← The Attention Machine

Search

The search engine is not a library. It is an advertising platform with a library function.

How to read this page. Tap any underlined word to see the precise academic term and a short definition. Expand any "Deeper" box for the evidence and contested points. The main text works on its own — you can skip both and still get the whole argument.

The scale

Google was incorporated in September 1998, its first office a garage in Menlo Park, California. By Google's own figure, it now handles more than 5 trillion searches a year — roughly 14 billion a day. Google's advertising revenue in 2024 was about $265 billion, three-quarters of it from search. This is not a side feature of search. It is the reason search exists as a free product.

What shapes the results you see

When you search for a product, a health condition, a political question, or a news event, you are not receiving a neutral ranking of relevant information. The results are shaped by three forces working simultaneously.

Commercial advertising pays to appear at the top of the searches advertisers care about — anything with a product, a service or a price behind it. Across four million results analysed by one large industry study, the first organic (non-paid) link received about 27.6% of clicks. Users who do not read past the paid results at the top of the page are seeing what advertisers paid to put there.

Clicks are factored into Google's ranking: the 2024 US antitrust ruling against Google describes a ranking signal, Navboost, trained on 13 months of users' click data. Content that generates clicks is rewarded — not just content that is accurate or useful. The same engagement dynamic that distorts social media feeds applies to search, in a different form.

Search engine optimisation (SEO) is an entire industry dedicated to making content rank highly — rank, not quality, being what it is paid for. Well-resourced actors — commercial companies, political organisations, coordinated influence campaigns — can and do shape what appears at the top of results for specific queries.

The autocomplete effect

Before you finish typing, Google suggests completions for your query. These suggestions are generated from aggregate search behaviour. An industry study of real search sessions found users pick an autocomplete suggestion about 23% of the time.

Autocomplete is an editorial decision made automatically, at scale, billions of times a day. What autocomplete surfaces, and what it omits, shapes the questions people ask before they have finished forming them. In 2016 one newspaper found it suggesting antisemitic and extremist completions to sensitive questions, and argued that a network of right-wing sites had learned to game the ranking behind them (Cadwalladr, 2016). That was one investigation, years ago; the mechanism it exposed — aggregate behaviour shaping the question before it is asked — is the same today.

The position effect

An eye-tracking study at Cornell found that people trust a highly placed result because of where it sits, not because of any independent assessment of its accuracy: they clicked the top results even when the summaries below them were plainly more relevant, because they assumed Google had ranked by relevance.

Research on health information is striking: a 2014 study that manipulated Google's output found that the ranking of search results — whether pro- or anti-vaccination sites appeared first — significantly influenced users' subsequent beliefs, with groups offered only anti-vaccination results showing increased concern regardless of scientific consensus. Position moves belief — in that study, measurably. Whoever controls position has a hand in what people come to think about a topic.

The zero-click problem

AI Overviews appeared on about 6% of Google searches in January 2025 and on roughly a quarter of them by mid-2025, before settling at about 16% in November 2025, according to Semrush's tracking of more than 10 million queries. They provide answers directly on the search page — drawn from sources across the web, synthesised by AI, without the user visiting any original source. This removes source context, correction policies, funding disclosures, and the ability to assess credibility. The intermediary becomes the effective author of the answer.

The Tow Center for Digital Journalism tested eight AI search tools in March 2025 and found that they answered more than 60% of queries about news articles incorrectly — inventing links, misattributing quotations, and stating wrong answers with unwarranted confidence. Google itself acknowledged in May 2024, soon after bringing AI Overviews to everyone in the US, that some had misread web pages and presented inaccurate information, and that others had drawn on sarcastic or troll-y forum posts — among them advice to use glue to make cheese stick to pizza. When you receive a synthesised answer rather than a list of sources, you lose the ability to evaluate it.

What you can do

Understand that the top positions in a commercial search usually include paid advertising. Use Google's "Web" filter (under "More") to see results without AI Overviews. For health information, start with Cochrane Reviews (cochrane.org) rather than general search. Alternative search engines: DuckDuckGo (no behavioural tracking), Kagi (user-supported, no advertising). Verify AI-generated summaries against original sources before treating them as settled. The specific engines, blockers and privacy guides are kept current on Tools & Sites, which is where these recommendations live.

How we know — the search-ranking experiment on vaccination beliefs +

The striking health-information finding is from Allam, Schulz and Nakamoto (2014), published in the Journal of Medical Internet Research. It was a controlled experiment: participants were shown search results for vaccination that the researchers had deliberately re-ordered — some saw mostly pro-vaccination sites ranked first, some mostly anti-vaccination. Afterwards, beliefs and attitudes shifted in the direction of whatever had been ranked highest, independent of the underlying scientific consensus. Because the ranking was the only thing manipulated, the study isolates the effect of position on belief.

What it does and doesn't show. It demonstrates that ranking order can move stated attitudes in a study setting. It does not measure how long the shift lasts, or how it behaves outside the lab against people's prior commitments and other information sources. Treat it as solid evidence that position matters, not as a measure of exactly how much it matters in the wild.

Tools & Sites — the search engines and blockers named here →
AI & The Next Layer →
Symptoms: Health Misinformation →