← Blog

Everywhere and nowhere: what is documented about visibility before Brazil's run-off

15 October 2026

On 14 October a report by NOS, the Dutch public broadcaster, said that, on social media, one of the two candidates in Brazil’s presidential run-off seemed to be everywhere and the other nowhere. It named no study and no data set. Its evidence was one campaign’s account, a single search its reporter ran on Instagram, an order from Brazil’s Superior Electoral Court (TSE), and Meta’s first reply.

The run-off is on 25 October. Here is what was measured, what Meta admitted, what the court and the federal prosecutors did, and what is not documented, so that you can weigh the next such claim yourself. Like our earlier post on influence online, this one names no candidate or party; some source titles do. “First-placed” and “second-placed” refer to the first round, on 4 October. Quotations from Portuguese are our translation.

What was measured

Engagement, counted by commercial firms. These firms sell social-media analytics and publish no full method: industry figures. HypeAuditor found that from 24 August to 23 September the first-placed candidate’s Instagram posts drew about 150,000 likes each, against about 50,000 for the second-placed candidate’s. DSC Lab, counting likes plus comments on Instagram from 14 September to 2 October, found the same gap per post, 329,367 against 137,675, but counted the second-placed candidate ahead on followers (15.4 million against 11.9 million), total engagement (28.9 million against 16.8 million) and views (349.8 million against 155.8 million), from 210 posts against 51. Few posts that each draw a crowd, or many that each draw less: both can be true at once, and neither is “nowhere”.

Tests of what the platforms show. TecMundo reported on 8 October that it had searched Instagram on several phones, each signed in to a different account. Neither candidate’s official account appeared among the first results for his name; typing the full username brought each to the top. Meta told TecMundo: “we do not recommend official candidate accounts registered with the TSE on the surfaces of our apps, including search.” That follows a TSE rule in force since 2024: platforms that use a recommender system “must exclude from the results” the profiles candidates have declared to the court, and what is posted on them, except paid boosting. An unofficial account carrying a candidate’s name can still be recommended. So one search, on one account, shows little.

The rule has not always held: a Metrópoles columnist found candidates’ accounts the column did not follow in 3, then 13, of the first 30 posts of the column’s own Instagram feed in August, and the NetLab UFRJ research laboratory found 24 among 1,873 TikTok recommendations in September. Too few to say whether either side gained.

An AI tool that answered one name and not the other. On 14 October Terra and Olhar Digital reported that Meta AI on WhatsApp produced a campaign image for the first-placed candidate’s first name and ballot number, and refused the same request for the second-placed candidate. Meta, citing the electoral rules, then extended its block on images of candidates to first names and ballot numbers. The tests show the asymmetry; they do not show its cause.

What Meta admitted, and what it denies

On 10 October the second-placed candidate’s campaign went to the TSE, alleging that Meta had restricted its reach to its opponent’s benefit: a contested claim, which Meta denies. Meta’s first answer: “The allegations are simply false and unsupported by the facts.”

Meta’s filing of 12 October is under seal; Metrópoles reported the Reels incident and Meta’s denial, and the complaining campaign’s party summarised the rest. Meta admits two failures. On 23 September it disabled the campaign’s ad account because two earlier restrictions “were wrongly classified as serious”, and restored it “in a little over an hour”. On 7 October Reels view counters were blocked from 15:50 to 20:49, Brasília time, for creators “in more than a hundred countries”. Meta says distribution was not affected, that the official account’s distribution that day was “substantially above its own recent baseline”, and that there is “no technically verifiable evidence of selectivity or manipulation”. Of the network of 255 profiles the complaint cites, from a survey by the Instituto Democracia em Xeque, Meta says it reviewed the 35 it could identify, disabled 7, and found no coordinated inauthentic behaviour among them. That covers 35 profiles, not 255. The data behind all of it is not public.

What the court and the prosecutors did

On 10 October a TSE minister gave Meta 48 hours to preserve its logs, report profile by profile every measure applied since 16 August, explain 23 September and 7 October, and say whether it applies the recommendation rule “identically for both candidacies”. “The facts alleged,” the decision reads, “occur in an environment that only the respondent controls.” He had not, at this stage, recognised any algorithmic discrimination, Congresso em Foco reported. Meta asked for 48 more hours. As of 15 October we found no ruling. Update, 19 October 2026: on 15 October the minister granted an injunction ordering equal treatment of both run-off candidates under a daily fine, while ruling out “for now” manipulation for or against either candidacy; the plenary confirmed it unanimously on 16 October.

The same week, the court suspended a site that offered cryptocurrency for posts supporting the first-placed candidate, finding “indications that the project went beyond spontaneous expression of political support” (9 October), and gave Google 24 hours to explain AI answers attributing to the second-placed candidate “false and non-existent statements” (13 October).

On 14 October the federal prosecutors’ office for citizens’ rights (MPF/PFDC) opened a procedure on Meta and X, after four civil-society organisations brought it “more than a thousand reports” of accounts and posts taken down. The platforms have five days to say whether these follow court orders or algorithms. Self-gathered reports do not measure how often each side is restricted.

What is not documented

No independent study with a published method shows that a platform held back one candidate or pushed the other in this campaign. Débora Salles of NetLab UFRJ told DW: “It is very hard to say there is interference in the campaign […] only on the basis of the data we can access today. It is not that such interference cannot exist, but platforms today work as a black box”. So deliberate favouritism remains contested: from outside, it can be neither shown nor ruled out, and nor can the denial be checked.

How one side can look everywhere without anyone choosing it

Each is documented; none has been measured for this election.

• The rule moves reach elsewhere. With official accounts kept out of recommendations, candidates reach people through followers, shares, paid ads, and allied or unofficial accounts. The side with the denser network of allies can look more present.
• Ranking rewards reactions. In US data, posts about the political out-group “were shared or retweeted about twice as often as posts about the in-group”. Twitter’s own researchers found that “in six out of seven countries studied, the mainstream political right enjoys higher algorithmic amplification than the mainstream political left”; Brazil was not among them.
• Meta changed what it recommends. In January 2025 it said it would “recommend more political content based on these personalized signals”, such as likes and views, without saying when in Brazil.
• Some amplification is paid or rewarded, like the cryptocurrency site. How much, nobody has measured.

So ask: what was measured, by whom, with what method? Per post or in total? One search, or many? Is the source a party to the dispute? Both parties have given the court their account; the court has asked for the data.

How a ranked feed decides what you see is on How It Works; what is known from other elections, on Elections & Democratic Harm. We will re-read these sources after the run-off and correct this post if any of them changes.

← All posts
About this project →