In a controlled 2026 Web Conference experiment, researchers filled a search pool until two-thirds of it was machine-written, SEO-tuned content. More than 80 percent of what the answer engine then cited turned out to be synthetic. That gap, between the pool’s makeup and the answer’s makeup, is the number search teams should worry about.

Duane Forrester ties the retrieval result to a separate line of peer-reviewed research in a July 9 Search Engine Journal analysis. One paper, presented at the ACM SIGIR conference, named the effect invisible relevance bias. Retrieval systems, the components that decide which pages an answer engine pulls from, ranked machine-written pages above equally good human-written pages. No relevance-based reason justified the promotion. The leading explanation is perplexity: machine-written text tends to read smoother and more statistically predictable, and retrieval models trained on similar text treat that smoothness as a trust signal.

Forrester frames the stakes with two further numbers. More than half of newly published English-language web articles are already AI-generated, according to a Graphite analysis of tens of thousands of pages. Microsoft search and AI lead Jordi Ribas has separately projected that AI agents could eventually issue a thousand times more queries than all human search combined.

That preference compounds over time, a mechanism the Web Conference authors called retrieval collapse. Answer accuracy in their experiment held steady near 68 to 70 percent through the entire contamination process. Nothing on a typical AI-visibility dashboard would have flagged the shift, because the metric teams actually watch, citation frequency, never moved. Underneath it, the diversity of sources feeding those answers had already narrowed to near-identical synthetic pages.

The finding lands inside a two-year SEO debate: does an AI-answer citation reward a page’s real authority, or does it merely reward text that mimics what a model expects a good answer to sound like? Forrester argues the second condition cannot hold. He points to separate Nature research on model collapse: models trained repeatedly on their own output degrade across successive generations. That is the structural reason he expects retrieval systems will eventually need to favor verified, human-produced sources over synthetic ones.

Google’s own guidance on AI features says it evaluates whether content is helpful, not how it was produced. That keeps platforms neutral, for now, on the synthetic-versus-human question. Forrester’s projection is that the neutrality is temporary, not that it has already changed.

Teams that treat AI-citation frequency as a health metric should pair it with a source-diversity check: how many distinct outlets, not just how many mentions, appear alongside them across repeated prompts. A rising citation count inside a narrowing, synthetic-heavy pool is not the win it looks like on the dashboard.

Search Engine Journal published this analysis, written by columnist Duane Forrester, on July 9, 2026.