US market research

How Often Are Sources Referenced by AI Cited? 14.2%

Published

This article was first published in Korean by Intrix Lab on September 13, 2026.

Of the sources AI referenced during web searches, 14.2% were cited in its answers. The remaining 85.8% entered the reference set but were not selected as answer links. These measurements were collected on 2026-08-31 from 432 calls: 144 queries across 6 industries, each run 3 times on a single model, Model A. Being read during a search and being cited in an answer should not be treated as the same metric.

How did we distinguish references from citations?

We separately counted sources AI read during its search process (references) and sources it linked to in its answers (citations). We treated entry into the reference set as discovery and adoption as an answer link as selection. All 1,764 instances of domains cited in answers were within the reference set. There were 0 citations outside that set.

What does 14.2% mean?

This result comes from repeatedly asking Model A 144 queries across 6 industries. Among responses that actually performed a search, the median number of sources referenced per response was 31, and the median number ultimately cited was 4. Of the referenced sources, 14.2% were cited in the answers, with a 95% confidence interval of 13.5% to 15.0%.

Why is looking only at the first search insufficient?

Model A performed multiple searches to produce a single answer. Of the referenced domains, 48.6% first appeared in the second or later search actions. The 95% confidence interval for this proportion was 46.4% to 50.6%.

This means that looking only at the first search results misses roughly half the candidates AI actually reviewed. Looking only at final citations also hides the 85.8% that were not selected. To understand where a source dropped out, you need to examine both stages.

What are the limits of these findings?

The population consists of 144 queries across 6 industries and a single model, Model A, collected on one day, 2026-08-31. These values came from the API's web search interface, so we cannot say that the 14.2% rate would recur in the same model's web app or in other models. With only 3 runs per query, the data is also too limited to assess individual sources' performance.

This analysis measured the observed rate between referencing and citation. It did not test which page elements led to selection. This single figure also cannot distinguish why a source was not referenced from why it was referenced but not selected.

The 95% confidence interval represents uncertainty from variation in the query sample within this question set. It does not include uncertainty from changes in industry, model, or timing. There is no guarantee that the same result will occur elsewhere.

What can sites in similar circumstances take from this?

Sites measuring AI citations should first separate entry into the reference set from final citation. An absence of final citations alone does not mean a site was not discovered during search. Conversely, an increase in references alone does not establish an increase in answer visibility.

This is the extent of what the data supports. Before deciding on an intervention, first identify whether discovery or selection is the bottleneck.

Intrix's free SEO and GEO audit is a starting point for checking a site's search and crawling conditions. AI references and citations require separate, repeated measurement using a fixed question set. Intrix does not guarantee specific rankings, citations, or visibility.

Frequently asked questions (FAQ)

Q. Does being included among referenced sources mean being cited in an AI answer?
No. In this measurement, only 14.2% of referenced sources were cited in the final answers. Discovery and selection must be counted separately.

Q. Does the uncited 85.8% represent low-quality sources?
We cannot draw that conclusion. This analysis measured only whether sources were selected; it did not test the page characteristics that led to selection. Factors such as relevance, source type, and answer composition may have acted together.

Q. Should we assume our site's citation conversion rate is also 14.2%?
No. This value is limited to API measurements of a single model, Model A, using 144 queries across 6 industries on one day, 2026-08-31. Assess your site using repeated measurements with the same fixed question set.

← Back to insights