alice; KREISS, Daniel. False equivalencies: online activism from left to right. Science, v. 369, n. 6508, p. 1197-1201, 2020.

GONZÁLEZ-BAILÓN, Sandra et al. Asymmetric ideological segregation in exposure to political news on Facebook. Science, v. 381, n. 6656, p. 392-398, 2023.

GUESS, Andrew M. et al. How do social media feed algorithms affect attitudes and behavior in an election campaign? Science, v. 381, n. 6656, p. 398-404, 2023.

HINCK, Ashley. Politics for the love of fandom: fan-based citizenship in a digital world. Baton Rouge: Louisiana State University Press, 2019.

JENKINS, Henry; FORD, Sam; GREEN, Joshua. Spreadable media: creating value and meaning in a networked culture. New York: New York University Press, 2013.

KALLA, Joshua L.; BROOCKMAN, David E. The minimal persuasive effects of campaign contact in general elections: evidence from 49 field experiments. American Political Science Review, v. 112, n. 1, p. 148-166, 2018.

KREISS, Daniel. Prototype politics: technology-intensive campaigning and the data of democracy. New York: Oxford University Press, 2016.

NYHAN, Brendan et al. Like-minded sources on Facebook are prevalent but not polarizing. Nature, v. 620, n. 7972, p. 137-144, 2023.

RESENDE, Gustavo et al. (Mis)information dissemination in WhatsApp: gathering, analyzing and countermeasures. In: THE WORLD WIDE WEB CONFERENCE, 2019, San Francisco. Proceedings [...]. New York: ACM, 2019. p. 818-828.

RIEDL, Martin J.; LUKITO, Josephine; WOOLLEY, Samuel C. Political influencers on social media: an introduction. Social Media + Society, v. 9, n. 2, 2023.

STROMER-GALLEY, Jennifer. Presidential campaigning in the internet age. 2. ed. New York: Oxford University Press, 2019.

TORET, Javier; GUERRERO, Rocío. El Ejército de Lula. Tecnopolítica a la altura de los tiempos. El Salto, Madrid, Oct. 27, 2026. Blog Atenea Cyborg. Available at: https://www.elsaltodiario.com/atenea_cyborg/ejercito-lula-tecnopolitica. Accessed on: Oct 27, 2026.

AGÊNCIA NACIONAL DE TELECOMUNICAÇÕES (Brazil). Painel de dados Meu Município: acessos e cobertura. Brasília: Anatel, 2026. Available at: https://www.anatel.gov.br/dadosabertos/paineis_de_dados/meu_municipio/. Accessed on: Oct 27, 2026.

TRIBUNAL SUPERIOR ELEITORAL (Brazil). Open data: votação por seção eleitoral 2022 e 2026; perfil do eleitorado por seção eleitoral 2026. Brasília: TSE, 2026. Available at: https://dadosabertos.tse.jus.br. Accessed on: Oct 27, 2026.

What the polls and markets missed · The scare that mobilized · Placar home page