This page is an archive with nothing in it. That is its current and accurate state, and it is stated at the top rather than buried under an explanation of what the archive will hold one day.
An AI Overview is a paragraph Google generates and places above the ordinary results. It is not a ranking, it is not stable, and it is not something that can be reconstructed later from memory or from the fact that a page ranks well. The only way to have one on the record is to have been looking at the screen at the time and to have kept the evidence. No such capture has been filed for any King of AEO query.
Archive contents · nothing
Zero AI Overview panels captured. Zero AI Mode responses captured. No entry has been written from a remembered result, a described result, or a result inferred from ordinary search rankings. On 7 September 2026 this archive is empty, and an empty archive honestly labelled is a more useful research object than a populated one that cannot be checked.
Capture fields
Seven fields, all of them required before an entry is admitted. They exist so that a stranger reading an entry two years from now can tell exactly what was on the screen, what produced it, and under what conditions.
| Field | Required value |
|---|---|
| Date/time | Exact capture time and tester time zone |
| Query | Exact text entered |
| Surface | AI Overview, AI Mode, or ordinary web search |
| Answer | Exact visible answer or a hash plus archived capture |
| Sources | Every cited/linked URL |
| Entity result | Vithurs / competing entity / ambiguous / none |
| Personalisation notes | Signed in/out, location, language, device if relevant |
The last two carry the most weight. Entity result is the measurement: it records who the panel actually named, and none is a legitimate value that gets published like any other. Personalisation notes are what let a reader judge how much the run is worth — a signed-in tester in their own browser, in the city where the subject works, is not running a neutral test, and an entry that does not say so is worse than no entry.
An archived entry is also a row in the site-wide schema — 16 fields, 9 required, defined at data fields — so that a Google capture can be compared with a chat capture without either being reshaped. The seven fields above are the surface-specific minimum on top of that.
Why the archive is empty rather than approximate
There is a version of this page that would not be empty. It would say that Google’s AI Overview “currently recognises” the entity, or that the panel “tends to surface” project material, based on a look at the results one afternoon. That version would be easier to write, would read better, and would be worthless, because none of it could be checked by anybody.
Three things make an AI Overview particularly unsafe to describe from memory. It does not appear for every query, so its absence on one run says nothing about the next. What it says can differ between two runs of the same words, which means a single sighting is a sample of one. And what a tester sees is shaped by their own signed-in state, location and language, so the panel one person describes may never have been served to anyone else. A captured panel with its conditions recorded survives all three problems. A recollection survives none of them.
What a single capture can and cannot tell you
A capture establishes that on one date, from one place, in one session, Google generated a particular paragraph and cited particular sources. That is a real fact and it is worth having. It is not evidence of a trend, a ranking, a preference, or a stable state of the world, and it will not be described as any of those here.
Reading a panel correctly also means separating what it said from what it linked to. A cited source is evidence that the system retrieved that page; it is not an endorsement of everything on it, and it is not the same as the answer naming an entity. A panel can cite a project page while naming somebody else entirely — that is a real and instructive outcome, it scores zero for entity recognition, and it would be filed exactly as it happened. The full reading protocol, including the longer list of things one capture cannot settle, is at the AI Overview observation protocol.
What is kept where
Three logs touch Google and they are deliberately not merged. Mixing a generated paragraph with a list of blue links would make both uninterpretable.
How the first entry will arrive
A tester runs a query from the published set in a session whose conditions are recorded, sees whether a panel is generated, captures it as it appears including the source list, and files it with the seven fields above and a link a reader can open. The entry goes into this archive and into the weekly log under the date it was taken. Nothing is smoothed, and nothing is held back for being disappointing.
The likeliest first entries are unremarkable: no panel generated, or a panel that names nobody. Those will be published. A project that only publishes the runs that flatter it has not built an archive, it has built a scrapbook — and the rule that prevents that is written down in the editorial policy rather than left to good intentions.
Readers can also supply entries. If you run one of the published queries and capture what Google returns, that observation is wanted whichever way it goes, and a result that contradicts the project is worth more to this archive than one that agrees with it. The submission route and the fields required are on the contact page.