If you already sponsor creators, save the AI answers that cite their videos. A source link is a useful place to start reading. Counting it as a recommendation leaves too much of the answer unchecked.
What the YouTube study measured
Scrunch’s September 10 study used answers collected from May through July 2026. Its main comparison held the question and assistant fixed within answers that cited YouTube. When a cited video’s description named a sponsor, that brand appeared in 40.0% of answers. With no cited video naming that sponsor, it appeared in 32.3%. The difference was 7.7 percentage points.
The outcome was a brand name in the answer. The researchers identified sponsorships from descriptions, excluding audio. Their observational design could not separate payment from prominent brand placement; mention rates were also rising before the first observed citation. It does not establish a sales return or a recommendation effect. Michael Iannelli’s account of the study explains the comparison.
Those dates matter. Scrunch later reported a sharp change in Perplexity’s YouTube citation behavior. A campaign report should describe the answers actually observed during its own reporting window.
A brand’s site can appear while its name disappears
Our October 6 cleanser pilot gives us a concrete reason to keep these fields separate. We asked, what’s the best facial cleanser for sensitive skin
. Every one of the 32 saved Google AI Overviews citation lists included cerave.com. None of the answer texts named CeraVe.
Four distinct answer texts across 32 records; repeated wording limits independence.
Perplexity supplied another distinction. CeraVe appeared under Good alternatives
in 27 of its 28 saved answers. We read that placement as an alternative recommendation. That is an editorial interpretation of the wording, separate from the stored automated score and not independently validated.
This cleanser pilot measured no sponsorship treatment. It supplies a separate example of how citations and brand outcomes can come apart. The exports also lack the original inline citation anchors, so a source link cannot tell us which sentence it supported.
Put the distinction in the report
Our practical recommendation is to keep the source link next to the answer and write down what happened to the brand. A reviewer should be able to open the record and check the label.
- Fix the unit of comparison. Keep the exact prompt and assistant with each dated observation. Report how many answers you inspected. Mark repeated text so a recurring answer does not look like fresh agreement.
- Save the source and the response together. Retain the cited URL, whether the brand is named, and the relevant wording. Preserve inline citation anchors when available. A domain appearing in a source list does not establish which claim it supports.
- Describe the recommendation. Distinguish a primary choice from an alternative. An incidental reference can remain a mention; unclear wording should stay unresolved. Keep an automated label separate from a reviewer’s interpretation.
- Make every rate readable. Show both the count and its denominator.
Brand named in answers citing the video
uses a different denominator fromvideo cited in all observed answers
. Track clicks or purchases separately if those are the campaign outcomes.
And keep the old records when you check again. If the source disappears next month, the earlier answer still matters to the audit. If the brand appears more often, read its role before calling the change a win.
Evidence and counting notes
The external study and Gist’s pilot are separate datasets. Gist’s numbers above reuse two completed native Deep runs from the October 6 two-question pilot; no new collection was run for this article. They describe one exact cleanser question in one batch, without a controlled country setting. They are not category rankings or product advice.
Download the 60 CeraVe observation rows (CSV). The file counts cerave.com and its subdomains once per answer, searches the answer text for the CeraVe name, and records its placement under the alternatives heading. It includes stored response hashes so the rows can be matched to the original 128-row table.
How the editorial labels differ from automated scores
The 27 Perplexity alternatives are assistant editorial annotations of the saved wording. The stored automated recommendation field was positive in four of the 28 answers. Both are retained in the CSV; neither is silently substituted for the other. Google did not discover CeraVe as an answer entity, so its native recommendation field is blank rather than an invented zero. A second-reviewer agreement rate has not been established.
Gist sells AI-visibility services. The checklist is our editorial interpretation of these observations. It does not promise a sponsorship return or attribute an answer change to a particular source.