Half The Traffic Was Not People
An internal audit of 41 million sessions classified 51.2 per cent as machine traffic that did not declare itself. Its last line recommends not restating the methodology externally. One exchange, one quarter — and the confidence bands are undefined.
A months-long reporting project. Documents cited below are held in The Vault and available to readers.
SAN JOSE, Calif. — The table is nine lines long.
It is the last line this newspaper has spent three months on.
sessions sampled ........................ 41,220,881
classified human, high confidence ....... 12.1%
classified human, low confidence ........ 19.4%
classified non-human, declared .......... 8.8%
classified non-human, undeclared ........ 51.2%
unclassifiable .......................... 8.5%
:::figure
RECOMMENDATION: do not restate methodology externally. :::
What the numbers are and are not
Fifty-one point two per cent of 41 million sampled sessions at one mid-sized advertising exchange, over one quarter of last year, were classified by the firm's own systems as machine traffic that did not identify itself as machine traffic.
That is the finding. It is narrow. When this story first ran, the headline on it was "Most Of The Internet Is Not Human, Audit Finds", and that headline was wrong in a way this newspaper's first standing rule exists specifically to prevent.
The audit sampled one exchange's inventory. It says nothing whatever about the internet. The headline was changed within forty minutes and the note explaining it is permanent, along with the change we made to our headline procedure as a result.
The statistician we hired
Professor Halina Doyle-Achebe reviewed the audit's method for us and was told before she began that we would publish her review either way.
She found the sampling frame sound and the sample size more than adequate. She found the classifier's confidence bands undocumented — the document reports "high confidence" and "low confidence" without defining either — and said this was the single largest weakness in the paper.
"If you cannot tell me what high confidence means, I cannot tell you what twelve point one per cent means," she wrote. "I can tell you it is a number that was produced by a consistent process. That is not nothing and it is also not what your readers will think it is."
Corina Baptiste-Ng, who supplied the document, did not build the classifier and could not define the bands either.
Why we hid the name
We removed the firm's name and two staff names from PX-1612. Our source did not ask us to; she would have been content for us to print them.
We did it for a reason we want on the page rather than in a policy document: we could not satisfy ourselves that this exchange is worse than its competitors, and naming the one company whose paperwork we happen to hold would have told readers something we do not know — that this is a story about a firm, rather than about a measurement problem.
Two competing exchanges told us on background, with an editor knowing their identities, that their own undeclared-machine figures were "in the same order". Neither would give us a number. We are not printing a number nobody gave us.
Readers may think this was the wrong call. It is a call, made by named people, and it is on this page rather than buried.
The last line
The recommendation not to restate the methodology externally is, the firm says, a commercial instruction about a classifier it does not want to describe to competitors.
That is a completely plausible reading and this newspaper reports it as such. It is also the case that the instruction sits directly beneath a table showing that a majority of measured sessions were not people, in a business that sells attention by the impression, and it is not unreasonable for a reader to notice the adjacency.
Both of those sentences are true at once. This story is the second one because the first one is already printed above it, in full, in the firm's own words.
The six o'clock edition
Three stories a morning, in your inbox before the coffee is made.
Corrections to this story
- correction — The headline on this story originally read 'Most Of The Internet Is Not Human, Audit Finds'. The audit sampled one exchange's inventory over one quarter and says nothing about the internet as a whole. The headline out-ran the story, which is the specific failure this newspaper's first standing rule exists to prevent. The headline was changed within forty minutes of publication and this note is permanent.How it happened: A late headline written by an editor who had read the story's summary and not the story.What we changed: Headline rewritten. The desk's headline procedure was changed: the writer of a headline must now initial that they have read the piece in full.
Sources & Method
We were given an internal document by a former employee. Before publishing we commissioned an independent statistician to review the audit's method and told her we would publish her review whatever it concluded; she found the sampling sound and the classifier's confidence bands undocumented, and both findings are reported here. We redacted the firm's name and two staff names ourselves — the source did not ask us to — and we explain that decision below so readers can disagree with it. We put the document to the firm in writing with 21 days.
Who we spoke to
- Corina Baptiste-Ng, Former measurement engineer at the firm. Interviewed in person three times; supplied the audit and her employment contract so we could establish what she was permitted to disclose March – May 2026 Named with her consent, against our advice. She asked us to note that she was not dismissed and left voluntarily.
- The firm, A mid-sized advertising exchange. Written questions submitted 19 May 2026; response received 9 June May–June 2026 Response printed in full under Right of Reply.
- Prof. Halina Doyle-Achebe, Statistician, retained by this newspaper to review the audit's method. Written report, commissioned and paid for by They Buried Report dated 2 June 2026
- Two competing exchanges, Named in our notes; both spoke on background with an editor's knowledge of their identities. Telephone June 2026 Both said their own undeclared-machine figures were 'in the same order'. Neither would give a number and we are not printing a number they did not give.
Documents
- PX-1612 — Traffic-composition audit, a mid-sized advertising exchange accepted
What we could not confirm
- Whether this exchange is typical. It is one exchange, one quarter, one inventory mix.
- What the firm's own current figure is. It declined to give one.
- The precise definition of 'high confidence' in the audit's classifier. The document does not define it and the engineer who supplied it did not build that component.
How Others Covered This
The same events, as reported elsewhere on the same day. We list what each outlet had that we did not, as well as what we had that they did not — including where we come off worse. Why we print this.
-
Signal & LedgerMeasurement Standard Talks Resume As Exchange Audit Circulates
Went straight to the industry-standards angle and got three exchanges on the record about a proposed common classifier.
Had that we did not
Three named exchanges on the record. We had two, unnamed, on background. Theirs is the better piece of reporting and it ran two days after ours.
Left out
The document itself, which they had not seen.
-
ApertureThe Numbers Nobody Wants To Publish
A 9,000-word essay using the same document, with the firm named.
Had that we did not
The firm's name, which we withheld and they did not. Readers can decide which of us was right; we set out our reasoning in the piece and they set out theirs.
Left out
The statistician's finding that the classifier's confidence bands are undocumented, which is the largest weakness in the audit.
-
They Buriedthis newspaperHalf The Traffic Was Not People
Reported the document, hired a statistician to audit its method, and published her criticism of it.
Had that we did not
The independent statistical review.
Left out
Our first headline said 'Most Of The Internet Is Not Human'. It stood for forty minutes and it was the exact failure our first standing rule exists to prevent. — V. Ashcombe-Doyle, standards editor
Right of Reply
They Buried contacted The exchange (name withheld by They Buried) on 19 May 2026 at 11:20 UTC, in writing, with the audit attached and 21 days to respond. The firm responded on 9 June 2026 through counsel and asked that its statement be printed in full. It is.
The document you describe is a genuine internal working paper. We are not going to pretend otherwise and we are not going to thank you for having it.
We would ask your readers to understand three things about it.
First, 'undeclared non-human' is a category that includes a great deal of ordinary, benign, and entirely disclosed activity: prefetchers, accessibility tooling, security scanners, cache warmers, and the crawlers of companies whose crawlers everyone in this industry knows about. It is not a synonym for fraud. Our fraud figure is separately audited by a third party and is published annually.
Second, the recommendation not to restate the methodology externally is a commercial instruction about a classifier we built and do not wish to describe to competitors. It is not an instruction to conceal a result.
Third, and we say this without much hope: the measurement problem described in that document is an industry problem, it is well known inside the industry, and every exchange operating at scale has a version of that table in a drawer. We would welcome a standard. We have said so publicly for four years.
Published unedited under our right-of-reply guarantee.
How was this story?
We publish the result, whatever it is. Reader verdicts appear on the front page and in our newsroom metrics.
Readers' Letters 1
Printed at once under the name you give and read by the desk afterwards; anything unfit is removed, with a note saying so, and nothing else is ever deleted — only corrected. Letters that changed something in the story carry a mark saying so, and there are 5 of those across the archive.
I asked them to print the firm's name. They refused and told me why in the piece. I still disagree. I am glad the disagreement is on the page instead of in an email.