Your AI citations are a setting on someone else's engine.

On September 13, Google AI Mode sent 11.0% of its citations to YouTube. On September 14 it was 6.8%. On September 15 it was 3.1%, and it sat between 3.4% and 3.9% from then until the figures were published a week later. That is a drop of about 72% in two days. I do not think it tells you anything about YouTube.

The number describes the engine, not the publisher

When a share moves that far in two days, the explanation sits with the engine or with whoever is measuring it. The one explanation I rule out is that the videos got worse between Sunday and Tuesday.

Which sources an AI answer borrows from is decided inside a system you cannot see, by people who do not owe you notice. You can lose a citation share you did nothing to earn, on a day you did nothing to lose it.

Google says as much in its own documentation: 'AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary.' That sentence reserves the right to change everything.

A research preprint from April, 'Don't Measure Once,' makes the measurement version of the same point. The authors write that AI search answers vary across runs, prompts, and time, and that visibility should be treated as a distribution instead of a single reading. It is a preprint, not a peer-reviewed paper, so treat it as a serious argument and not a settled one.

Where these numbers come from, and how far to trust them

The figures are from OtterlyAI, which published them on September 22. Otterly runs 1,600 prompts every day against seven AI engines and counts which sources get cited. Otterly also sells AI search monitoring, so it has a commercial reason to want you to believe citation share is worth watching. This is one vendor's panel of prompts, built around 16 US industries, over about three weeks. The plateau that carries my argument is shorter than that: roughly one week, from September 15 to publication.

Otterly does not say why the drop happened, and Google did not call me (haha).

I also cannot rule out a measurement artifact. With one vendor and one prompt panel, a change in Otterly's prompt set or in how it parses answers would produce the same step. So the careful version of the claim is this: if the drop is real, it had nothing to do with the content.

There is a second data point, from a different vendor and in the other direction. On September 23, Digiday reported agency data from Tinuiti showing YouTube's share of citations in Google AI Mode more than quadrupled between January and April of this year. Tinuiti is an agency with services to sell, so the same caution applies. The Digiday headline was 'LLMs keep citing YouTube in search results.' That was a fair reading of January to April. It ran one day after Otterly published a panel showing the same share in the same engine at 3.4% to 3.9%, down from 11.0%. The two panels are not the same and do not line up neatly. Both show a share that travels a long way with no change in the content.

The obvious move, and why I would not make it

The same Otterly report says LinkedIn's share of citations across all seven engines rose 39% between September 1 and September 20. You can already hear the advice. YouTube is down, LinkedIn is up, move your effort to LinkedIn.

Look at what the 39% is made of. LinkedIn went from 0.407% of all citations to 0.564%. That is a real increase, and it is also a move from about four tenths of one percent to under six tenths. In the same report, Gemini did not cite LinkedIn once. So the pivot being recommended is to rebuild your plan around half a percent of citations on a vendor's panel. Clear the calendar (haha).

Even if the number were ten times bigger, I would say the same thing. If YouTube can more than quadruple in a few months and then drop about 72% in two days, whatever is up this month can do the same.

Google's documentation says there are 'no additional requirements' and no 'special optimizations necessary' to appear in AI Overviews or AI Mode. Google has its own reasons for saying that. It is also the only party in this conversation that knows the settings.

What the evidence supports, and what it does not

Think about what an AI answer is. Someone asks which firm handles a certain kind of problem. The engine pulls from whatever mix of sources it favors that week and writes a few sentences. The sources are the borrowed part. The subject of those sentences is the part a business cares about.

What the September numbers support is narrow. The source mix is not yours, and on this panel it changed by about 72% in two days. They do not show that what an engine says about a company is any steadier. Nothing I have cited measures that across the September 13 to 15 change, and the same preprint says results for brands vary from run to run as well.

So here is the claim at the size the evidence allows. Citation share is the wrong thing to manage, because you do not control it and it does not describe you. Whether the description of your company holds through a change like this one is an open question.

My bet, and it is a bet, is what I call Trust-First Marketing: put the effort into what buyers and others in your field say about you on the record, and treat what the engines repeat as a result of that and not a target. The September data does not prove the bet. It shows that the alternative is managing to a number someone else sets.

Where I could be wrong

The strongest objection is that which company an answer talks about, and how, is also decided by the engine. If so, I have only moved the problem. I do not have an answer to that. I have a test. Run the same company queries before and after a change like the one on September 13 to 15. If the descriptions swing as hard as the source mix did, my bet is wrong and it is all one dial.

The second objection is practical. The engine has to borrow from somewhere, so you still have to be somewhere. Yes. Publish where your buyers already are, in the format you are good at. If that is YouTube, the September numbers are not a reason to leave. They were never the reason to be there.

And the evidence is thin. One vendor, one panel, about three weeks, one week of plateau, no stated cause. Two results would weaken the argument. If the drop turns out to be a change in Otterly's panel or parsing, the September example is gone. If citation share measured over months turns out to be steady enough to plan around, then a two-day drop is a blip and managing the share is reasonable after all.

You can lose a citation share you did nothing to earn, on a day you did nothing to lose it.

If you want something to do with this, run a small test. Ask an engine about your company by name several times, in separate sessions, and write down what it says. A good result is boring: the answers agree with each other, and they describe you the way your best customers would. A bad result is answers that contradict each other, stay vague, or describe a company you do not recognize. That is a problem with what is on the record about you, and no change of platform fixes it. Then ask about the problem you solve, without your name, and note who the engine talks about and what it says about them. Save both sets of answers and run them again after the next big swing in the citation numbers. If your description moves as much as the source mix did, you will have better evidence against my argument than anything in this essay.