How to read AI visibility with sample size, intervals, and snapshots

One answer can contain three different events. The brand name appears in a sentence: that is a mention. A URL in the answer points at your site: that is a citation. Your name and a competitor’s name appear in the same answer: that is co-occurrence. All three can happen together, or only one of them can.

The OpenAI web search guide separates inline citations from the sources list, and says the sources list is often longer than the citations. In the Perplexity Sonar schema, citations may be null. A missing URL list means this response did not include that list. It is not, by itself, proof that the mention rate is zero.

A proportion is a count of events divided by a sample size. When the sample size is 1, you have one observation. You do not have a percentage to report.

NIST/SEMATECH e-Handbook section 7.2.4.1 discusses confidence intervals for a proportion. It says the Wilson method comes from Wilson (1927) and was recommended by Brown, Cai, and DasGupta (2001) and by Agresti and Coull (1998). The handbook says one advantage is that the lower limit cannot be negative, which the usual normal approximation can produce. For a very small sample or very few events, it also shows an interval based on the binomial distribution, and notes that the interval is not symmetric.

This note does not compute an interval for any brand. A percentage with no sample size and no interval is not the kind of estimate that section describes.

The GEO paper defines visibility with citation word count and position, and compares pages before and after edits inside its own experiment. That definition can be cited. It cannot be relabeled as an official ChatGPT share. Neither OpenAI’s search post nor the API guide uses that metric.

• The calendar date
• The full question
• The language of the question
• The country or location value actually sent, if that call had the field
• The product surface, or the API model and tool name
• The raw answer and the URLs it actually returned

The OpenAI guide says search_context_size does not guarantee a number of sources or citations. The snapshot should keep the citations that came back, not only the name of the setting. Perplexity’s disable_search turns web search off. That response does not belong in the same citation rate as a response that searched.

A later run is a new snapshot. If the page changed between two snapshots, the older answer is not a result for the newer page. Snapshots in different languages or countries also do not add up to one number, because the API fields checked in the earlier notes are separate.