How Google autocomplete works (and what it is not)

Start typing into a search box and a list drops down before you finish: the engine’s guess at what you are about to ask. GuessTheSearch is built entirely on that list. Every day we show you the start of a question and ask you to guess the top prediction. To play well — and to read the answers sensibly — it helps to know where those predictions come from.

This article sticks to what Google has said publicly about autocomplete, plus what we have observed from capturing thousands of predictions ourselves. We are an independent game and are not affiliated with Google, so we will not pretend to know anything about its systems beyond that.

Predictions, not suggestions

Google describes autocomplete results as predictions rather than suggestions, and the distinction matters. The feature is not recommending what you should search for. It is trying to predict what you are most likely typing, so you can finish your search faster — which is especially useful on a phone keyboard, or when you are not sure how to spell something.

According to Google’s own explanations, those predictions are drawn from searches that have actually been made on Google. The system looks at what people have typed that begins the same way you have begun, and offers the most likely continuations.

What goes into a prediction

Google has named several broad factors that shape which predictions you see and in what order:

• Real queries. Predictions reflect searches people have actually typed, not phrases written by an editor.
• Popularity. Common searches are more likely to appear than rare ones.
• Freshness. Searches that are trending — rising in interest right now — can appear even if they would not top an all-time list.
• Language and location. The same letters typed in a different language or country can produce very different predictions.
• Your own history. If you are signed in and have search history turned on, some predictions can reflect your past searches.

Popularity is the easiest to see in action. Here are two differently phrased questions — one about a country, one about its people — that we captured on the same day:

The phrasing is different but the top two predictions are identical, because the curiosity behind them is the same. The same prediction tops our captured lists for Germany, Italy and the UK: a great many English-speaking searchers want to know why so many European homes go without air conditioning. A popular question tends to surface whichever way you ask it.

What predictions are not

Because predictions look like sentences, it is easy to read too much into them.

• They are not answers. “Why is X so Y?” being predicted says only that people ask it. It does not mean X is Y, and the question may rest on a false premise.
• They are not Google’s opinion. They reflect what searchers type, filtered by Google’s policies. They are not an editorial view about any person, place or group.
• They are not a poll. Someone who types a question may be curious, sceptical, joking or checking a claim they heard. Search volume measures interest, not belief.
• They are not complete. Google removes some predictions under published policies, and in some contexts shows none at all. We look at that in detail in autocomplete and bias.

Why the same prefix gives different results in different places

Location and language make a bigger difference than most people expect. When we first ran our capture from a server in Zagreb, the prompt “japan vs” came back full of Croatian football fixtures. That is perfectly sensible for a Croatian searcher and useless for a quiz.

So GuessTheSearch pins every request to US English, using the standard language and country parameters (hl=en and gl=us). That has two effects. First, everyone playing the same puzzle is guessing against the same data, wherever they happen to be. Second, the daily capture does not drift depending on where our server runs.

Why some prefixes return nothing

One of the first things we learned building the game is that autocomplete is fussier about wording than you might think. Plenty of perfectly natural questions return no predictions at all. In our testing, “why are nurses so” came back empty, while “why is nursing so” returned a full list. “Why are lawyers always” returned nothing, but “why don’t lawyers” returned plenty:

We can only describe the pattern, not explain it from the inside. Some empty results are probably just a lack of popular searches that begin that way. Others may reflect Google’s policies, which allow it to show fewer or no predictions for certain kinds of prompts. Either way, it has shaped the game: the pairing of question shapes and themes in GuessTheSearch is measured, not chosen. We probed each shape against live autocomplete and kept only the combinations where Google actually answers. That is why jobs never get a “Why are … always ___?” question, and why some subjects are phrased the way they are.

Grammar matters too. Each subject carries its own verb, so we ask “why are printers so” rather than “why is printer so”. The mismatched version returns a much thinner result set, because far fewer people type it.

What we get back, and what we keep

The raw response is messier than the tidy dropdown you see in a browser. A request for one prefix can include near-misses that only look like matches, the bare prefix with nothing after it, and pointers to forums and video sites. Before anything reaches a puzzle, our code:

1. checks that the prediction starts with the exact question shape we asked for;
2. checks that the subject sits directly before the completion;
3. drops answers too long to fit on a button;
4. removes platform noise, name collisions and anything our content filters reject;
5. keeps the list visually distinct, so two options never differ by a single word.

The order of what survives is Google’s order. The #1 prediction that remains is the correct answer; the others become decoys on the harder difficulty levels. Our editorial standards describe the content filters, and every recapture is reviewed by a person before it ships.

Predictions change over time

Because freshness is one of the factors, autocomplete is a moving target. Some predictions are evergreen — “why do Greeks” is topped by breaking plates, a question with no news hook at all. Others are visibly tied to the moment they were captured:

Three of those four are about memory prices, which says a lot about when they were captured. (The odd one out is about a character from an animated series; name collisions like that are a constant feature of autocomplete, and part of the charm.) Our captured technology questions are full of the same kind of time stamps: whether a particular operating-system version is out yet, whether an older one is still supported, whether programmers are being replaced by AI. Ask the same questions a year from now and many of them will have moved on.

How that becomes four different quizzes

That drift is exactly why GuessTheSearch offers four time periods. We capture predictions every day and keep the snapshots, so each period reads from a different slice of history:

The weekly consensus uses a simple ranked-vote method: each day’s list votes for its completions by position, and the totals decide the week’s order. Where our archive does not reach back far enough yet, a period falls back to our offline answer set, rotated so that the four periods are still genuinely different puzzles. As the archive grows, those older periods fill in with real snapshots automatically.

Playing the same theme across periods is one of the more interesting things you can do in the game. Shifts in the top answer are small, real records of what people were curious about at the time.

What this means when you play

• Think like a searcher. The top prediction is what many people type, not the cleverest or truest answer. Our strategy guide goes into this in depth.
• Think American English. The data is pinned to US English, so US spellings, sports and news cycles show up.
• Think about the moment. In the “Last week” period, recent events can push a fresh question to the top.
• Read the answers as questions. A prediction about a nationality or a star sign is a record of curiosity, not a verdict. Our posts on star signs, generations and European stereotypes show what that looks like in practice, and where stereotypes come from covers the psychology behind it.

The full rules, including how difficulty levels choose decoys, are on how to play.

▶ Play today’s tech quiz