How long after a YouTube video is published does a generative search engine start citing it? This study measures time-to-first-citation for the PushLeads YouTube channel using twelve weeks of weekly AI answer snapshots, and documents a measurement artifact that more than doubled the apparent latency in the first pass. It is a companion to How to Measure YouTube Citations in AI Overviews, which covers the attribution method used here.
Headline result
For long-form videos published between October 23 and November 8, 2026, the median time from publication to the first observed citation in an AI answer was 8 days (mean 8.4, range 1 to 20). 39 of 73 videos (53%) had been cited by the November 15 snapshot. 18 (25%) were cited within 7 days and 34 (47%) within 14 days. Shorts published in the same window were cited once in 48.
Data
The source is a weekly audit table. Each row is one prompt run through one engine on one snapshot date, with the full list of cited URLs stored, not just domains. Engines over the study period were Google AI Overviews (from August 23), Google AI Mode (from September 11), Perplexity, ChatGPT and, from late October, Claude. Gemini appears in early snapshots only.
A citation is attributed to a video when a cited URL contains the ID of a video on the brand’s channel, using the matching rules in the companion piece (redirect decoding, watch and Shorts URL patterns). Publication dates come from the channel’s own video list. Videos deleted from the channel were excluded.
No new data was collected for this analysis. Every figure below is computed from snapshots that already existed.
Definitions
- First citation date: the earliest snapshot date on which any engine cited the video for any prompt.
- Time-to-citation: first citation date minus publication date, in days.
- Resolution: snapshots run weekly, mostly on Sundays, so time-to-citation is an upper bound with up to roughly 7 days of slack. A video cited “in 8 days” may have been citable on day 2.
The artifact: prompt-set expansion
The first pass covered every channel video published on or after the brand’s first snapshot: 506 videos with at least one subsequent snapshot. It found 170 cited videos with a median time-to-citation of 24 days and a mean of 28.6.
Those numbers were wrong in a specific way. Before October 23 the weekly prompt set for this brand held roughly 22 to 58 distinct prompts. On October 23 it grew to 157, and by November 15 it held 269, most of them written one per published video to test that video’s topic. Snapshots before October 23 found at most a few video-citing rows per week. The October 23 snapshot alone found 78.
The effect is a left-censoring problem in disguise. A video published in September whose topic was not in the prompt set could not be observed as cited until a matching prompt existed. When the expanded set went live, a large backlog of already-citable videos all received the same “first citation” date. Their measured latency reflects when the instrument started looking, not when the engine started citing.
The month-of-publication breakdown makes it visible. Among long-form videos at least 30 days old, the share cited within 30 days was 3% for August publications, 15% for September and 55% for October. Nothing about the videos changed that sharply. The measurement did.
Corrected cohort
To remove the artifact, the analysis was restricted to videos published on or after October 23, when the expanded prompt set was in place, and on or before November 8, so every video had at least one full weekly snapshot after publication. Results for that cohort:
| Group | Videos | Cited | Within 7 days | Within 14 days | Median days |
|---|---|---|---|---|---|
| Long-form | 73 | 39 | 18 | 34 | 8 |
| Shorts | 48 | 1 | 0 | 1 | 11 |
Mean cohort age at the November 15 snapshot was about 15 days, so the 53% cited figure is still right-censored. Some uncited videos will be cited later. The latency figures for cited videos are not affected by that, but the cited share is a floor.
By production format
Within the corrected long-form cohort, by production format:
| Format | Videos | Cited | Median days |
|---|---|---|---|
| Narrated slides, no on-screen presenter | 24 | 17 | 8 |
| AI presenter | 27 | 16 | 8 |
| Presenter over slides | 7 | 2 | 14.5 |
| Screen recording | 5 | 0 | n/a |
The cells are small and the formats were not randomly assigned to topics, so these are descriptive only. They do not support the idea that an on-screen presenter is required for citation.
Engine distribution
Counting every video ever observed as cited on this channel, Google AI Overviews cited 149 videos, Google AI Mode 98 and Perplexity 6. ChatGPT, Gemini and Claude cited none of the channel’s videos in any snapshot, although Claude answers did cite other YouTube content. Video citation in this data is almost entirely a Google phenomenon.
Prompt bias check
Because most prompts were written per video, one objection is that the result only measures whether a video surfaces for its own question. To test that, Google AI answers since October 23 were split by prompt type. For prompts written to match a video’s topic, 21.5% of answers (471 of 2,193) cited one of the brand’s videos. For general buyer prompts not tied to any video, the rate was 18% (30 of 167). The general set is small, but the rates are close, which suggests the videos are being retrieved for adjacent queries rather than only exact-match ones.
Limitations
- One brand, one niche (marketing for contractors and home-service businesses), one twelve-week window.
- Weekly sampling bounds latency resolution at about 7 days.
- AI answers are non-deterministic. A single weekly run per prompt and engine will miss some citations.
- The corrected cohort is 73 long-form videos. Format comparisons are small-sample.
- Prompt coverage still determines what can be observed. Latency for topics outside the prompt set is unmeasured.
Practical implications
- When measuring time-to-citation, freeze the prompt set or restrict analysis to content published after the set was last changed. Otherwise prompt additions masquerade as citation events.
- For Google’s AI surfaces, a long-form video that answers a specific question is typically picked up within one to three weekly cycles. A video uncited after about four weeks is a candidate for re-angling, not waiting.
- Shorts should not be counted on for citation in this vertical.
- Store full cited URLs. Domain-only tracking would have recorded none of these video citations as brand citations.