Which YouTube videos does Google's AI Overview cite? We checked 1,777 how-to answers

Published 29 September 2026. Data collected 22 September 2026. Last updated 29 September 2026. [TODO: add named author and credentials to the byline.]

Key findings

In a snapshot of 1,777 Google AI Overviews for how-to questions, 83% cited at least one YouTube video. The cited videos had about a fifth of the views of the videos YouTube ranks first for the same questions, although they were also newer. When a citation pointed to a moment, that moment usually came early in the video.

• 83% of AI Overviews cited YouTube: 4,902 citations to 1,460 different videos.
• 36% of citations linked to an exact second of the video.
• 78% of question and video pairs were cited in all three runs, so the pattern is stable.
• Cited videos were newer than YouTube's top results in 445 of 506 questions, so age explains part of the view gap.

This is one snapshot. It shows what gets cited, not what makes Google cite it. Each finding below comes with a bet you can test and the catch that stops it being a rule.

How often do AI Overviews cite YouTube for how-to questions?

Most of the time: 83% of the 1,777 AI Overviews cited at least one YouTube video. Across all of them, we logged 4,902 YouTube citations pointing to 1,460 different videos, so the same videos recur but the pool is wide.

That figure is a share of AI Overviews, not a share of citations, so it is not directly comparable with industry numbers. Ahrefs, for example, reports YouTube taking the largest slice of AI Overview citations across a broad set of US queries, at 22.9% in its September 2026 ranking (https://ahrefs.com/blog/most-cited-domains-ai-overviews/). Our set is narrower (how-to questions only), which is a likely reason YouTube appears so often. The direction matches other work: OtterlyAI found that more than half of the social media citations in Google's AI Overviews and AI Mode come from YouTube, according to its 2026 study (https://otterly.ai/blog/youtube-ai-citation-study-2026/).

[TODO: insert Figure 1, the bar chart of the four headline rates. Caption: "Figure 1. Four rates from the study. AI Overviews citing YouTube 83%, citations linking to an exact second 36%, pairs cited in all three runs 78%, questions where the cited video was newer 88% (445 of 506). Source: Dodox search data, 22 September 2026."]

Are the videos Google cites the most-viewed ones?

No. The cited videos had about a fifth of the views of the videos YouTube ranks first for the same questions. View count does not look like the gate to a citation.

The catch is age. The cited videos were newer than YouTube's top result in 445 of 506 questions (88%), and newer videos have had less time to collect views. Some of the gap is simply time. [TODO: add median views, cited vs top-ranked, plus views per day since upload. That comparison removes the age effect and is the number that decides whether the headline holds.]

The result is at least consistent with a wider shift away from rankings. An Ahrefs analysis summarised by LoudScale (https://loudscale.com/blog/ai-overviews-sources/) reports that only 38% of AI Overview citations now come from pages in Google's top 10, down from 76% in July 2025. That figure covers web pages, not video, so treat it as context, not confirmation.

Which second of the video does Google link to?

The cited moment usually comes early in the video [TODO: median timestamp X seconds; Y% within the first 30 seconds]. Only 36% of citations carried a timestamp at all, so this covers the linked subset.

A parallel pattern shows up in text. CXL's analysis of 100 AI Overview citations found 55% came from the first 30% of the cited page, per its AEO guide (https://cxl.com/blog/answer-engine-optimization-aeo-the-comprehensive-guide/). That sample is small, but it points the same way.

Bet: give the answer in the first 30 seconds. Catch: how-to videos may put their answer early anyway, so early moments may be cited only because that is where the answers are.

What does Google quote: the creator's words or the description?

The cited lines matched what the creator says out loud, not the video description.

Bet: say the words people search for, out loud. Catch: the moment that answers a question naturally uses the question's words, so the words may only mark where the answer is.

What kind of line gets cited?

The cited line often sets up the answer instead of delivering it. An AI model classified the lines [TODO: add the share for each category].

Bet: say what you are about to show, right before you show it. Catch: the classifier can be wrong, and we spot-checked [TODO: N lines by hand, with an error rate of X%].

What should creators test?

Each bet has a confound. Each one below is paired with the reason it might be an illusion and a test that could separate the two explanations.

Bet: give the answer in the first 30 seconds

• Why it might be an illusion: answers may simply sit early in how-to videos.
• Test that would settle it: compare against the timestamp of the answer in videos that were not cited.

Bet: say the search words aloud

• Why it might be an illusion: answers naturally repeat the question's words.
• Test that would settle it: compare cited lines with equally relevant uncited lines from the same video.

Bet: announce what you are about to show

• Why it might be an illusion: the AI classification may be wrong.
• Test that would settle it: hand-label a random sample and report the error rate.

Bet: a small channel can win a specific question

• Why it might be an illusion: cited videos are newer, so they have fewer views.
• Test that would settle it: compare views per day since upload.

None of these bets costs anything if it is wrong, since each is also good how-to teaching.

How we ran the study

• Questions: 600 how-to questions across 12 topics [TODO: list the 12 topics], sourced from [TODO: where the questions came from].
• Searches: each question searched on Google three times on 22 September 2026, using [TODO: device, country, language, logged-out or logged-in]. That returned 1,777 AI Overviews.
• Citations: 4,902 YouTube citations; a timestamp was recorded where the link pointed to a specific second.
• Comparison set: for the view comparison, cited videos were compared with the video YouTube ranks first for the same question, across 506 questions [TODO: explain why 506 of 600]. Views were read on [TODO: date].
• Line classification: an AI model labelled the cited lines [TODO: model and criteria].
• Data: [TODO: link to the question list and results file].

What this study cannot tell you

It is correlation, not cause: it shows what cited videos look like, not what makes Google cite them. It is one day, and AI Overviews change often. It covers how-to questions only, and it covers Google only. A 2026 NJIT study of 14,212 queries, summarised in AirOps' AEO guide (https://www.airops.com/blog/aeo-answer-engine-optimization), found that Google Search, AI Overviews and Gemini retrieve sources that overlap very little, so a video cited here may be absent elsewhere.

Re-run planned: [TODO: date]. This page will be updated with the new data.

Sources

1. Dodox search data, 22 September 2026 [TODO: dataset link].
2. Ahrefs, The 50 Most-Cited Websites in Google AI Overviews (https://ahrefs.com/blog/most-cited-domains-ai-overviews/), September 2026.
3. OtterlyAI, YouTube AI Citation Study 2026 (https://otterly.ai/blog/youtube-ai-citation-study-2026/), May 2026.
4. LoudScale, AI Overviews Sources (https://loudscale.com/blog/ai-overviews-sources/) (summarising Ahrefs' analysis of 863,000 SERPs), May 2026.
5. CXL, Answer Engine Optimization guide (https://cxl.com/blog/answer-engine-optimization-aeo-the-comprehensive-guide/), August 2026.
6. AirOps, AEO guide (https://www.airops.com/blog/aeo-answer-engine-optimization), citing an NJIT study presented at SIGIR.

Third-party figures come from vendor blogs with differing methods, and we have not independently verified them. They are context, not proof.