Research

Why this page exists first

A great deal of what circulates as “GEO research” is a screenshot, a small unrepresentative prompt set, or a vendor’s own funnel described as a finding. It spreads because the subject is new and the numbers are memorable. We would rather publish nothing than add to that.

The standard for anything published here

• The exact prompts or queries used, published in full — not summarised.
• The surfaces tested, and the dates they were tested on.
• How many observations, and how often each was repeated.
• The raw variation between runs, not only the average.
• What the result does not show, stated as plainly as what it does.
• A clear statement where we have a commercial interest in the conclusion.

What makes this genuinely hard

1. 01Answers are not stableThe same prompt can return different sources minutes apart. Any single observation is a sample, and a study that treats one as a fact is measuring noise.
2. 02Results are personalised and regionalWhat one account sees is not what another sees. Without controlling for account state and location, a comparison across brands is not comparing like with like.
3. 03The systems change without noticeRetrieval behaviour shifts with model and product updates that are rarely announced. A finding is a statement about a period, not a permanent property.
4. 04Correlation is cheap hereSites that get cited also tend to be established, well-linked and well-structured. Attributing citation to any single tactic usually cannot survive that confound.

The research you can do yourself

Your own prompt set, checked on a schedule against your own market, is worth more than any published benchmark. It is specific to your buyers, it controls for your competitors, and you can act on it. That is exactly what AI visibility tracking is for.