ASKED
AI MORE?

Everything published about what people ask comes from query logs, and the numbers in them are stranger than the aggregate figures suggest. In the largest query log whose full frequency distribution has ever been published, the majority of distinct search queries were asked exactly once in the six weeks it covers.

That single fact does more work than any volume estimate. It means the picture of "what people search for" is built from a small, heavily repeated head sitting on a tail that is, by construction, invisible until after the fact. And the problem shifts shape one step along, because prompts typed at an AI assistant are question-shaped almost by construction, while the public record for those describes subject matter rather than repetition.

Key Takeaways

• The public evidence base is query logs, and the one whose full frequency distribution is published is an AltaVista log of roughly a billion requests over six weeks, analysed in 1999.
• Of the 154 million distinct queries in it, 63.7 percent were asked exactly once in that window, and only 13.6 percent more than three times.
• Queries then averaged 2.35 words. For 85 percent of them only the first screen of results was ever viewed, and 77.6 percent of sessions contained a single query with no refinement.
• Google has said publicly that 15 percent of the searches it sees each day are ones it has not seen before.
• A question-shaped query is a minority form in search with a published definition: the Natural Questions corpus keeps only Google queries of eight words or more that survive filters written to "discard a large proportion of queries that are non-questions".
• Million-conversation assistant corpora do exist, but they are donated collections rather than production logs, and none of them publishes the frequency distribution that makes search volume a number at all.

Where the numbers on search queries come from

From logs, and from a much smaller number of them than you would expect.

The reference study is by Craig Silverstein, Monika Henzinger, Hannes Marais and Michael Moricz, in SIGIR Forum in 1999. They analysed "an AltaVista Search Engine query log consisting of approximately 1 billion entries for search requests over a period of six weeks". That covered "almost 285 million user sessions, each an attempt to fill a single information need".

Query logs are commercially sensitive and personally revealing, which is why open releases at that scale are rare. What the public has instead is vendor estimates, panel data and the occasional figure a search company chooses to state.

So when anyone tells you what people ask about, the underlying evidence is one of three things. A proprietary log they cannot show you. An estimate modelled from clickstream panels. Or a statement from a search engine. It is always worth knowing which, and that is the same instinct behind how we count anything on this site.

Why most search queries are asked exactly once

Because language is generative and questions are not drawn from a menu.

The Silverstein figures are the clearest demonstration in print. Across 154 million distinct queries:

How often a distinct query was asked in six weeks	Share of distinct queries
Exactly once	63.7 percent
Twice	16.2 percent
Three times	6.5 percent
More than three times	13.6 percent

The average distinct query in that log was asked 3.97 times. The most common one was asked 1,551,477 times. Both numbers are true at once, which is what a long tail looks like when you write it down.

Google has said the modern version of this publicly. Announcing BERT, a change to how Search reads language, in October 2019, Pandu Nayak wrote that of the billions of searches Google sees, "15 percent of those queries are ones we haven't seen before".

Fifteen percent, every day, brand new. That is the part no keyword tool can have a number for, because the number does not exist until somebody types the thing.

What makes a question-shaped query different

Length, word order, and the fact that word order suddenly matters.

Start with what a query is for. Andrei Broder's 2002 taxonomy, still the standard framing, splits search intent three ways. Navigational: "give me the url of the site I want to reach". Informational: "acquire some information assumed to be present on one or more web pages". Transactional: "perform some web-mediated activity".

The line usually quoted from the paper is that "informational queries constitute less than 50% of web searches". Broder's log analysis put the split at roughly 20 percent navigational, 48 percent informational and 30 percent transactional, with the remaining 2 percent unclassified.

A question-shaped query is almost purely informational, which already makes it a minority form. The sharper definition comes from the people who had to build a corpus of them.

The Natural Questions corpus, published in Transactions of the Association for Computational Linguistics in 2019, is built from "real anonymized, aggregated queries issued to the Google search engine". To find the questions inside that stream, the authors had to filter. Their length rule: "All the questions in NQ are queries of 8 words or more that have been issued to the Google search engine by multiple users in a short period of time." Then five syntactic tests on top. Three of them look for queries beginning with "who", "when" or "where" followed by particular verb forms; the other two key on the shape of the query rather than on an opening question word.

The purpose of all that machinery is stated plainly: "to discard a large proportion of queries that are non-questions". Even inside Google's own query stream, a question you would recognise as a question is something you have to go looking for.

Which sets up the awkward part. A prompt typed at an AI assistant is question-shaped almost by construction: it is a sentence, it carries function words, and it is often well past eight words without anyone filtering for it. That is the rare form in the search record and the ordinary form in an assistant.

Public collections of those prompts do exist, and they are large. WildChat, presented at ICLR in 2024, holds a million user conversations with ChatGPT comprising over 2.5 million turns, gathered by offering people free access in exchange for consent to share the transcripts. LMSYS-Chat-1M, released in 2023, holds another million across twenty-five models, "collected from 210K unique IP addresses in the wild on our Vicuna demo and Chatbot Arena website".

That turn count is worth setting beside the 1999 figures. Two and a half turns per conversation is a conversation; 77.6 percent of search sessions carried one query and no refinement at all. People do not so much search an assistant as talk to it.

There is still no assistant equivalent of the table above. Operators have published accounts of their own production traffic, but those describe what people bring by subject and by task. The corpora describe the shape of the prompts. Neither kind of publication reports repetition, so no public source, donated corpus or operator write-up, says what share of distinct prompts is asked only once. The shape of the question stream is documented and its distribution is not, which is the reverse of the position for web search.

Why people used to avoid asking questions

Because from the 1999 log through to at least 2019, asking properly made your results worse.

The 1999 log records the habit precisely. The average query was 2.35 words long. For 85 percent of queries "only the first result screen is viewed". And 77.6 percent of sessions contained exactly one query, meaning the searcher typed something, looked, and left without rephrasing.

Nayak has a name for the resulting dialect: "people often use 'keyword-ese,' typing strings of words that they think we'll understand, but aren't actually how they'd naturally ask a question". Keyword-ese was rational. An engine treating a query as an unordered bag of words gains nothing from "how do I", "for" or "to", and those words could actively dilute the terms that mattered.

What changed is that word order started carrying meaning. Nayak's worked example is a query about whether a Brazilian traveller needs a visa for the United States: the word "to" fixes the direction of travel, and before the change the results were about Americans going to Brazil. That is the whole difference between a bag of words and a sentence, in one preposition.

If you want to see how badly people estimate the head of that distribution, Higher or Lower is the experiment, and its explainer covers what a search-volume figure is measuring in the first place.

What still cannot be counted

Three things, and they are structural rather than temporary.

1. Anything new. A query nobody has typed has no volume, and 15 percent of a day's searches are in that category.
2. The same question asked differently. "Can I take a dog on a train", "dogs on trains uk" and "are dogs allowed on trains" are one information need and at least three rows in any counting system. A tool that reports them separately under-counts the need; one that merges them has made an editorial decision you cannot inspect.
3. Intent. Broder is candid that even his own log classification was "at best an inexact science, but usually a wild guess", and he built it by having 400 queries read by hand after removing the non-English and sexually oriented ones. Nobody has a clean way to know why a string of words was typed.

None of that makes the numbers useless. It makes them a measurement of the head, reported with a confidence the tail does not support.

Frequently asked questions

How does anyone know what people search for?

From query logs held by search engines, plus estimates modelled on top of them. The one public analysis that publishes a full query-frequency distribution covers an AltaVista log of about a billion requests over six weeks, released by Silverstein and colleagues in 1999. Most of what circulates now rests on a proprietary dataset, a panel-based estimate, or a figure a search company has chosen to state publicly. When you see a search volume, it is worth knowing which of the three you are looking at.

What proportion of searches are brand new?

About 15 percent a day, on Google's own account. Pandu Nayak wrote in October 2019 that of the billions of searches Google sees, "15 percent of those queries are ones we haven't seen before". The 1999 AltaVista log shows the same shape historically: of its 154 million distinct queries, 63.7 percent were asked exactly once.

What is a question-shaped query?

Operationally, a long query with the grammar of a question rather than a list of keywords. The clearest published definition comes from the Natural Questions corpus, which keeps only Google queries of eight words or more, issued by several users in a short period, that also pass syntactic filters written "to discard a large proportion of queries that are non-questions". The filtering is necessary because most queries are not questions.

Why did people search in keywords rather than sentences?

Because it worked better. Early engines treated a query as an unordered set of terms, so function words added nothing and could dilute the words that mattered. Google's Pandu Nayak calls the resulting dialect "keyword-ese". The 1999 log shows how far it went: an average of 2.35 words per query, and no refinement at all in 77.6 percent of sessions.

Are most searches questions?

No, and they never have been. Broder's 2002 taxonomy found that "informational queries constitute less than 50% of web searches", with his log analysis splitting roughly 20 percent navigational, 48 percent informational and 30 percent transactional. Question-shaped queries are a subset of the informational share, which is why a corpus builder has to filter for them rather than simply collect them. What people type at an AI assistant is far more often question-shaped by construction, though nothing in the published record puts a number on the difference.

Guessing at any of this is harder than it looks, which is the point of the quizzes. If you would rather have a puzzle with a single right answer, the codes and ciphers are the other end of the site.

Sources

• Craig Silverstein, Monika Henzinger, Hannes Marais and Michael Moricz, Analysis of a Very Large Web Search Engine Query Log, SIGIR Forum 33(1), 1999, 6-12. Read as the PDF hosted by SIGIR. Supplies the billion-entry AltaVista log and 285 million sessions, the 2.35-word average, the 85 percent first-screen figure, the 77.6 percent single-query sessions, and the distinct-query frequency table - retrieved 10 August 2026.
• Andrei Broder, A Taxonomy of Web Search, SIGIR Forum 36(2), 2002, 3-10. Read as the PDF hosted by SIGIR. Supplies the navigational, informational and transactional definitions, the "less than 50%" statement, the log-analysis split, and the author's own caveat about inferring intent - retrieved 10 August 2026.
• Pandu Nayak, Understanding searches better than ever before, The Keyword, Google, 25 October 2019. Supplies the 15 percent never-seen-before figure, the "keyword-ese" description, and the Brazilian-traveller visa example - retrieved 10 August 2026.
• Tom Kwiatkowski and colleagues, Natural Questions: A Benchmark for Question Answering Research, Transactions of the Association for Computational Linguistics 7, 2019. Read as the PDF in the ACL Anthology. Supplies the description of the queries as real anonymized aggregated Google queries, the eight-word threshold, the five syntactic heuristics and the stated purpose of discarding non-questions - retrieved 10 August 2026.
• Wenting Zhao and colleagues, WildChat: 1M ChatGPT Interaction Logs in the Wild, ICLR 2024. Supplies the million user-ChatGPT conversations and the collection method, free access offered in exchange for consent to share transcripts - retrieved 10 August 2026.
• Lianmin Zheng and colleagues, LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset, 2023. Supplies the million conversations across twenty-five models and the verbatim "collected from 210K unique IP addresses in the wild on our Vicuna demo and Chatbot Arena website" - retrieved 10 August 2026.