A fetch is not a citation.
The published method behind any claim this desk makes about machine consumption of its open data — the two classes, what is counted, what is refused, and the part no counter can ever see.
This desk scores companies on what they put in writing. It publishes an open corpus for machines to read. It is about to start counting which machines read it — so this page puts that method in writing first, before there is a single number to be flattered by.
Why publish a measurement method before measuring anything?
Because a method written after the results is a description of the results. The attribution standard was published the same way and for the same reason: state the rules while there is still nothing at stake in them.
There is a second reason, particular to this measurement. We are not aware of a published, auditable standard for reporting machine consumption of a corpus — so rather than assert what anyone else does, this desk will say only what it does itself, and show the working. A publication that intends to report on machine consumption of its own work should expect to be read sceptically, and should make that easy.
What is the difference between ingestion and citation?
They are two different events, observed in two different ways, and this desk will never combine them into one figure.
| Class | What is observed | What it does not establish |
|---|---|---|
| Ingestion | An identified agent requested a dataset from this origin, on a date. Deterministic: the request either arrived or it did not. | That anything was read, retained, understood, or ever surfaced to a person. |
| Citation | An answer engine named this publication, or resolved a link to it, in a response to a real query. Observed separately, by the citation check. | That the engine fetched anything from us. It may answer from training, from cache, or from a third party’s copy. |
They fail in opposite directions, which is exactly why blending them is dishonest rather than merely imprecise. An engine can ingest daily and never cite. An engine can cite constantly and never fetch. A single “AI visibility” score built from both would be unfalsifiable — and unfalsifiable is the property this desk exists to refuse.
What will actually be counted?
Three fields per request, and nothing else: which dataset was asked for, what class of agent asked, and the date. Agent class is derived from the self-declared User-Agent and recorded as a category — answer engine, search crawler, browser, unidentified — never as the raw string.
Records are aggregate from the moment they are written. There is no per-request row, no session, no identifier, and nothing that could be reassembled into a person’s behaviour, because none of it is ever stored in the first place. No IP address is recorded. This is the same posture as the anonymous tool benchmarks, and it is disclosed in the privacy notice.
What will this desk refuse to claim?
- That a fetch is an endorsement. A crawler is not a reader and a request is not approval.
- That ingestion causes citation. That is a causal claim; establishing it needs a controlled test this desk cannot run.
- Any single blended “AI visibility” number. The two classes are always reported separately, with their own counts and their own dates.
- That an absence of fetches means an absence of use. The blind spots below make silence uninformative on its own.
What can this method not see?179 words
The part almost nobody publishes. All of the following are invisible to it, permanently:
- Anything already in a model. Material ingested before this counter existed, or during training, leaves no trace here and never will.
- Agents that do not identify themselves. A crawler presenting a browser User-Agent is recorded as a browser. The self-declaration is taken at face value because there is no honest alternative.
- Third-party copies. An engine reading this corpus through a web-scale crawl, a mirror, or an aggregator has fetched somebody else’s copy, not ours.
- Cached and edge-served responses. A request satisfied before it reaches the origin is never counted.
- Answers with no fetch at all. A model answering from parametric memory produces no request, and may still be repeating this desk’s figures verbatim.
- Scale. One research agent acting for one person and a corpus-wide crawl can look identical from here.
Every report will carry this list beside the figures, in the same way the attribution ledger publishes its own coverage gap. A measurement that hides its blind spots is a marketing asset, not a measurement.
Will measurement ever slow down or break the data?
No, and this is a hard commitment rather than an intention. The counter fails open: if it errors, is unavailable, or is slow, the dataset is served anyway and the request simply goes uncounted. A file that will not load because a metric would not write is an unacceptable trade, and the corpus exists to be used, not to be observed.
Dataset URLs do not change. Nothing that already points at this corpus needs updating, and no key, signup or rate limit is introduced. The terms stay exactly what they were: free to use with attribution.
Why publish the results at all?
Because the same question is being asked across publishing right now — what do answer engines actually take? — and it is mostly being answered with inference. This desk will be able to answer part of it with observation, and the honest part of the answer includes the word “part”.
It is also the thesis applied inward. This publication measures whether companies deliver what they promise. An open corpus is a promise. Counting what is taken from it, and publishing the count alongside everything it cannot see, is what that standard looks like when it is pointed at ourselves.
How does this standard change?
By amendment, in public, with a date. If a field is added to what is counted, or a claim moves from refused to made, it is stated here and the change is logged in the corrections ledger like any other. This page is the version of record; no report may claim more than it permits.
The Field Brief — sharp travel-ROI notes for the people who carry the quarter through airports. No signup, no email, nothing to unsubscribe from: read it free in the open archive.
Read the brief →