YouTube appeared in 37.54% of 349 sampled Google AI Overviews responses with citations

A visible presence in a defined sample

YouTube was a cited domain in 131 of the 349 observed Google AI Overviews responses containing at least one valid citation: 37.54%. That is the main finding of this Bee LLM Observatory article, published on September 10, 2026. The observation period covers seven complete days, September 3 through September 9, in Madrid local time.

The percentage applies only to responses with citations in this sample. It does not represent all Google AI Overviews answers or all conversations with artificial intelligence. The segment covered 60 prompts, or queries, with repetition and unequal weights across queries and accounts. Those characteristics limit how far the finding can be generalized.

What a YouTube citation means here

The measure counts responses containing at least one valid citation to youtube.com or its subdomains. A domain counts only once in each response. An answer with several YouTube links therefore contributes just one occurrence to the total. The 37.54% figure is the share of responses with citations in which that domain appeared.

It is not the percentage of all links that led to YouTube. Nor does it reveal how many people opened a link, watched a video or found the source useful. The aggregate evidence cannot establish whether YouTube was an answer’s main source, what information it supported or whether the citation was accurate.

Equal weight for each query barely changes the finding

The primary measure gives every observed response equal weight. A query run many times consequently has more influence on the result than one run only a few times. That accurately describes the collected responses, but it is different from a sample in which every query has equal importance.

The available sensitivity check gives each prompt equal weight. Under that calculation, YouTube’s presence in Google AI Overviews moves from 37.54% to 37.94%, a difference of 0.40 percentage points. The interpretation barely changes: the domain appears in just over a third of responses with citations under both approaches. Stability under this check does not remove possible effects from the selection of queries, accounts or collection methods.

The other published figure describes a different set

In the segment labeled Gemini, YouTube appeared in 10 of 533 responses with citations, or 1.88%, across 104 prompts. Giving each query equal weight produced a figure of 2.01%. Here too, the adjustment preserves the descriptive reading: YouTube had a limited presence in that observed set.

The gap between these percentages is not a controlled comparison of engines. The sets contain different queries and may differ in account weights and collection methods. Engine labels identify collection channels; an API channel is not the consumer app. These figures cannot rank product quality, establish universal source preferences or measure market share.

What readers can take from the result

The finding supports a specific reading: for this group of queries during this week, YouTube was a visible citation destination in the observed Google AI Overviews responses. Its appearance shows that a video platform can feature among the sources linked in an AI answer. The aggregates do not explain why it appeared in those instances.

The observed difference makes the context of an AI statistic worth examining: which questions were collected, through which channel and using which denominator. It does not demonstrate that publishing videos increases the likelihood of being cited. It also cannot identify topics, formats or editorial practices that explain the result.

How the denominator was built

Processing began with 3,287 raw records, with no duplicates removed. There were 2,870 valid responses from known engines; 31 lacked citation telemetry, leaving 2,839 measured responses. Of those, 2,478 contained at least one valid citation. They form the overall coverage used to study cited domains: 126 prompts, 17 projects and 15 accounts.

The process selected completed executions and retained the latest response for each execution and stored model, requiring nonempty text without errors. Branded prompts and accounts with “demo” in their email address were excluded. Only URLs recorded as citations with a known citation method were used, rather than links merely detected in the text. Empty or missing citation records were excluded from the headline denominator.

The boundaries of the published evidence

Published segments must contain at least 100 responses, 10 prompts, three projects and three accounts; domain occurrences must span at least three accounts. Only predeclared public platform domains are exported. An absent or suppressed engine must not be read as a zero, and even zero observed occurrences would not establish general absence.

The Bee LLM monitored sample does not represent all AI activity. Publication thresholds do not correct unequal weights or query selection. No personal, customer, project, individual query or individual response data is published. The publication date does not extend the observation period either: the analysis stops at the start of September 10 in Madrid.