ChatGPT stopped citing you

AI search is consuming more of the web than ever, but giving less back. With the rollout of GPT-5.3 in March 2026, ChatGPT now runs over ten fan-out searches per prompt, pulling from a wider range of sources than previous models. Yet the number of links it actually shows users has dropped by roughly 20%. It searches more, but cites less. This isn't a bug. It's a design choice. And it has serious implications for anyone who publishes content on the open web.

The numbers behind the shift

Resoneo, a French SEO consultancy, tracked 400 prompts daily over 14 weeks using an AI visibility platform called Meteoria, producing about 27,000 comparable responses. Their findings were stark: average unique domains per response fell from 19 to 15 after the GPT-5.3 transition. Average unique URLs per response dropped from 24 to 19. That's not a subtle change. One in five previously cited sources simply vanished from the output. OpenAI's framing? A better user experience. Fewer, more relevant links. But the mechanism behind the change tells a different story. GPT-5.3 now evaluates what it calls "authority signals", essentially credentials, trust indicators, and institutional weight, before deciding whether to include a link. The model isn't just summarizing information anymore. It's deciding who deserves to be seen.

AI search as a credentialing system

This is the part that should make independent publishers uncomfortable. When an AI model decides which sources are "authoritative" enough to cite, it becomes a gatekeeper. And unlike Google's PageRank, which at least had a transparent-ish logic (more links to you = more authority), the criteria here are opaque. Who decides what's authoritative? The model does, based on patterns in its training data and whatever signals OpenAI has baked into its retrieval pipeline. There's no public documentation on how these authority signals work. There's no way to audit them. And there's no appeals process when your site gets quietly dropped from the citation pool. The result is predictable: established institutions and large publishers get cited. Everyone else gets scraped. As BotRank's analysis put it, "a smaller citation set gives more weight to the pages that already look trustworthy and easy to parse." If two competitors publish similar guides, the one with cleaner structure and institutional credibility captures the citation while the other disappears entirely from the answer.

The zero-click acceleration

This trend didn't start with GPT-5.3. It's an acceleration of something that's been building for years. More than 65% of Google searches now end without a click, according to aggregated data from Semrush and Search Engine Land. Google's own AI Overviews have crushed organic click-through rates, with Seer Interactive's September 2025 study showing organic CTR plummeting 61% (from 1.76% to 0.61%) for queries where AI Overviews appear. Paid CTR dropped 68% on those same queries. And it's not just Google. A Q4 2024 report from TollBit found that AI chatbots like ChatGPT and Perplexity drive 95 to 96% less referral traffic to publishers than traditional Google search. Click-through rates from AI answers measured below 1%. Now OpenAI is following the same playbook. More AI-generated answers, fewer reasons to click through, and a shrinking pool of sources that get mentioned at all.

Small publishers get hit hardest

The pain isn't distributed evenly. Chartbeat data reported by Axios in March 2026 showed that small publishers lost 60% of their search referral traffic over two years, compared to 22% for large publishers. That gap is enormous. Large publishers have brand recognition, direct traffic, and the institutional weight that AI models now reward. Small publishers, indie blogs, niche experts, they're the ones who built the web's information layer, and they're the ones being squeezed out. The Reuters Institute's 2026 report, drawing on views from 280 media leaders across 51 countries, found that executives expect search engine referrals to fall by 43% over the next three years. Publishers have been "surprisingly candid," as AdExchanger noted, about losing 20%, 30%, and in some cases as much as 90% of their traffic and revenue. Chartbeat's broader network data showed that between November 2024 and November 2025, traffic from Google Search to over 2,500 sites decreased by 33% worldwide and 38% in the US.

The extraction paradox

Here's the irony that sits at the center of all this: the web content that trained these models is the same content being de-linked. LLMs like GPT-5.3 were built on the open web. They learned from blog posts, news articles, documentation, forums, and everything in between. The quality of their answers depends entirely on the quality of the content they consumed. But the system they've created actively discourages the production of that content. If publishing on the web no longer drives traffic, and if AI search no longer links back to your work, then the economic incentive to create high-quality public content collapses. This isn't a theoretical concern. It's already happening. Chegg, the education platform, saw its business model gutted when AI systems started replacing structured learning content. NPR reported that online publishers face what amounts to an "extinction-level event." This is extraction at scale. The models take everything, give back almost nothing, and then optimize for giving back even less.

The incentive problem

It's tempting to frame this as "AI is evil," but that misses the point. The real issue is a broken incentive structure. The web has always operated on an implicit deal: you create content, search engines send you traffic, and that traffic sustains your business through ads, subscriptions, or conversions. AI search breaks that deal. It takes the value of your content (the answer to someone's question) and delivers it without sending the user to your site. Some curation is genuinely useful. Nobody wants to sift through ten blue links when a well-synthesized answer is available. The problem isn't synthesis, it's the lack of reciprocity. The value flows one way. Ross Simmonds, writing about GPT-5.3, argued that "traffic from ChatGPT" was always the wrong metric and is now "officially dead." His framing is that AI visibility is more like PR or thought leadership than performance marketing. But that's cold comfort for the small publisher whose business model depends on actual visitors.

What does a content strategy even look like now?

If AI search won't link to you, what do you do? Some options are emerging, though none are perfect. Building direct audiences through newsletters, communities, and social platforms reduces dependence on search referrals. Creating content that AI can't easily replicate, original research, unique datasets, personal experience, raises your chances of being cited when citations do happen. But the honest answer is that nobody has fully figured this out. The rules are changing faster than anyone can adapt to them. A blog strategy built on SEO fundamentals two years ago might be irrelevant today. What's clear is that the relationship between content creators and AI platforms needs to be renegotiated. Right now, the platforms are setting the terms unilaterally, and those terms amount to: we'll use your content, but we won't tell anyone where we got it. For anyone who publishes on the web, that should be deeply concerning. Not because AI search is inherently bad, but because the economics of content creation are breaking down, and nobody with the power to fix it seems particularly motivated to do so.