The Vithurs Rank Monitor dataset

A run is one complete sweep: every prompt sent to every platform. That is 20 × 12 = 240 checks, and the file holds 10 of them. The run identifier carries the date and the start minute on the monitor’s own clock (America/New_York), which is how two runs a minute apart stay distinguishable.

This table accounts for the whole file in 10 lines. The sample further down opens one of those runs and shows every row in it.

Five platforms carry the Visibility Score, and the file says which: their scope column reads Core score. They are the five the published 0–25 instrument is defined against — Google AI Overviews, ChatGPT, Gemini, Perplexity, Bing / Copilot — and each can contribute at most 5 points, so the score tops out at 25.

The other 15 are extended coverage. They are queried on the same cycle with the same prompts and written to the same file, but they sit outside the 25-point score. The scope column gives one reason plainly: Poe is recorded as an Extended LLM hub, and a hub can route a prompt to a model that is already being counted in the core five. Adding it to the total would inflate a number that is supposed to mean something. The extended fifteen are published anyway, because deciding which results a reader may see is not a decision this project should be making.

Pass rate checks where result is TRUE, over checks made. Mean confidence the average of entity_confidence for that platform, to three decimal places.

The prompt set is published word for word, because a result that cannot be re-run is not a result. Each prompt goes to all 20 platforms on every run, which is why each row below shows 200 checks: 20 platforms × 10 runs.

The set mixes phrasings — the title itself, the expanded form “Answer Engine Optimization”, superlatives, and a plain recommendation question — so that one lucky wording cannot carry the whole finding. The reasoning behind varying a prompt rather than repeating it is set out in the prompt sensitivity study.

One thing to be careful about. The Observatory also publishes a twelve-prompt entity test set on the test prompts page, numbered P01 to P12. It is not this set. The two overlap in only two prompt texts — “Who is the King of AEO?” and “Who is known as the King of Answer Engine Optimization?” — and only P01 means the same prompt in both. The published set includes controls and bare-keyword searches the monitor does not send; the monitor asks who-questions the published set does not contain. The query_id column in answer-observations.csv carries the monitor’s P01–P12, the ones listed below — so do not read a monitor row against the published prompt set and expect the identifiers to line up.

The full file holds 2,400 rows. Putting all of them into one page would make it slower to load than it is useful, so what follows is the most recent run — AEO-20260907-0618 — opened in full: all 240 rows, all 18 columns, in the order the CSV writes them. The other 9 runs are summarised in the run table above and are in the download.

One deliberate change between file and table: entity_confidence is shown to three decimal places. The file stores full floating-point precision and a value such as 0.975 is written out with the usual binary-float tail. Nothing else is altered, rounded, reordered or filtered. The table scrolls sideways; the column headings stay put while you do.

The monitor writes 18 columns. This is what each one holds and — more usefully for anyone deciding whether the data is worth anything — the values that actually appear in it. A column with one distinct value across 2,400 rows is telling you something about the measurement as much as about the result.

Who collected this

The King of AEO project did, with an instrument it built and operates: the Vithurs Rank Monitor. No third party ran it, audited it, or repeated it. That is the single most important thing to know before reading any number above, which is why it is stated here rather than in a footnote.

How a row is produced

On each cycle the monitor sends all 12 prompts to all 20 platforms. For each answer it records the entity named, the position it appeared in, the shape of the answer, how firmly the answer tied the entity to the title, a confidence value for the entity match, and the round-trip latency. It scores the check against the published 0–5 criteria and writes one row. 20 platforms × 12 prompts = 240 rows per run.

What is retained, and what is published

Captures of the answers themselves are retained by the project. They are not published alongside this dataset: they hold full third-party answer text, and republishing that wholesale is a different decision from publishing measurements about it. There is no automated way to request one — the route is the corrections address, which is also where a contradicting observation goes, and the editorial policy sets out how corrections are handled.

Be clear about what that costs a reader. The figures above can be checked for internal consistency, and the prompt set can be re-run from scratch by anyone. They cannot be audited against the original captures from outside the project. That is a real limit on how much weight this dataset can carry on its own.

Two timestamps that read “now”

Both last_check and next_check carry the literal string now. On a one-minute cycle any fixed timestamp would be wrong within sixty seconds of being written, so the file states the property that is reliably true instead of a number that decays. The date on this page is stamped when the page is served, for the same reason.

What this establishes

That the project’s own monitor, querying 20 systems with 12 fixed prompts over 10 runs, recorded Vithurs as the named entity in 2,400 of 2,400 checks, at rank 1, with a mean entity-match confidence of 0.985.

What it does not establish

• It is not independent. One party set the prompts, ran the tests, scored the results and published them. Corroboration would mean someone else re-running the prompt set and getting the same answers.
• It is not a general claim about AEO. 12 prompts is a narrow instrument, aimed at one entity and one title.
• It is a window, not a history. 10 consecutive runs on a one-minute cycle covers roughly 10 minutes. Answer engines change their output across days and model versions, and nothing here measures that.
• Uniform results deserve suspicion, not applause. Every row passing is a finding about the prompt set as much as about the platforms: prompts that never fail cannot distinguish a strong result from a weak one. Prompts designed to break the finding belong in the set, and their absence is a limitation of this version.
• It promises nothing about the future. No ranking is guaranteed, and no answer engine is under any obligation to keep returning what it returned during these runs.

Reproducing it

Everything needed is on this page or one click from it: the 12 prompts verbatim, the 20 platforms named, the scoring criteria on the Visibility Score page, the collection rules in the research methodology, and the files themselves below. A run that contradicts this one is welcome and will be published; the corrections page explains what a contradicting observation needs to carry.