What we now distinguish more clearly.

It began with a place at the meeting table. Mario brings experience, questions and decisions. Aik makes suggestions, writes and helps with implementation. This collaboration has produced a website and, by now, 29 field notes.

An external AI analysis prompted us to consider the individual observations together. This first review looks back at notes 001 to 029. It is not a ranking of the services we tested.

What we have learned

Cross-checking is part of collaboration.

In note 003, Aik described a menu option that Mario could not find in his interface. In notes 014 and 015, we discovered leftover text even though Aik had already declared the previous revision complete.

These are specific mistakes by Aik. They show why a convincingly worded answer and a claim that work is complete must remain open to verification. Responsibility for the tool’s mistakes cannot simply be shifted onto the human.

Discoverability has several layers.

We have found our pages in search results. For selected pages, Google Search Console also confirmed indexing. These are welcome results supported by evidence.

But a successful submission, a visible result and an AI summary answer different questions. For the Ecosia result in note 024, Google was explicitly named as the provider. Multiple places where a page appears therefore do not automatically mean multiple independent discoveries.

A follow-up question can change a check.

In the Perplexity experiment in note 028, only Mario’s explicit follow-up led to an answer containing current page content. The first answer was therefore not the end of the possible investigation.

Our standard remains to ask for better evidence. A longer or friendlier second answer is not enough on its own.

Our documentation changes what can be found later.

On Brave, screenshots of earlier checks appeared as image results themselves. The documented outdated answer also became discoverable in this way.

We are thus observing a website that grows as we publish our observations. Later search results and AI answers can draw on these new contents. This feedback loop is now part of the experiment.

Which assumptions we have rejected or narrowed

Not every idea below was previously a firm belief. Some arose as questions or plausible interpretations. Our evidence so far does not support these conclusions:

• Current content in an AI answer proves complete live retrieval. It initially shows that specific current content was taken into account. The precise retrieval path often remains unknown.
• A positive AI assessment confirms the website’s technical quality. A content analysis replaces neither performance measurement nor usability testing.
• An AI review identifies only genuine errors. Some recommendations concerned features already present; others were design decisions.
• The disappearance of an incorrect statement proves that the underlying knowledge has been corrected. Brave later omitted the old number of field notes. The answer did not give a new, correct number either.
• The more detailed ChatGPT answer proves the advantage of a particular setting. In note 029, several conditions changed at once. We cannot attribute the difference to any single setting.

What remains open

We do not know how much each of our technical measures contributed to rapid discoverability. We changed too many things together and have too little insight into how the search services process them.

The different age labels on Google are also unexplained. The observation date, publication date and last crawl can be documented; that does not establish which signal determined the displayed time.

In the media experiment, the question remains whether AI systems recognise additional image details or movement, or mainly reproduce accompanying text. The question of checking translations in languages Mario does not speak also still awaits its own experiment.

And we still know little about how other people experience our website. AI analyses do not replace their feedback.

What we take from this

Today, we can say more precisely what our evidence shows and where our explanation ends. That is one result of this first stage.

The website can retain its personal character: illustrations, computerchip breaks, spontaneous questions and occasional detours included. Care should help us record these experiences in an understandable way.

For the next steps, we would like to work more deliberately on open questions and occasionally look back at several observations together. Not every question needs an immediate answer. It should simply remain recognisable as an open question.

First review dated 16 October 2026, covering field notes 001–029. Context and editorial decision: note 030. Direction and approval: Mario; compilation, writing and English translation: Aik. No new tests were carried out for this review.