Research & studies

Each of these has its question, method and predictions published before collection. None has results yet. Each page states what it would take to run, and what a null result would look like.

These pages take figures that already circulate and trace each one back to its origin. Where the chain dead-ends in a vendor blog citing another vendor blog, the page says so. That observation is usually worth more than the number.

What the evidence supports, what it does not, and where the field is running ahead of its data. Start with what GEO actually is or the tactic evidence scoreboard.

Crawlers, access controls, rendering and the plumbing that decides whether your pages are reachable at all. Begin with the AI bot registry or the technical audit.

A research programme is defined as much by what it refuses to produce as by what it ships. These are standing commitments, not preferences.

No composite visibility score. Blending citation rate, mention volume and sentiment into one number produces something that moves for reasons nobody can explain. A score you cannot decompose is not a measurement, and it is the single most common product in this category.

No statistic without a traced origin. If a figure cannot be followed to a named study with a disclosed method, it does not get a number on these pages. It gets a sentence explaining where the chain broke.

No first-party claim without first-party data. Every experiment described here is labelled with whether it has run. Where a study is pre-registered and unrun, the page says so in the opening lines, not in a footnote.

No prediction presented as a finding. Forecasts about where AI search is heading are cheap and unfalsifiable. Where this site reasons ahead of the evidence, that reasoning is tiered as a hypothesis and can be checked later.

No quiet corrections. When a figure here turns out to be wrong, the page is corrected and the change is stated. Pre-registering predictions guarantees some will fail in public, which is the price of writing them down first.

Most of the value of this programme is transferable. Five questions will tell you whether almost any AI-search figure means anything.

Who collected it, and what do they sell? Not a disqualifier. A vendor measuring the thing they sell access to has an interest in the direction of the answer, and you should know that before you quote it.

What was the query panel? A consumer panel and a professional panel produce different numbers from the same engine. The panel usually drives the result more than the engine does.

What is the denominator? A citation rate computed over queries where the engine never searched is diluted by definition. Most published rates never state the base.

Was variance measured? Ask the same engine the same question twice and the sources can change. Without a repeat run, part of any difference is noise being reported as a finding.

When was it collected? These products change without announcement. A figure with no date is not a statistic. It is a rumour with a decimal point.

The worked version of this is where AI SEO statistics come from, which traces the field's most-repeated figures back to their origins one at a time.

Two separate systems run across these pages. They answer different questions and are not interchangeable.

A grade describes where a number came from. A tier describes how much weight a claim can carry. A traceable number can still support only a hypothesis. The full method is in where AI SEO statistics come from.

Quote anything here, including the parts that are inconvenient. Two requests, both of which make the citation more useful to your own reader.

Carry the grade with the number. A figure graded partial is not the same as one graded traceable, and stripping the qualifier is how a soft number becomes a hard one across three reposts.

Carry the date. These surfaces change without announcement. An undated figure cannot be checked against anything, including itself a year later.