YouTube in LLMs: the study that shakes up certainties

A new study from Minddex has just shaken up the certainties of SEO on YouTube. Out of 55,631 quotes analyzed, AI models like ChatGPT, Perplexity or Gemini do not work at all like the YouTube algorithm. Views don’t count. Neither do subscribers. What matters is what the video says.

What Minddex measured, and what it means

Minddex is a platform specializing in AI visibility, also known as GEO (Generative Engine Optimization). In May 2026, the Minddex team published a study on how YouTube appears in responses from large language models. It follows an equivalent study on Reddit, which we relayed at the end of April.

The corpus is considerable: 57,871 raw YouTube quotesof which 55,631 could be enriched and analyzed, or 96.1% of the total. These quotes are extracted from 33,706 unique LLM responses, produced by ChatGPT, Perplexity and Gemini, on 526 B2B and B2C client projects spanning numerous industries. In total, 22,180 unique videos and over 12,000 distinct channels are represented.

The objective was the following: to understand what determines, concretely, that a YouTube video is cited in an AI response. The usual assumptions (views, subscribers, channel size) were tested statistically. The results almost systematically contradict intuitions from classic SEO.

Human audience does not predict LLM visibility

This is probably the most counterintuitive result of the study. Minddex calculated the correlation between audience metrics (number of views, number of channel subscribers) and the number of LLM citations a video received. The Spearman coefficients obtained are respectively +0.088 for views and +0.019 for subscribers. Both are well below the threshold for practical relevance set at 0.10.

Translation : a video with 50 views has as much chance of being cited by ChatGPT as a video with 50 million views, if its content precisely answers the question asked. The median observed is a single quote per video, in all audience segments without exception.

What LLMs read are the transcriptsnot the counters. They evaluate the semantic relevance of the content with regard to the user’s query. Human popularity does not enter into this logic.

This result has a direct implication for brands who have long thought that their AI visibility depended on their performance on the platform. This is not the case. The two systems operate according to distinct logics.

Brand channels capture almost nothing

Another striking figure in this study: 97.6% of LLM citations point to content produced by third parties. Independent creators come first with 36% of citations, followed by the media and press (32%) and individuals (30%). Official brand channels only capture 2.4% of citations.

The difference with Reddit is striking. On the latter, the third party/brand split was 78% versus 22%. On YouTube, it is 97.6% versus 2.4%. The ratio is almost binary.

Practical conclusion: publishing on your own YouTube channel is a complement, not a main lever for AI Search. To exist in the LLM answersyou must appear in third-party videos : those of creators, specialized journalists, users who talk about your products or your sector. The work of press relations, placement and seeding with creators becomes a structuring component of the GEO strategy, and no longer a peripheral option.

The format that works: between 5 and 15 minutes

The study also provides a precise answer to the question of format. 41.9% of LLM citations go to videos lasting between 5 and 15 minutes. This is approximately twice their actual weight in the YouTube catalog (where 5 to 10 minute videos represent approximately 19% of uploads according to Statista 2025).

In contrast, Shorts (less than 60 seconds) only capture 3.6% of quotes, while they represent 36% of daily uploads on the platform. The under-representation is a factor of 10.

The rule is simple: Shorts serve Discovery virality, according to a logic close to TikTok. The medium-long format, between 5 and 15 minutes, serves AI Search. A short video does not provide enough textual content for LLMs to extract a substantive answer to a specific query. A well-structured 8-minute video that covers a topic in depth provides an actionable transcript. This is what models are looking for.

For teams who have built their content strategy around Shorts in 2025-2026, this is something to quickly integrate: this format capitalizes for algorithmic discovery, not for visibility in AI responses.

The YouTube long tail is accessible

The study also brings good news for small chains. 71% of LLM citations go to channels with fewer than 100,000 subscribers. The dominant segment is that of channels between 10,000 and 100,000 subscribers, which concentrates 33.2% of citations. Better yet: very small channels with fewer than 1,000 subscribers account for 11.5% of the total.

Conversely, very large channels (more than a million subscribers) only represent 6% of citations. Certainly, channels exceeding 100,000 subscribers remain proportionally over-represented in relation to their weight in the overall catalog (they constitute approximately 1% of YouTube channels but collect 29% of LLM citations, i.e. an over-representation of a factor of 30). But access to LLM citations is not reserved for them. A recent, modest channel, which produces precise and relevant content, can be cited as much as an established channel.

This is a notable difference from classic YouTube SEO, where channel size and number of views create powerful leverage. In AI Search, the relevance of the content partly evens out these differences.

The evergreen effect is stronger on YouTube than on Reddit

The temporal dimension of the study reveals a third important lesson. 65% of videos cited by LLMs are more than a year old at the time of citation. On Reddit, that figure was 44%. The evergreen effect is therefore significantly more pronounced on YouTube.

Analysis of the age distribution of the videos cited shows that the most represented age groups are 1 to 3 years old (33.6%) and over 3 years old (31.5%). Very recent content (less than a week old) is almost absent. Videos less than three months old together represent only 6.8% of citations.

Minddex deduces that it is necessary between 12 and 36 months to a video to enter the pool of content taken up by LLMs. The strategic horizon is therefore 1 to 3 years, not quarterly. This calls into question the logic of steering by short-term KPIs, common in teams which measure the performance of each video in the weeks following its publication.

Producing long-lasting content, on stable and recurring subjects in its theme, is a decision that pays off in the long term. The view rate at D+30 is not the right indicator to anticipate visibility in AI Search in 24 months.

What it changes in practice

The Minddex study quite clearly draws the gap between historical YouTube SEO logic and what AI Search now requires. The two are not directly opposed, but their respective logic is different enough to justify a separate strategy.

Optimize a title and thumbnail, aim for virality, accumulate subscribers and post Shorts for reach: all of this remains relevant to the YouTube algorithm. It does not control visibility in ChatGPT or Perplexity.

What drives visibility there is the content spoken in the video, its precision, its clarity, its ability to answer a question that users ask of LLMs. It is also the presence in third-party content, among creators or media whose videos deal with your sector, your brand, or the subjects on which you want to exist in AI Search.

The good news for players who do not have significant resources: the entry ticket is not to have a large channel. It is well-targeted, well-structured content, and old enough to have been indexed and integrated into the training data or sources consulted in real time by the models.