A new study from GEO vendor Graphite finds that AI search systems fed enough machine-written content begin circling back to their own prior outputs, producing answers that gradually converge into near-identical text. The research doesn’t prove production AI search is already collapsing, but the simulations offer a concrete warning: the GEO gold-rush may be quietly eroding the diversity of answers that makes AI citation worth chasing in the first place.
What the simulations actually show
Graphite ran 1,528 simulations across three experimental designs using GPT-5.2, Gemini 3 Pro, and Claude Sonnet 4.5, covering 1,019 unique questions and over one million LLM API calls. In 79.6% of those simulations (1,216 of 1,528), outputs converged: the models began producing responses that were essentially paraphrases of each other. In one test configuration using entity questions with full replacement, the share of response pairs classified as “essentially paraphrases” climbed from 22% to 89% across rounds.
The share of ChatGPT references drawn from AI-generated content in Graphite’s prompt dataset rose from 38.9% in January to 42.7% in June. Graphite estimates that roughly half of article-style web pages may already be AI-produced, though that is an estimate, not a measured figure.
Graphite is explicit about scope: “We do not definitively prove that AI search collapse is already happening.” The study is a controlled simulation, not a live audit of ChatGPT or Google’s AI Overviews. That distinction matters. The full methodology and results are at Graphite’s AI Search Collapse report.
What is AI search collapse?
AI search collapse, as defined in the Graphite study, is a feedback-loop failure mode: when AI retrieval systems draw increasingly from AI-generated sources, the diversity of information in their outputs erodes, causing answers across queries and models to converge toward near-identical responses. The collapse is driven not by AI-generated content in general, but specifically by self-authorship bias, where models disproportionately cite their own prior outputs.
Self-citation is the mechanism, not AI content broadly
The most operationally significant finding is the one that cuts against the common framing. AI generation alone adds approximately zero percentage points to a reference’s citation rate once quality and self-authorship are controlled for. What drives disproportionate citation is self-authorship: even a single self-authored reference in the retrieval pool can start a cascade. Self-authorship alone adds about 26 percentage points to a reference’s citation probability after controlling for eight quality dimensions.
That is a different problem than “AI content is flooding the web.” It is a structural bias in how retrieval works, and it means that brands flooding the web with AI-optimized content to capture AI citations may be feeding a loop that makes all citations less differentiating over time. Axios covered the broader implications of this risk for publishers and brands in its June 25 report on AI search and GEO strategy.
Independent academic work corroborates the underlying mechanism. Researchers Hongyeon Yu, Dongchan Kim, and Young-Bum Kim describe the same dynamic in “Retrieval Collapses When AI Pollutes the Web,” a paper submitted to The Web Conference 2026 (WWW ’26). Their findings show a two-stage retrieval collapse: AI-generated content first dominates search results, then erodes source diversity. Under SEO-style contamination scenarios, a 67% pool contamination rate led to over 80% exposure contamination. The paper is not affiliated with Graphite, an independent academic check on the phenomenon Graphite’s simulations model commercially.
What this means for GEO measurement
Graphite is a GEO vendor with a direct commercial stake in how brands respond to this research, so the practical framing in the study deserves scrutiny alongside the data. Still, the measurement implication is real regardless of vendor motive: if AI answers homogenize, appearing in them stops being a differentiator. Tracking AI Overview impressions via Search Console’s generative AI reports tells you whether you’re cited, but not whether that citation delivers distinct visibility when every competitor answer looks the same.
The durable play suggested by the data is original, first-hand content that models cannot generate from their own prior outputs: proprietary data, direct expert attribution, primary research. That is precisely the kind of source that self-authorship bias cannot replicate and that Google’s own guidance on gaming AI answers as a spam violation draws a line around: the policy distinguishes content that earns AI inclusion through genuine substance from content engineered to trigger citations without it.
The simulation does not test whether mitigations work or cover cross-model retrieval. What it establishes is a plausible feedback loop with a specific, testable driver. For teams investing in GEO, the priority question is no longer just “are we cited” but “are we cited because we have something no machine recycled from itself.”
