What autocomplete actually is
Autocomplete predicts the rest of the query you have started typing. That sentence contains the two things people most often get wrong about it.
First, it predicts a query, not an answer. The dropdown is guessing what you were about to type, not telling you what is true, popular, or worth reading. A prediction appearing under your cursor is a statement about how sentences that begin like yours have historically continued — nothing more.
Second, it is a prediction rather than a suggestion, and engines are careful about the distinction for a reason that turned out to be legal rather than semantic. A suggestion implies endorsement. A prediction implies inference. When autocomplete attaches a damaging word to a person's name, the difference between those two framings is the difference between publishing a defamatory statement and reporting a statistical pattern — and courts in several jurisdictions have been asked to decide which it is.
The feature is old and the name has moved around: Google Suggest as a Labs experiment, then autocomplete proper, then a brief and much-disliked detour called Google Instant. Underneath, the mechanism has changed remarkably little.
The mechanism: prefix matching under a millisecond budget
Autocomplete is a hard engineering problem for one reason: the latency budget. The dropdown must update between one keystroke and the next, which means the whole operation — network round trip included — has to complete in a few milliseconds. That constraint dictates the design.
The core structure is a prefix tree, or trie: a tree in which a node's position rather than its contents determines the string it represents, with each edge labelled by a single character. Walking down from the root spelling out "sea" lands you at the node beneath which every candidate beginning "sea" is stored. Tries are, as the standard description has it, "commonly used in predictive text or autocomplete dictionaries, and approximate matching algorithms" — because retrieving all completions of a prefix is a walk rather than a search.
What is stored in the tree is the interesting part. Candidates come from aggregated past queries: what large numbers of people have actually typed, deduplicated, counted, and pruned. This is why autocomplete quality is a function of query volume, and why it is one of the hardest advantages in search to replicate. An index can be crawled by anyone with a few servers; a corpus of what a hundred million people typed last Tuesday cannot be crawled at all. It accrues only to engines that already have the traffic.
Candidates are then ranked, typically by some combination of historical frequency, recency and trend, and the searcher's language and approximate location. Modern implementations blend in freshness signals so that a name entering the news rises within hours rather than weeks.
Two refinements matter. Predictions must tolerate typing errors, since users are mid-word and often mid-mistake; the standard techniques are phonetic matching of the Soundex family and edit-distance measures such as the Levenshtein algorithm, which is language-independent. And many engines show zero-prefix predictions — a list offered before you have typed anything at all, drawn from your recent searches or from trending queries. That list is not autocomplete in the strict sense, because there is no prefix to complete.
A short history, including the version everybody hated
Autocomplete arrived on Google as a Labs experiment and became a default feature of the search box in 2008, initially listing predictions beneath the box along with an approximate count of results for each.
In September 2010 Google went considerably further with Google Instant, which did not merely predict the query but re-ran the search on every keystroke, rewriting the entire results page as you typed. Marissa Mayer said it would save two to five seconds per search. It was disorientating to use, punishing on servers, and incompatible with the way people actually compose queries — many searchers type a phrase, pause, and reconsider, and Instant made the page thrash while they did it. Google removed Instant results on 26 July 2017, citing the growing share of searches on mobile devices, where a full-page rewrite per keystroke is neither useful nor cheap.
Ordinary autocomplete survived that episode unchanged and is now near-universal across search engines, browsers' address bars, e-commerce sites and operating system search fields. Its persistence is a reasonable indicator that predicting the query is genuinely useful even where predicting the answer is not.
Why your predictions are not the same as anyone else's
Autocomplete is conditioned on more than the characters you typed. The reliable inputs are language, approximate location, recency and trending signals, device, and — on engines that personalise — your own search history. Type the same three letters in Manchester and in Melbourne, in English and in German, on a phone and on a laptop, and you will see four different lists.
This has a consequence that is routinely ignored in public argument. A screenshot of autocomplete is not evidence of what "the internet thinks". It is evidence of what one engine predicted for one person, in one place, at one moment, given a partial string. Autocomplete screenshots are nonetheless used as if they were survey data — in news reports, in marketing decks, and in litigation — and they cannot bear that weight.
There is a second, quieter consequence. Because predictions are fetched from the engine as you type, the engine sees what you typed even if you never press enter. A query abandoned halfway has still left the device. Engines that do not retain query logs have less to work with when building predictions, which is a genuine trade-off rather than a flaw: an engine that keeps nothing about you cannot personalise the dropdown, and an engine that personalises the dropdown is keeping something.
The structural point behind all of this is that autocomplete is owned by whoever has the query stream. Google, with by far the largest, has the best predictions almost by definition. Microsoft Bing runs its own and supplies partners. Engines that resell another engine's results — Startpage, which serves Google's rankings without Google's tracking, and DuckDuckGo, which takes its web links from Bing — face the question of where predictions come from separately from where results come from, and the honest answer is that the arrangements are rarely documented in public. Mojeek, which crawls independently and does not build user profiles, has a far smaller query stream to learn from, and that is an unavoidable consequence of the privacy position rather than a deficiency of engineering.
What gets filtered, and why it is contested
No engine shows raw predictions. Every one of them filters, and the filtering is where autocomplete stops being an engineering topic.
The public record of Google's filtering is patchy and mostly assembled by outsiders noticing absences. In January 2010, predictions for searches beginning "Islam is" stopped appearing, which a Google spokesperson attributed to a bug that was subsequently fixed; the predictions returned later that month. From 26 January 2011, terms including "BitTorrent", "Torrent", "uTorrent", "Megaupload" and "Rapidshare" were filtered from Autocomplete. In September 2012, Google removed "bisexual" from terms blacklisted for Instant Search — an example of a filter that was itself the problem. Google's position throughout has been that it filters words it considers potentially inappropriate, and 2600: The Hacker Quarterly compiled a list of restricted terms, most vulgar or derogatory but including some whose exclusion had no obvious rationale.
The sharpest pressure has come from defamation claims. The best-documented is Bettina Wulff, wife of the then German president, for whom Google's autocomplete produced terms including "escort" and "prostitute" — derived from unsubstantiated rumours circulating on gossip sites and blogs. In 2012 she took legal action against Google, the broadcaster Günther Jauch and 34 other German and foreign bloggers and outlets, obtaining cease-and-desist orders. Her argument was precisely the one at the top of this page: that users would read the predictions as statements of fact rather than as inferences about typing behaviour.
The mirror image of filtering is manipulation. Because predictions derive from query volume, query volume is a surface that can be pushed on, and a reputation-management industry exists around companies linked to prediction terms such as scam, complaints and fraud. It is, as the standard reference dryly notes, "an area that is open to manipulation". Engines respond with rate limiting, bot detection and manual removal, and the contest does not end.
What autocomplete is not
Four features sit close to autocomplete and are constantly confused with it.
Spelling correction — the "showing results for" or "did you mean" line — runs after you submit, on the whole query, and can silently change what was searched. Autocomplete runs before submission and changes nothing unless you pick a prediction.
Related searches, at the foot of the results page, are computed from the completed query rather than from a prefix, and are about the query space around what you asked.
People Also Ask is a results-page feature that extracts answers from indexed documents. It shares autocomplete's conversational look and none of its mechanics.
Browser address-bar suggestions mix at least three sources: your own history and bookmarks, held locally; your default engine's predictions, fetched over the network; and the browser's own heuristics. This is why the address bar sometimes proposes a page you visited once and never searched for. It is also why changing your default search engine changes your address-bar predictions — the network half of that list comes from whichever engine is set.
The single useful habit is to treat the dropdown as what it is: a keyboard shortcut. It saves typing. It is not a ranking, not an endorsement, not a measure of public opinion, and not an answer.
Frequently asked questions
How does search autocomplete work?
It matches what you have typed so far against a stored set of candidate queries, usually held in a prefix tree that makes retrieving all completions of a prefix a simple walk. Candidates come from aggregated past searches, ranked by frequency, recency, trend, language and approximate location. The whole operation must complete between keystrokes, in a few milliseconds.
Are autocomplete predictions the same for everyone?
No. Predictions vary with language, approximate location, device, trending signals, and — on engines that personalise — your own search history. The same three characters typed in two countries produce different lists. This is why an autocomplete screenshot is evidence of what one person saw once, not of what an engine shows or of what the public believes.
What is the difference between autocomplete and suggestions?
Engines describe the dropdown as predictions rather than suggestions, and the distinction is legal as much as linguistic. A suggestion implies endorsement; a prediction implies inference from how similar queries have historically continued. The framing matters most when a prediction attaches a damaging term to a person's name, which is exactly what defamation claims over autocomplete have turned on.
Does a search engine see what I type if I never press enter?
If autocomplete is active, yes — predictions are fetched from the engine as you type, so partial queries leave your device even when abandoned. Engines that do not retain query logs have correspondingly less material for prediction. That is a genuine trade-off: an engine that keeps nothing about you cannot personalise the dropdown.
What was Google Instant and why was it removed?
Announced in September 2010, Google Instant re-ran the search on every keystroke and rewrote the whole results page as you typed. Marissa Mayer said it would save two to five seconds per search. Google removed Instant results on 26 July 2017, citing the growing share of searches on mobile devices, where rewriting a full page per keystroke is neither useful nor cheap.
Do search engines filter autocomplete predictions?
All of them do, and disclosure is thin. Documented cases include predictions for "Islam is" disappearing in January 2010, attributed by Google to a bug and later restored; filtering of BitTorrent and file-sharing terms from 26 January 2011; and the removal of "bisexual" from blacklisted terms in September 2012. Google's stated basis is filtering potentially inappropriate words.
Can autocomplete be defamatory?
It has repeatedly been claimed to be. The best-documented case is Bettina Wulff, wife of the then German president, whose name produced predictions including "escort" and "prostitute" derived from unsubstantiated rumours. In 2012 she took action against Google, broadcaster Günther Jauch and 34 other bloggers and outlets, arguing users would read predictions as statements of fact.
Why do smaller search engines have worse autocomplete?
Because the raw material is aggregated query volume, and that cannot be crawled or bought the way an index can. It accrues only to engines that already have traffic. An independent engine such as Mojeek, which crawls its own index and does not build user profiles, has a far smaller query stream to learn from — a consequence of scale and of the privacy position, not of engineering.