A small shift in a model’s behaviour is rarely front-page news. But every so often one lands that you have to sit with before the implications catch up with you.
Recent ChatGPT releases have been citing measurably fewer domains per answer than they did before — a meaningful slice of the citation layer gone within a release cycle — while at the same time running more internal search queries per response, not fewer. The pattern is clear: the model is looking in more places and linking to fewer of them.
Searching harder, filtering harder
The instinct is to read fewer citations as retreat — the model doing less work. The fan-out numbers say otherwise. More queries plus fewer citations means the model isn’t doing less; it’s filtering more aggressively. It examines more sources and keeps fewer.
Anyone who has watched search since 2008 will recognise the shape. Every major Google update did some version of the same thing: crawl more, rank fewer, keep perceived quality high by cutting the long tail. The difference is that Google’s cuts fell mostly on spam and affiliate farms — you could see what was dropped, and nobody missed it. With AI answers, the criteria for “worth citing” are less visible, the pool of potentially citable content is far broader, and the narrowing happens in the dark.
What “authority” means when the model decides
The research coming out of GEO studies this year is converging on a consistent picture of what language models prefer to cite. Four signals keep appearing: strong domain authority, structured and unambiguous factual content on the page itself, broad consensus across independent sources, and — most interestingly — original data or analysis the model can attribute to you specifically.
What the research is not finding is weight behind the things many brands have been paying for: mentions in listicles, guest posts, inclusion in “top 10” articles on second-tier blogs. The citation layer doesn’t appear to care much about those. And as the citation pool tightens, it cares less.
The implication is uncomfortable. The old playbook — build backlinks, place mentions, get covered anywhere — is being replaced by something closer to: be a primary source, or be invisible. You don’t earn a citation by being talked about. You earn it by being worth reading directly.
The narrowing is already visible in shopping answers
Ask an assistant for a product recommendation and you get three names. Not twenty. Not page one of Google. Three brands, ranked and reasoned, with a clear preference — and the direction of travel is towards fewer, not more.
If you’re one of the brands named, this is extraordinary news: the model has pre-qualified the shopper before they reach you, and the less the model cites, the higher the intent of the traffic you receive. If you’re a brand that used to live in the long tail — reachable through a comparison article, a forum thread, an outbound link from a review site — you are being quietly excluded from conversations you used to at least be part of. Not by anyone deciding you don’t deserve it. By a filter that gets more ruthless with each release.
What to actually do
Three practical moves follow from this.
First, stop over-investing in the signals models weight least. Paid mentions on mid-tier blogs and keyword-density exercises were already flattening; a narrowing citation layer flattens them further.
Second, become a primary source worth citing. Own data about your category, publish it, keep it current, and make it structured and unambiguous. The brand most likely to be cited on “best winter gloves for cycling commuters” is the one that measured thermal performance and published the results — not the one with the most affiliate placements.
Third, plan for scarcity. Assume the assistant names one, two, maybe three brands in your category, and work backwards: what would you have to be, and be known for, to hold one of those slots? Making your product information coherent and machine-readable is the foundation everything else stands on.
For the whole of the SEO era, the web’s economic model assumed abundance — page one, then page two, then a long tail that gave smaller brands a way in. The AI era is moving towards scarcity. There is no list. There is an answer, and you are in it or you aren’t.
Find out whether you’re in the answer. Get your free GEO Score.