Executive Summary

Key figures

Source: Zhou, McCloskey, Srinivasan, "You Can Tell Who's Asking", arXiv:2609.24106 (2026-09-21).

The material is FineWeb, a public corpus built from the web pages Common Crawl collects by keeping the English body text and running it through quality filters. It is in wide use as pretraining data for today's large models. The researchers took 110 snapshots, from the 2013 crawls through the 2025 ones, and kept only the sentences that end in a question mark, run 5 to 100 words, and either open with a question word or carry an interrogative expression. After Unicode normalization and case folding, 13.4 billion were left.

Half of this study sits in the unit of counting. The researchers counted not how many different questions exist but how many times a question appeared on a page. The same sentence sitting on 100 pages counts as 100. Anyone scraping questions off the web has generally read that number as the size of a demand. Appear often, wanted often. The paper nails this down at the outset. Occurrence counts, it writes, "measure how often a string was published and not how often it was asked."

So each question got a tag for where it was written. Five classes, and the first rule that matches wins.

• Posted by a person — a host classified as a forum or message board, or a page whose address carries a path such as /forum, /thread, /topic or /community.
• Copied — not caught by the rule above, and the same sentence appears word for word on two or more different domains.
• Commerce FAQ — a host classified as a shop or a customer support page.
• Editorial — a host classified as news, a blog or a magazine.
• Unclassified — whatever matched none of the four.

Reading the classification rules keeps you from over-reading the numbers that follow. The copied tag is not a verdict that a machine wrote the sentence. It is an observation that the identical string sits on several sites. The paper states that it ran no machine-generated-text detector at all, on the grounds that a question averages about twelve words and no detector is trustworthy at that length.

The paper names its own weak spot first. The unclassified share, the questions that matched none of the four rules, grew from 61% to 73%, and the researchers wrote that this is the weakest part of their scheme. They also published a comparison against human annotators. Agreement between annotators on where a question came from was κ 0.52, which is passable; agreement on whether a person genuinely wanted the answer fell to κ 0.23. People cannot settle that among themselves. So the paper claims nothing about genuine or fake, and claims only where the question was written.

The researchers took the 1,000 most frequent questions of 2015, 2020 and 2024 and counted the tags. Copied strings were already 71.7% of the 2015 list, then 86.8% in 2020 and 93.2% in 2024. In those same years, copied strings account for only 3 to 4% of distinct questions corpus-wide. A class that is rare across the web as a whole occupies nearly all of the top of the chart.

The other side emptied at the same pace. Among the top 1,000, questions coming from the places where people post ran 6.3% in 2015 and 0.3% in 2020, and they stayed at 0.3% in 2024. The top of the chart became a slot that fills by a logic other than human curiosity.

Looking at the sentences themselves is the fastest way to see what sits up there. The most duplicated question of 2020 was "what are you waiting for", with 72,331 occurrences. Next came "how can we help you". Neither is a sentence someone threw out wanting an answer; both are the boilerplate you drop into a banner or a chat widget. They only wear the shape of a question.

The genuine-curiosity side was never peaked to begin with. 97% of distinct questions appear once or twice across the whole period. All of the top 1,000 together make up less than 3% of occurrences. Human asking lies scattered along a long tail, and copy-paste builds the peaks.

This is where the study touches practice directly. Public datasets built by harvesting questions off the web already number several, and the paper's appendix gathers the family into one table. CCQA scrapes the schema.org question markup embedded in pages; WebFAQ narrows that to FAQ markup only. GooAQ takes its questions from search-box autocomplete, and PAQ says up front that it was machine-generated. Of the first two the paper writes that they "treat the result as natural asking".

This study counted the questions written into body text with no such markup, and by count those are the majority. No prior work, the researchers say, has looked at provenance and change over time at the level of the question. Anyone who weights questions scraped off the web by how often they occur, or feeds the most frequent ones first into training and evaluation, is weighting copied strings rather than human interest. The paper overturns no single dataset's quality. It overturns the habit of reading frequency as demand.

The time series is cut into five eras: a 2013–2015 baseline, the search optimization era of 2016–2019, the COVID period from 2020, the early AI period from late 2022 when ChatGPT arrived, and 2024–2025. Across that span the share of occurrences posted by people dropped from 9.3% to 1.9%. As a fall, 79%. Over the same years copied strings rose 25.9%, and commerce FAQ, starting from a small base, doubled.

The next step is the one that makes the study worth trusting. Twelve years of crawling grew and changed what it reached, so the first job is to split a change in the web from a change in the crawl. The researchers checked it two ways: regressing out crawl size, and recounting on a fixed panel of the 79,826 hosts present in at least 90% of all snapshots.

They wrote up the result in the direction that cuts their own claim. The claim that copied strings grew as a share did not survive the adjustment. The regression coefficient came out at −2.0% with a p-value of 0.49. Measured instead as how many land on a single page, a 13.3% increase held, at p 0.006. So the paper pinned itself down: it claims "the growth of the manufactured layer as a rate and not as a share."

They held the same ruler against the decline in questions posted by people. On the fixed panel the forum decline shrinks to 42–56%. The two corrections converged at roughly 45%, and the researchers recorded that value as their best estimate while stating that the 79% measured crawl-wide should be read as an upper bound. The direction does not change; the size is about half.

The side that filled up has its own evidence, not only the side that emptied. The share of pages whose address carries a path such as /faq or /questions-and-answers rose from 0.82% to 1.42%, steepest between 2017 and 2019. That window overlaps the spread of the structured markup that folds an FAQ into a search result. Two separate strands of data point the same way. Once putting a question on the page earned a slot in search, question-shaped text multiplied.

At the point where you want to read AI into this current, the paper hits the brake. Verbatim duplication across hosts, the evidence behind the copied tag, peaked at 21.6% in the March 2021 snapshot and came down to 18.3% through the AI eras. The peak precedes the ChatGPT release by twenty months. The paper states that it cannot attribute any question to model-written text. The force that turned the web's questions into something built for machines arrived before generative models did. The name the paper itself gives that stretch of years is the search optimization era.

The paper's title says you can tell who is asking. So what do you go by? The researchers ran part-of-speech analysis and regular expressions over a sample of 2.2 million questions to measure the shape of each one, then tested whether those features alone could predict where it came from.

One expectation missed first. Whether a question opens with a question word or takes a yes-or-no shape told the classes almost nothing apart. The association came out at 0.03, the only one of five pre-registered hypotheses to fail. Length did the work instead. Questions posted by people average 12.9 tokens, copied strings 8.5. The share carrying an explanatory sentence around them splits 7.3% to 3.9%, and the share using first-person words such as I or we splits 29.7% to 21.4%. Long, with the circumstances attached, and with something of the writer in it: that is the human side.

A model fed those three signals separated questions posted by people from copied strings at AUC 0.725. Useful, and some way from perfect. Swap the opponent to commerce FAQ and it sinks to 0.554. A coin flip is 0.5, so it effectively cannot tell them apart. The paper's conclusion is honest here. Form picks out the slapdash question farm; a commerce FAQ that someone wrote with care gives nothing away by its looks.

Hold the same signals against the twelve-year series and you can see which way the web's questions moved. Taking questions as a whole, without splitting them by class, mean length fell from 13.41 tokens to 11.85, down 11.6%. Over the same twelve years the share with background attached fell from 9.95% to 6.12%, down 38.5%. The paper describes the change as the human around the question diminishing. A line of background, a note about one's own situation: that part goes first.

How far a question stands on its own was measured on the same sample. Split by part-of-speech rules, a little over half are complete in a single sentence, 38% are fragments of a conversation, and 12% carry a word whose referent you have to look back for. The portion needing surrounding context to resolve fell over the series from 14.0% to 8.6%. The other side, the rise in the self-contained portion, was not statistically significant, so the paper made no claim about it.

The context analysis hands an especially uncomfortable result to anyone who cleans data. Questions posted by people depend most on surrounding context to make sense, at 19%, and commerce FAQ is the most complete in a single sentence, at 59%. A filter that keeps only clean, self-sufficient questions cuts the human side first and lets more of the manufactured side through. A quality filter can work in the direction of lowering quality. The context labels themselves reach only 63% accuracy against a human check, which is why the paper calls this part exploratory.

4.1A Detection Method Is Also an Evasion Method

A conflict-of-interest disclosure sits near the end of the paper. All the authors are with Bodhium Labs, a company that builds commercial products in generative engine optimization, and the same company funded the work. Generative engine optimization is the business of shaping content so that an AI answer cites your page. The team that measured how the web's questions are being manufactured for machines works at a company selling the product that helps with that manufacture.

The researchers did not hide the tension; they wrote it down as dual use. A classifier that separates questions posted by people from manufactured ones serves the side filtering data, and it serves just as well the side writing text the classifier will not catch. Read the signals from the first part of section 4 backwards and you have an evasion manual. Make the sentence longer, attach the circumstances, mix in the first person.

The sentence from this paper that will last longest is not a number but the definition in section 1. What the web lets you count is publication, not asking. Anyone working with data can restate it in other words. Volume is no substitute for provenance.

When we talk about the quality of training data, we mostly start by counting size. How many hundreds of millions, how many terabytes, how much duplication was stripped out. This study shows that inside one and the same corpus, the more frequent a sentence is, the further it stands from a person. Weighting by frequency, taking the most frequent as representative: the practice itself manufactures a bias. How much of a thing you have does not reveal that bias. Counting where each piece came from does.

This is why Pebblous talks about provenance whenever it talks about AI-Ready Data. Reading the paper leaves one question to turn back on our own data. In the corpora we use for training and validation, can we count today the share a person actually posted? The four items below are not items from the paper; they are questions raised by carrying its method over to our own work.

• Can you look up, right now, which host each collected document came from? Provenance classification is possible only while that record survives.
• Have you counted how many times the same sentence appears across different domains? Cross-domain duplication survives document-level deduplication.
• What does your quality filter throw out first? A rule that strips short, context-dependent sentences can throw out the human-written side first.
• Where do you use frequency as importance? Sampling weights, evaluation set selection and frequently-asked-question lists are all such places.

None of the four asks for a new model or a new tool. They ask what you wrote down alongside the data at collection time. Separating provenance at the moment of collection is something an earlier article on separating training data by source took up from the data audit side. Provenance is not recoverable later. Write it down when the document arrives, or counting the human share in that corpus stays an estimate forever.

Thank you for reading this far. Every figure this article cites can be checked by anyone in the full text of arXiv:2609.24106. We are curious how far you record provenance in the data you handle. If there is something you set out to count and could not because the record was missing, we would like to hear about it.

Primary source

• 1.Zhou, C., McCloskey, V., & Srinivasan, K. (2026). "You Can Tell Who's Asking: What the Web's Questions Are Made Of, and Where They Come From." arXiv:2609.24106 [cs.CL].

Datasets

• 2.Penedo, G., et al. (2024). "FineWeb." Hugging Face. The corpus this study drew its 110 snapshots from.
• 3.Common Crawl Foundation. "Common Crawl." The crawl underlying FineWeb.