What kind of YouTube videos does AI pull into its answers?

YouTube ranking

Do AI engines pick the videos that rank highest on YouTube?

Mostly not. We searched YouTube with each prompt and saved the 20 highest-ranking videos. Only 17% of the videos an engine pulled for a prompt rank in YouTube’s top 20 search results for that prompt, and 5% rank in the top 3. Even counting the top 20 results for any of our 1,209 prompts, only 31% of the pairs turn up. Most of the videos AI engines use are ones YouTube’s own search doesn’t rank for the question at hand.

Perplexity and ChatGPT stay closest to YouTube’s ranking, with 29% and 26% of their picks ranking in the top 20. Google’s two engines are further away: 21% for AI Overview and 15% for AI Mode.

When an engine’s pick does rank in YouTube’s top 20, it tends to be the stronger video on the page. In the 494 prompts where that happens, the picked videos have a median of 13,100 views, against 6,700 for the other videos in the same results, and they rank sixth on average rather than twelfth.

Views and channels

Do AI engines prefer popular videos or big channels?

Big channels, not popular videos. To see what sets the AI picks apart, we compared them with 1,998 other videos from YouTube’s top 20 search results for the same prompts, videos the engines could have picked but didn’t. The median video AI engines pulled has 4,700 views, the same as the other videos in YouTube’s search results. Six in ten have under 10,000 views. Viral videos are, if anything, less common: 1% of AI picks have over a million views, against 3% of the other videos.

The channels behind them are another story. The median AI pick comes from a channel with 21,500 subscribers, more than four times the 4,900 of the other videos. Very small channels are where the gap shows most: 11% of AI picks come from a channel with under 1,000 subscribers, compared with 31% of the other videos. Channels with 10,000 to a million subscribers make up 56% of AI picks and 37% of the rest. The very largest channels, with a million or more, are equally common in both groups (4%).

Freshness

How old are the videos?

Older than you might think, but newer than the other videos. The median video was 14 months old the first time an engine pulled it in our data. The other videos in YouTube’s search results had a median age of 26 months. Videos from the past year make up 45% of AI picks and 33% of the rest, while videos older than five years are half as common among AI picks (13% vs 25%).

Very new videos are not favoured. 8% of AI picks were under a month old when first pulled, a little below the 12% of the other videos.

A year in the title usually means a recent video. 15% of AI picks have a year in the title, against 12% of the other videos. In 80% of cases the year matches the upload year, and in another 17% it is the following year. Videos with a year in the title were a median of 6.5 months old when first pulled, compared with 16 months for videos without one.

Length and format

How long are the videos, and do Shorts get picked?

A little longer than average. The median AI pick runs 8 minutes, against 6 minutes for the other videos. Videos of 8 to 30 minutes make up 44% of AI picks and 35% of the rest, while videos of three minutes or less are less common (25% vs 32%).

Shorts are a Google thing. They make up 21% of YouTube sightings in AI Overview and 17% in AI Mode, but only 7% in ChatGPT and 2% in Perplexity.

Timestamps

Where in the video do AI engines link?

Google’s engines often link to a specific moment rather than the whole video. 39% of YouTube sightings in AI Overview and 30% in AI Mode come with a timestamp. ChatGPT and Perplexity almost never do this.

The moments are mostly early. 39% of timestamps point into the first tenth of the video, 27% into the first 30 seconds and half into the first two minutes. The median link starts 1 minute 43 seconds in.

Chapters go with timestamp links, but don’t decide where they land. Videos with chapters in the description make up 41% of Google’s timestamped sightings, against 24% of sightings without a timestamp. Yet only 12% of the timestamps on those videos start exactly at a chapter. The engines seem to pick the moment from the video itself, not from the chapter list.

Titles

Do titles make a difference?

Less than you might expect. AI picks and the other videos have titles of the same length: a median of 57 characters and 10 words. Most title styles are about as common in both groups.

A few are a little more common among AI picks: titles with “best” or “top” (19% vs 17%), how-to and guide titles (20% vs 17%), titles with a year (15% vs 12%) and comparisons (5% vs 3%). These are the kinds of titles that answer a question directly. Separators such as “|” or “-” are a little less common among AI picks (29% vs 33%). None of the differences is large, so a title alone is unlikely to explain why a video is picked.

Descriptions and subtitles

Do descriptions and subtitles matter?

Descriptions a little, subtitles more. AI picks have slightly longer descriptions: a median of 1,164 characters (158 words) against 1,051. Short descriptions under 500 characters are less common among AI picks (22% including empty ones, vs 27%). 28% of AI picks list chapters in the description, against 23% of the other videos.

The clearest difference is subtitles. Only 4% of AI picks have no subtitles at all, against 15% of the other videos. Most of the subtitles are YouTube’s automatic ones. Videos with subtitles added by the creator are about as common in both groups (15% vs 14%). That fits with engines that read what is said in a video: a video without a text version gives them less to work with.

AI picks also have a few more tags (a median of 10 vs 8) and a higher like rate (1.5% of views vs 1.0%).

Engines

Do the engines pick the same kind of video?

Broadly yes, with a few differences. Google’s two engines pull far more YouTube videos than the others and account for 90% of the sightings. They are also the ones that use Shorts and timestamps. ChatGPT picks the biggest videos: a median of 15,500 views from channels with 39,100 subscribers, and the newest, at a median of 8 months old. Perplexity picks the oldest, at a median of 20 months.

Medians per video. Age is the video’s age the first time the engine pulled it. Shorts and timestamp links are shares of sightings; top 20 is the share of prompt-video pairs that rank in YouTube’s top 20 search results for the prompt. ChatGPT pulled few YouTube videos, so its figures are less certain.

Repeats

Do the same videos keep coming back?

A small group does. 62% of the videos were pulled only once in three months, but the top 10% of videos account for half of all sightings. 13% were pulled five times or more, and two thirds of those for more than one prompt.

Videos rarely travel far. 17% were pulled for more than one prompt, 16% by more than one engine and only 3% for more than one brand’s prompts.

What it means

What should brands and creators take from this?

Methodology

How we did it

Data

YouTube videos among the sources of answers from ChatGPT, Perplexity, Google AI Overview and Google AI Mode to 1,209 prompts tracked in Vercite, collected 5 August to 3 November 2026. 10,426 sightings of 3,728 videos with a known video ID, from the prompts of 54 brands.

Sightings

One video in the sources of one AI answer. If the same video is pulled into answers on ten different days, or by two engines, that is ten or two sightings. We count every video in the sources, whether or not the answer text refers to it. Some sightings from September and October could not be matched to a video ID and are left out.

Video data

Views, likes, subscribers, upload date, length, title, description, tags and subtitles were collected from YouTube’s public video pages on 3 November 2026. View and subscriber counts are as of that day, not the day of the sighting. 28 videos had been removed or made private and are left out.

YouTube ranking

We searched YouTube with the wording of each prompt, in the prompt’s country and language, and saved the 20 highest-ranking videos in the search results (1,362 prompts, 3 November 2026). Prompts are written as questions to an AI, so they are longer than typical YouTube searches.

Comparison group

1,998 videos drawn at random from YouTube’s top 20 search results for the 494 prompts where at least one AI pick also ranks in the top 20, excluding every video any engine pulled. This keeps the comparison to the same topics.

Age

For AI picks, age is measured on the day an engine first pulled the video in our data. For the comparison group, age is measured on 3 November 2026.

Prompt mix

The prompts come from brands tracked in Vercite, in English, Swedish, Portuguese, Norwegian, Dutch, Finnish and Romanian. A few brands account for many of the videos, so we also checked the results with every brand counting equally. The shares moved by a few points at most and the patterns stayed the same. The results describe this set of prompts, not all AI use.

Timestamps

A timestamp link is a YouTube link with a start time. Position is the start time divided by the video’s length. Chapters are counted when a description lists at least three times in the form 0:00.

Titles

Features were matched with simple word and character patterns. Words such as “best” and “how to” were matched in English, Swedish, Portuguese and Norwegian.

Causes

This is a comparison, not an experiment. Channel size, length, subtitles and the rest go together with being picked; we can’t say which of them, if any, makes an engine pick a video.