Method
What we do not claim.
We do not claim these figures describe AI systems generally. They describe one engine.
We do not claim they generalise beyond this population of thirteen companies in one category.
We do not claim the measurement would return the same values in a different window. Our own second finding indicates it would not.
We do not claim pre-registration has been validated inside AI-discovery measurement. It has not been. It is imported from fields where it has been, and that import is an argument, not a proof.
We do not claim our instrument is correct. We claim it is sealed, that its corrections are published, and that its output can be checked against a record we cannot alter.
This page describes how the measurement on the Findings page was produced. It is written so that a reader who does not believe the findings can identify exactly where to attack them.
Three controls govern everything below: the method is fixed before the evidence exists, absences are recorded as outcomes rather than discarded, and every value is audited against the basis it rests on.
None of the three is our invention. Each is the standard remedy in a field that has had to survive the accusation of finding what it went looking for.
Population.
Thirteen companies whose product is the measurement or improvement of a business's visibility in AI-generated answers. The population was defined and fixed before measurement.
The measurement reported here was taken in a single session on 7 August 2026, beginning at 15:54 UTC and ending at 16:04 UTC.
Under ten minutes. It is one session, not a campaign, and nothing on these pages should be read as describing sustained observation over time. Our own second finding — that the same query returns a different answer set on repetition — is precisely why a single window is a single window and is reported as one.
Companies are not named. The finding concerns the behaviour of the answer engine across a category, not the performance of any individual company, and naming would convert a methodological result into a competitive claim.
The population is small, it is one category, and it is not a sample of anything broader. Any reader treating these figures as generalisable beyond this population is reading them incorrectly.
Instrument.
Measurement was performed on Perplexity. One engine. Every figure on the Findings page carries that qualifier, and it is not stylistic — a result measured on one engine is not a result about AI systems generally.
The unaided appearance rate is the rate at which a company appears in an answer to a query that does not name it. Aided appearance — what happens when you ask an engine about a company by name — is a different measurement and is not reported here.
Repeated runs were used to measure the stability of the answer set. Agreement between runs is reported as a range rather than a point estimate, because a point estimate would imply a precision the underlying behaviour does not have.
Absences are results.
When a surface returned nothing, that was recorded as data and retained.
Nothing was dropped as a failed run. Nothing was re-queried until it produced output. The most common way a visibility figure is inflated is by discarding the runs that returned nothing, and a median of zero is only reachable by an instrument that refuses to do that.
Values are audited against their basis.
Every value carries the material it rests on. Where the engine returned a value with no supporting basis, that was recorded as a distinct state — not as a low score and not as a missing one.
The Findings page reports what happened when four such values were removed: nothing observable. That result is the argument for this control. An instrument that grades only the value cannot distinguish a fabricated number from an earned one of the same magnitude, and will report their removal as no change.
Pre-registration.
The specification — population, instrument, queries, counting rule, and the conditions under which a result would count — was written, sealed, timestamped and signed to an append-only ledger before any surface was queried.
The order of operations is the evidence. A method fixed after the results are visible can be fitted to them; a method fixed before they exist cannot.
This sequence is standard in medicine, where trial registration is required at or before the first patient's consent as a condition of publication; in psychology, where Registered Reports receive acceptance before results are known; and in assurance, where independence preconditions must be established before the engagement rather than asserted after it.
The transfer of that practice into AI-discovery measurement is an institutional analogy and we mark it as one. Nobody has validated pre-registration inside this specific domain, because nobody has run it here.
Verifying a JFD-marked document (SPEC-JFD-DOCUMENT-MARK-v1 §4).
Every instrument we issue carries a three-line mark: what we assert (Line 1), what you can check (Line 2), and how (Line 3). The document hash is computed so that YOU can reproduce it from the file in your hand:
1. In the delivered PDF's raw bytes, locate the two mark strings exactly as printed on the document — the SHA256 line and the VERIFY line.
2. Replace each with a run of '#' characters of identical byte length.
3. SHA-256 the result. Its first 16 hex characters must equal the value printed on the document.
4. Load the verification page named on Line 3 (jetfyul.com/v/{document number}) — it renders the full digest, the ledger and OpenTimestamps state, and whether the document has since been SUPERSEDED or WITHDRAWN. A document that cannot say it was withdrawn is not a record.
5. Where a .ots proof accompanies the document, ots verify anchors the hash in Bitcoin — no contact with us required.
The hash is computed over the document with Lines 2-3 as placeholders because a hash cannot cover itself; the substitution is byte-length identical, which is what makes step 3 reproducible by a stranger. A hash a verifier cannot reproduce is decoration.
Changelog
Changes to this method.
Every change to the method is recorded here permanently, with the date and the reason. Entries are not removed. Where an entry has itself been corrected, the correction appears as its own dated entry below and states exactly what was struck.
A method that changes silently is not a method.
| Date | Change | Reason |
|---|---|---|
| 7 August 2026 | Composite scoring corrected. Median composite fell from 62.5 to 50.0. | The prior composite treated an empty component as a scored zero, which projected a value from an empty set. Under the corrected instrument an empty component is not scored and does not project. Every composite produced before this change was inflated. The corrected figure was published and the earlier one retired. |
| 11 August 2026 | The claim that the measurement reproduced 40/40 was withdrawn from this page, from the Findings page and from the Receipts page. | It was not supported. The 40/40 figure was a re-score of a different population, not a re-capture of the thirteen companies, and the run behind the published figures carried no seal at all. A sealed run has since been performed against our own estate; it is a first measurement of a different subject and does not reproduce anything on the Findings page. The claim was removed rather than restated at a lower number. |
How does the sealed method work?
The method commits its pre-registration to an append-only ledger before any capture, then measures a single engine across a fixed question set. The seal timestamp precedes every evidence timestamp on the chain.
Sources — entity catalogued at Wikidata; U.S. service mark at the USPTO (Reg. 8009699). Last modified 2026-08-17.