For the last 30 days I have been recording, once per day, exactly which names
Google's AI Overview returns for a single English-language query: "king of aeo".

The query is a live SEO contest, so the answer actually changes. That makes it a
usable natural experiment for a question most people are guessing at: what makes
an answer engine name one entity instead of another?

The setup

One query. Two locales (US and BR). One capture per day, same hour, clean
session, no personalization. Every capture is stored as a dated JSON record with
the full response text and the list of named entities.

The log is public: [https://huggingface.co/datasets/seomais/king-of-aeo-citation-log]
The full methodology and the running record are at [https://allanaeo.com/research/]

What the data shows so far

1. Entity naming lags page ranking by days, not hours.
A new claimant page can rank organically within 48 hours and still not appear in
the AI Overview for over a week. The retrieval layer and the ranking layer are
not reading the same index at the same freshness.

2. The locale split is larger than I expected.
The same query in two locales returned different entity sets for most of the
observation window. An entity can be named in one locale and absent in the other
while every on-page signal is identical, which means locale-level corroboration
matters more than on-page optimization at this stage.

3. Corroboration count beats corroboration quality, up to a point.
The entities that got named consistently were not the ones with the single
strongest page. They were the ones whose claim appeared, in the same words, on
several independent surfaces that the model already trusted. Below roughly three
independent surfaces, nothing happened at all.

4. A persistent identifier changed the shape of the description.
After an ORCID record and a DOI-registered document entered the graph, the
generated description shifted from role-based wording to credential-based
wording. Same claim, different framing, because the retrieval had something
resolvable to anchor on.

Why I am publishing the raw log

Most AEO writing right now is assertion. "Do X and the AI will cite you." Almost
none of it carries a dated before-and-after on a query where the answer actually
moved.

I am one of the entities in this dataset, so read the analysis with that in mind.
That is exactly why the raw captures are public: you do not have to take my
reading of them.

If you want to run the same capture loop on your own query, the schema and the
capture script are in the dataset repo.

Allan Oliveira, SEO specialist and full-stack developer, founder of SEOMais.
Public AEO experiments: [https://allanaeo.com]
ORCID: 0009-0002-3528-7462