The search that knew too much

Two people searched for the same four words on the same morning and received different first results. Neither had done anything unusual.

Personalisation of this kind is usually described as a privacy question, which it is, but it is also an evidence question. If a search engine ranks results partly by what you have clicked before, then the page you see has been selected in part by your own previous agreement. For a factual query — the boiling point of a liquid, the date of an event — this barely matters. For a contested one, it matters a great deal, because the strongest version of the case you already favour will reliably appear first and the strongest version of the other case may never appear at all.

The effect is easy to underestimate because it is invisible. No result is hidden with a label saying so; the page simply contains ten links, as it always does, and nothing on it indicates which of the ten were chosen for you rather than for everybody.

There are cheap corrections. Searching the same phrase in a private window gives a result less shaped by history. Searching the opposing phrasing — 'evidence against' as well as 'evidence for' — reaches a different part of the index. Both take seconds and neither requires trusting any particular company's settings page.

The broader point is about what a search result is. It is not a list of what exists. It is a ranked prediction of what will satisfy you, produced by a system optimised for your approval, and approval is not the same as accuracy. On anything you expect to argue about, the ranking is the first thing to distrust.