Five social platforms appeared as sources in 25% of 360 sampled Google AI Overviews responses with citations

One in four responses with citations

Social platforms also appear among the sources readers can follow from an AI response. Of 360 Google AI Overviews responses with citations observed by Bee LLM Observatory, 90 cited at least one domain in this group: reddit.com, linkedin.com, facebook.com, instagram.com or tiktok.com. That represents 25% of the sampled responses with citations from this channel.

The denominator matters. This percentage does not describe every Google AI Overviews response or every question it receives. The segment covered 67 distinct prompts, repeated at different frequencies. Accounts also carried unequal weights. This is a finding about a monitored sample, rather than a representative estimate of general usage.

What a platform appearing as a source means

The measure asks a straightforward question: does a response contain at least one citation to a domain belonging to any of these five platforms? Subdomains are included. A response citing several platforms in the group still counts only once among the 90 responses meeting that condition.

The group consists of five domains specified in advance; it does not encompass all social media. The result also does not show which platform appeared most often, how many links each received, or what share of all citations belonged to the group. The 25% measures presence across responses, not a quarter of all cited links.

Presence does not establish endorsement or quality

For readers, the finding shows that a social platform can be part of the visible source trail in these responses. However, the aggregate does not identify which statement each citation accompanied, who published the material, or whether the response endorsed, challenged or simply mentioned it.

These figures therefore cannot establish the accuracy of the responses or the reliability of the linked posts. A cautious interpretation is that social sources appeared with some frequency in this observed segment. Concluding that they shaped the AI's conclusions would require information that this measure does not supply.

Equal prompt weights preserve the broad finding

The primary rate of 25% gives each observed response equal weight. A prompt run more frequently can therefore contribute more to the result than one run less often. As a sensitivity check, a calculation giving every prompt equal weight puts the group's presence at 23.66%.

That is a reduction of 1.34 percentage points, preserving the broad finding of a presence close to one in four responses with citations. This check helps show how much the result changes when prompt weights change. It does not equalize account weights or make the sample representative.

Other channels provide context, not a ranking

The other published segments recorded 14 of 576 responses with citations for OpenAI API, or 2.43%; 5 of 184 for Perplexity API, or 2.72%; and 13 of 197 for Google AI Mode, or 6.6%. Their rates with equal weight per prompt were 2.27%, 2.43% and 5.8%, respectively.

These are descriptive differences between samples whose prompts and collection methods may differ. They are not a controlled test using matched questions and cannot rank engines by their preference for social sources. OpenAI API and Perplexity API identify API collection channels: their figures should not be transferred to the consumer applications.

How the observation was defined

This article is published on September 24, 2026. The observation covers seven complete days in the Europe/Madrid time zone: from September 16 at 00:00, inclusive, to September 23 at 00:00, exclusive. Selection was limited to completed executions, non-branded prompts and accounts whose email addresses did not contain “demo.”

The latest response for each execution and stored model was retained, requiring nonempty text without an error. Only citation records with a known collection method were used; merely detected links were not. Each public domain, including its subdomains, counted at most once per response.

The scope of the published evidence

Of 3,070 valid responses from known channels, 35 lacked citation telemetry. This left 3,035 measured responses, of which 2,698 had citations. The set with citations covered 154 prompts, 27 projects and 24 accounts. Responses without valid citations or without citation records were excluded from the headline denominator.

Publishing a segment required at least 100 responses, 10 prompts, three projects and three accounts; matching responses had to come from at least three accounts. An absent or suppressed channel does not mean zero appearances. These filters support a description of the sample, but they do not remove selection limitations or support conclusions about market share.