The gatekeepers you don't see: search ranking and autocomplete

Series: How minds are moved · The modern machinery · 7 of 7

On 11 November 2026 hundreds of schools in Italy were evacuated after emailed bomb threats, and similar messages reached schools in Ireland and the UK. My news engine, LSN News, collected 2,776 articles about it from 1,250 outlets in 48 languages. Among the articles my engine collected in those two days, it was the number one story in Italian. In English it ranked 8th. In Romanian 13th, in Polish 25th, in Russian 44th, in Hindi 70th. Same event, same facts, very different importance depending on which language you read the news in. Ranking is a quiet kind of power: it does not tell you what to think, it decides what you see first. My AI agent and I went through what is known about the biggest rankers of all, search engines, and what is only claimed.

Contents · 12 sections

TL;DR #

• Search results and autocomplete are ranked and filtered by rules most people never read. The rules are partly public, and they include deliberate removals.
• A 2015 study by Robert Epstein claimed biased rankings could shift undecided voters’ preferences by 20% or more. The lab effect is real; the later claims that Google moved millions of real votes are contested and were criticised by other researchers.
• A 2023 study of real Google users found that what people clicked was more partisan than what Google showed them. Choice mattered more than the algorithm.
• Documented ranking abuse does exist: the EU’s top court confirmed in 2024 that Google demoted rival shopping services in its results.
• AI summaries now sit above the links. In one 2025 study, people clicked a result on 8% of visits with an AI summary, against 15% without.

How ranking shapes what you see #

Nobody reads page two. Studies of search behaviour have found for years that the first few results get most clicks. Whoever controls the order controls attention, and the order is set by a mix of relevance signals, quality rules, commercial interests and safety policies. That is not a conspiracy; it is how any list works. The question is which rules are applied, and whether we can see them.

The search engine manipulation effect: claim and critique #

In 2015 the psychologist Robert Epstein and Ronald Robertson published a study in PNAS. In five experiments with more than 4,500 undecided voters in the US and India, people searched a mock search engine whose results were biased toward one candidate. The authors reported that biased rankings shifted voting preferences by 20% or more, much more in some groups, and that most people did not notice the bias. They called it the search engine manipulation effect (SEME).

The critiques matter as much as the finding:

• The experiments used a simulated engine, invented rankings and people with little prior knowledge, often about an unfamiliar foreign election. Real voters have many other sources.
• Katharina Zweig of AlgorithmWatch argued that in the most realistic of the studies the single effects were around 2 to 4%, and that the headline measure added the shifts toward both candidates together.
• Google’s then head of search called the idea that it would secretly influence elections a “flawed elections conspiracy theory”. That is a company defending itself, so weigh it as such.

Then the claim grew. In 2019 Epstein told a US Senate subcommittee that biased Google results may have shifted at least 2.6 million votes toward Hillary Clinton in 2016, and possibly up to 10.4 million. He has described himself as a Clinton supporter. President Trump cited it, inflating the upper figure to 16 million, which Epstein said was not his number. Reporters found that the 2.6 million estimate rested on a small study of 95 people and had not been peer-reviewed. So: the lab effect is published and plausible; the real-world vote counts are contested estimates, not findings.

What real users do #

The best counterweight comes from a 2023 study in Nature by Ronald E. Robertson and colleagues, which tracked what Google actually showed real users during the 2018 and 2020 US elections, and what they clicked. People’s political identity made little difference to how much partisan or unreliable news Google showed them. It made a big difference to what they chose to click. The authors concluded that exposure and engagement were driven mainly by users’ own choices, not by algorithmic curation.

This does not prove search is neutral. It shows that the strongest filter is often the person at the keyboard.

Autocomplete: the suggestions you did not ask for #

Autocomplete is ranking before you even finish typing. Google publishes its autocomplete policies. Its systems try to block predictions that are violent, sexually explicit, hateful, disparaging or dangerous, and those that are unlikely to lead to reliable content, such as unconfirmed rumours after a news event. For elections, it says it does not allow predictions that can be read as a position for or against a political figure or party, or as a claim about the integrity of the electoral process.

The rules have side effects. In July 2024, after the assassination attempt on Donald Trump, users noticed that autocomplete did not suggest it. Critics, including members of Congress, called it election interference. Google said its protections against predictions about political violence were out of date and had blocked those predictions, that no manual action had been taken, and that it fixed them. Both facts are on record: a filter designed to avoid harm left a major real event out of the suggestions for a period after it happened, and the company says it was a policy gap, not a choice about Trump.

Documented ranking abuse #

There is a case where the courts did find manipulation. In 2017 the European Commission fined Google €2.42 billion for favouring its own comparison-shopping service in search results and demoting rivals. On 10 September 2024 the Court of Justice of the EU upheld the fine in a final ruling. This was commercial, not political, but it is the clearest legal proof that ranking can be bent to serve the ranker.

The new gatekeeper above the links #

AI summaries now often answer before any link appears. A 2025 Pew Research Center study tracked about 900 US adults’ Google searches. When an AI summary appeared, they clicked a regular result on 8% of visits, against 15% when there was none, and only about 1% clicked a link inside the summary itself. Whoever writes the summary now shapes the answer more than whoever ranks the links.

I see the shift on my own sites. In the last seven days my web-threat sensors logged 3,485 requests from Perplexity’s crawlers, 1,565 from Anthropic’s and 572 from OpenAI’s, against 153 from Google. AI answer engines read my pages more often than the classic search engine did that week. That is one small site and one week, not a trend line, but it shows where the reading is happening.

Same story, different rank: my own engine #

My news engine is a ranker too, and I owe you its limits. The ranks in the first paragraph are counted per language, over the 48 hours after the first article, by how many articles each story had: the bomb-threat story was 1st of 23,194 stories in Italian, 8th of 133,292 in English, 13th in Romanian, 44th in Russian. Language is not the same as country, and my engine holds more outlets in some languages than in others. The story page, while it stays public: https://lsn-news.shuffle-on.com/s/955147. But the pattern is the point: importance is local.

Framing is local too. For the shooting in Kraków on 7 November (2,168 outlets; https://lsn-news.shuffle-on.com/s/228066 while public), of 2,484 English headlines, 490 mentioned the suspect’s far-right views and 4 mentioned Russia or Russians. Of 600 Russian-language headlines, none mentioned far-right views and 74 mentioned Russia or Russians, focusing on the angle of the one Russian citizen among the wounded and on whether other Russians were hurt. Same event, 60 languages. Each audience got a different first line, and the first line is what most people remember.

What it feels like from the inside #

Search does not feel like being steered. It feels like looking something up. The top result feels like “the answer”, the autocomplete feels like “what people ask”, and an AI summary feels like the end of the question. That is exactly why ranking is powerful: it borrows the trust you give to your own curiosity.

How to notice it #

• The first result is an ad, a shopping box or a product from the search company itself.
• The autocomplete is missing something you know happened.
• An AI summary answers confidently and cites sources you have never heard of, or none.
• Your results on a contested topic all come from one direction, and you have never clicked the other.
• A story seems huge or tiny, and you have only checked news in one language.

What helps: scroll past the first screen; try a second search engine; search the same topic in another language; open the primary source behind an AI summary; and notice which results you skip, because that is your own filter.

Frequently asked questions #

Can search engines swing elections? #

Lab experiments show biased rankings can shift undecided people who know little about the candidates. No court or independent study has shown that a search engine actually swung a real election. The multi-million-vote estimates are contested.

Does Google censor autocomplete? #

It removes predictions under published policies, including any that take a side on a candidate or question the integrity of elections. Those rules sometimes hide real events, as in July 2024. Whether that is responsible filtering or censorship depends on whom you ask; the policies themselves are public.

Are filter bubbles in search real? #

The best real-world study found that Google showed people of different politics fairly similar news, and that users’ own clicks were more partisan than their results.

Is any ranking manipulation proven? #

Yes, commercially: the EU’s top court confirmed that Google demoted rival shopping services. Political manipulation by a major search engine has not been proven in court.

The lesson #

Whoever ranks the list does not need to lie to you. They only need to decide what comes first. You cannot read every list, but you can read past the first line: one more result, one more engine, one more language. That habit breaks more gatekeepers than any law.