Lesson 3 of 8
How AI engines decide what to cite
Why does an AI answer name some businesses and not others?
Two findings from our own corpus of 12,778 AI answers. First, citation is heavily concentrated: 8,577 different domains were cited across 109,674 citation events, but the distribution is extremely top-heavy, with a Gini coefficient of 0.8257 and the ten most-cited domains taking 16.16% of all citations. Second, the same question asked twice does not reliably produce the same sources. Both facts should change what you expect from anyone selling you visibility in AI answers.
AI answers analysed
measured
Citation events recorded
measured
Distinct domains cited
measured
Concentration (Gini)
measured
Share taken by the top 10 domains
measured
This is the lesson where we can show our work, so it is worth saying exactly what the work was. Over 164 days we asked five AI engines the same questions repeatedly and stored every answer along with the sources it named. That produced 12,778 answers containing 109,674 citation events across 8,577 distinct domains.
The first finding is concentration. A Gini coefficient of 0.8257 is a measure of inequality: 0 would mean every cited domain is cited equally often, and 1 would mean a single domain takes everything. At 0.8257 the citations are concentrated in a small number of sources, and the ten most-cited domains alone account for 16.16% of every citation we recorded. There is a long tail of thousands of sites, but it is a thin one.
The practical reading of that number is not despair, it is expectation setting. Being cited is not a matter of ticking a box that makes you eligible. You are competing for a slot in a distribution that is already top-heavy, and the sites at the top are mostly the ones that were already established references in their subject.
The second finding matters more for how you judge reports. Ask an engine the same question twice and you frequently get a different set of sources. Not a slightly reordered set: a different one. This is not a malfunction, it is how these systems work. It has a direct consequence that almost nobody accounts for: a screenshot of an AI answer naming your business is one sample of an unstable process. It is real, and it is nearly worthless as evidence on its own.
That is the discipline this lesson is arguing for. If a single reading can come out differently next time, then a single reading cannot support a claim about a trend. Measuring this surface honestly means asking repeatedly and reporting the spread, which is exactly what most reporting on AI visibility does not do.
Takeaway: One screenshot is one sample. Ask how many times they looked.