ChatGPT is citing YouTube and Facebook far less. The lesson is about what you do not own.

Between August and September 2026, YouTube lost 91% of its share of ChatGPT citations, Facebook 88.1%, Wikipedia 78.4% and Forbes 71.7%, according to measurements from the GEO tooling vendor Peec AI, reported by Abondance on September 28, 2026. None of those four sites was penalised and none of them changed how it publishes. ChatGPT changed how it picks its sources. For a business owner the lesson fits in one sentence: visibility that sits on a platform you do not own can halve in a matter of days, with nothing you did as the cause.

What actually happened between August and September 2026?

Four of the largest domains on the web saw their share of ChatGPT citations collapse from one month to the next: YouTube down 91% in relative terms, Facebook down 88.1%, Wikipedia down 78.4% and Forbes down 71.7%, based on Peec AI figures published by David Konitzny and reported by Abondance on September 28, 2026. Citation share measures what proportion of the links ChatGPT displays in support of an answer point to a given domain. Losing 91% of that share does not mean YouTube lost 91% of its audience. It means ChatGPT now shows YouTube as the source of what it asserts roughly nine times less often.

These numbers come with a caveat the Abondance piece states itself. Peec AI sells AI visibility tooling, its data is not independently audited, the method rests on hundreds of thousands of prompts tracked over time without the detail being published, and the compared periods are not dated to the day. A domain can also be retrieved by the model without being displayed as a source: retrieved and cited are not the same thing. The direction of the move is solid. The decimal point is not.

Is this the first time a platform has vanished from citations?

No, and the August 2026 precedent is better documented. Reddit went from an average 3.83% of ChatGPT citations between July 18 and August 7, 2026 to 0.52% between August 14 and 17, a relative drop of 86.4%, according to Klaas Foppen's analysis for Promptwatch published on August 20, 2026. The shift took four days. The same analysis notes that ChatGPT's use of the site: operator inside its own fan-out queries jumped from around 0.37% to 16.8% on August 8, 2026, a forty-six-fold increase.

The decisive finding sits elsewhere in that analysis. Over the same window, Google AI Overviews shows only an 11.3% relative decline in Reddit citations and Google AI Mode roughly 30%, with no single-day cliff anywhere. When one answer engine trims a domain slightly and another drops it overnight, the problem is not the domain. It is how one product decided, that month, which sites it would bother asking.

Why are these drops not penalties?

A penalty can be contested, documented and fixed. A search engine publishes its rules, announces its updates and lets a site come back into compliance. None of that applies here. What changed is an internal source-selection layer, invisible from the outside and revisable without notice. A technical analysis by Tomek Rudzki for Peec AI, published on September 28, 2026, describes an index of ChatGPT's own, organised into separate verticals including a video vertical, alongside at least eight external providers. That work is reverse engineering rather than OpenAI documentation, but it makes the movement legible: when the way a model asks changes, the list of sites it cites changes with it.

The practical consequence for a business owner is blunt. There is no complaints desk for a lost share of citations, no announced recovery window, no published criterion to comply with. That is a difference in kind from classic search, and it is what makes any strategy that hands your visibility to a single intermediary fragile.

What does this change for a small or mid-sized business?

Three concrete things. First, if your presence in AI answers runs mainly through a YouTube channel, a Facebook page, a directory listing or a Reddit thread, that presence is borrowed. September 2026 has just shown it can be taken back without notice. Second, search engines have not gone away and remain the base layer: what is happening is a shift in the mix of entry points, not a clean replacement of one by another. Treating the classic channel as dead is how businesses lose both at once.

Third, standing still has a measurable price. When an AI summary appears above the results, the click-through rate of the top organic position is 58% lower than it was, according to the Ahrefs study of February 4, 2026 covering 300,000 keywords compared between December 2023 and December 2025. Clicks are getting scarcer on both surfaces at once: inside the AI answer, where a citation is the only way to exist, and underneath it, where the blue link returns less than it used to. A company with neither citation nor click does not have a channel problem. It has an eligibility problem.

Should you stop publishing on YouTube, Reddit or LinkedIn?

No, and that would be a poor reading of these numbers. A YouTube channel that sells, a review page that reassures buyers, a thread that answers your customers: all of those hold exactly the value they held last month, because they bring an audience and social proof that owe nothing to an AI index. What has just been disproved is the idea that publishing there is, on its own, a visibility strategy for AI answers.

• A third-party platform is useful for reach and proof. It is not an AI visibility asset.
• Your own domain is the only surface whose readability, structure and files you control.
• Content that exists only on a platform disappears from citations at the same time the platform does.
• The question is not where to publish, but whether what you publish still makes sense pulled out of context.

What do you actually control inside an AI answer?

You do not control the citation, and nobody can promise you one. You control eligibility: pages an answer engine can read without executing code, claims that hold their meaning once lifted out of their paragraph, a company identity declared without ambiguity, public files that state who you are and what you sell. That is the part of the work that depends on no product decision taken in San Francisco, and the only part that survives a change in source selection.

We set out what that eligibility covers in "How do you get cited by ChatGPT when you are a small business?". And since the reflex, on reading news like this, is to open ChatGPT and ask whether it knows you, it is worth repeating what we wrote in "Asking ChatGPT whether it knows you measures nothing": a single answer varies from session to session and measures neither your position nor your progress. What can be measured is what your site exposes, signal by signal.

There is also a case where the right answer is to do nothing. If your business has no content of its own worth citing, if your site is three presentation pages with no verifiable information on them, no technical setting will create the substance that is missing. The budget is better spent on having something to say than on making a thin site machine-readable.