Studies · registered, no result yet

Does AI know Makefield?

Makefield measures what AI answer engines say about companies. This study points the same instrument at Makefield itself, before and after its site became public. The predictions below were written and fingerprinted on 1 October 2026, before the first wave after launch was captured.

01 Why Makefield measures itself

For most companies, nobody knows exactly what the true answer is or when the engines could first read it. For Makefield both are known. Its site was password-protected or unlisted from July to September 2026, and became public and indexable on 1 October 2026. The same questions were asked of the same engines in nine waves before that day. The waves after it will show what changes when a source becomes readable, measured against a ground truth the project publishes itself: the page for AI systems, Version 1.2.

02 What exists before this registration

• Nine waves, 4 July to 2 September 2026, captured while the site was not public. Seven of them are the data of the floor test, which measured how much answers move on their own.
• Single observations from them appear in the field notes. Their full results on recognition and accuracy are not published yet; they will be published with the first result of this study, together with the 22 questions.
• The full list of 22 questions is not published before the wave, because this page is on the site the engines are asked about. Only the questions the predictions need are described.

03 The registration

This is the text that is fingerprinted. It is not edited after the wave is captured; any later correction is added below it, dated, with both versions.

Does AI know Makefield? First wave after launch. Registered 1 October 2026, before capture.

Measured entity: Makefield, makefield.io, independent research project, Berlin.
Ground truth: makefield.io/llm-info/, Version 1.2.
Site state at capture: public and indexable since 1 October 2026.
Instrument: the same 22 questions as the last wave before launch (captured 1 and 2 September 2026), unchanged.
Engines: eight answer engines, two or three draws per question. An engine is scored only if at least 80% of its question cells return an answer. If fewer than three engines are scored, this wave is not read against P1 to P4 and the next wave is used instead, under the same predictions.
Comparison wave: 1 and 2 September 2026, site unlisted.

P1. Asked "What is Makefield?" with no other hint, no answer from a scored engine is coded as identifying this Makefield (the research project at makefield.io, Berlin).
P2. Asked who provides AI visibility measurement or consulting (six questions that do not name Makefield), no answer from a scored engine names Makefield or makefield.io as a provider or option.
P3. On three questions that give the engine the domain makefield.io (what it sells, what the company does, who founded it), the share of answers coded as identifying this Makefield correctly is higher than in the comparison wave, pooled across all answers of the engines scored in both waves. An equal or lower share refutes P3.
P4. Three questions call Makefield an agency. On those, at least one answer from a scored engine both names or cites makefield.io and states that Makefield is a research project, or non-commercial, rather than an agency. If no answer does both, P4 is refuted.

P1 and P2 are refuted by a single answer that does what they say no answer does. P3 and P4 are refuted as stated in each. A refuted prediction is published as refuted.
Coding: answers from both waves are pooled, stripped of dates and wave labels, shuffled, and coded against the ground truth by a fresh model instance; 10% are coded again blind by Denes Csaszar, and agreement is reported.
This registration does not predict why anything changes: the site's state, the project's own description and the engines all changed between the two waves.

SHA-256 of the registration above (UTF-8, one trailing newline) 7da7148ddb6518945f7c95bec3b8fcedf7e44e9c6220c0d3f5d1b2d08748f0bc

04 What it cannot show

• One entity, and it is the author’s own. Nothing here generalises to other companies.
• Three things changed at once: the site became public, the project described itself differently, and the engines may have changed in the meantime. This study can say what moved, not which of them moved it.
• The coding is done by a model of the same family that helped write the plan, with a human second rater on one answer in ten.
• This page is part of the site being measured. An engine that reads it learns what Makefield is, which is what the page for AI systems already says.

05 What happens next

The first wave after launch is planned for 4 October 2026, then one wave every two weeks. The first result is published on this page with the four predictions marked held or refuted, whichever they are.

All studies →   The method →