One cited host appeared in 76 of 409 sampled OpenAI API responses with citations

A concentrated set of citations, within a defined sample

In the Bee LLM Observatory sample, 76 of 409 responses with citations collected through OpenAI API cited exactly one distinct hostname: 18.58%. That is slightly fewer than one in five responses in this group. The percentage applies only to observed responses containing valid citations, rather than every response generated through that channel.

The main limitation belongs alongside the finding: questions and accounts do not carry equal weight in this collection. When each question receives equal weight, the proportion falls to 15.25%. Citations concentrated on one host remain part of the picture, but their measured frequency depends on how the sample is weighted.

What citing one host actually means

A host is the name identifying a web location within an address. This measure counts the distinct hostnames appearing in a response's citations. Several references to different pages on the same host still count as one. Subdomains can have different hostnames.

The 76 cases therefore do not necessarily contain just one URL, one document or one citation. Nor does the count establish that a single organization stands behind the references. The observation is narrower: all the valid citations counted in each of those responses shared the same hostname.

Why equal weighting changes the result

The primary measure is 18.58% because it counts each response once. Under that calculation, a question appearing repeatedly can influence the result more than a question collected only a few times. The OpenAI API segment contains 120 distinct questions across its 409 responses with citations.

The alternative calculation gives each question equal weight and produces 15.25%, a decrease of 3.33 percentage points. This indicates that repeated questions affect the aggregate result. It does not identify which questions account for the difference or why they were repeated. Equal weighting also does not turn this collection into a representative estimate of what the wider population asks.

The other segment provides context, not a ranking

For ChatGPT (web), 10 of 113 responses with citations met the same condition, or 8.85%. That segment contains 62 distinct questions. Giving each question equal weight produced 8.87%, almost unchanged from the response-weighted percentage.

These figures describe two different segments; they are not a controlled comparison. Questions, account weights and collection methods may differ. OpenAI API identifies a collection channel and is not equivalent to the consumer app. These data cannot establish a quality ranking among engines or measure market share.

What citation concentration can tell us

The finding describes how the references accompanying responses are distributed: in part of this sample, they all point to the same host. That is an observable property of the emitted citations. It does not establish that the model consulted, accessed, reviewed or verified the pages mentioned.

As an interpretation, a concentrated reference list can help frame questions about the diversity of the visible supporting material. However, this count does not measure editorial independence, accuracy or adequacy. Multiple hosts do not guarantee different perspectives, and a single host does not establish that an answer is poorly supported. Those questions would require additional evidence.

How the observation was defined

This article is published on October 15, 2026. The observation covers seven complete local days: from October 7 at 00:00 up to, but not including, October 14 at 00:00 in the Europe/Madrid time zone. The publication date is outside the observation period.

The collection included only completed executions, non-branded questions and accounts whose email addresses did not contain “demo.” It retained the latest response per execution and stored model, requiring nonempty text without an error. Citations came from the reference record with a known collection method; addresses merely detected in the text were not used.

Coverage and the limits of the count

The collection began with 1,656 records, with no duplicates removed. There were 1,369 valid responses from known engines; 15 lacked citation telemetry, leaving 1,354 measured responses. Of these, 1,045 contained citations and make up the overall coverage: 125 questions, 27 projects and 25 accounts. That coverage extends beyond the two segments shown.

Responses without valid citations and missing citation records are excluded from the main denominator. Reporting a segment required at least 100 responses, 10 questions, three projects and three accounts; cases meeting the measure had to span at least three accounts. An absent or suppressed engine does not mean zero cases. This monitored collection does not represent all AI conversations, and no personal data or individual customer, project, question or response content is published.