Breaking

Graphite reports AI search answers converge in 79.6% of tests as models cite their own output

June 25, 2026 at 11:09 EDT

Research & Papers
Foundation Models
AI Agents

A report published on June 24, 2026 by Graphite, a consultancy focused on AI search, found that "AI search collapse"—where answers become uniform as AI recursively cites its own generated content—occurred in 79.6% (1,216) of 1,528 simulations. Media outlets have picked up the findings, warning that AI search answers may grow blander and more easily manipulated over time (Axios).

Simulation Study · AI Search

The "AI Search Collapse" Risk: A Feedback Loop Feeding On Itself

A simulation warns that AI search tools increasingly cite AI-written pages — pushing answers toward becoming "narrower, blander and easier to manipulate."

~50%

of article-style web pages already estimated to be AI-generated

model families tested — OpenAI, Gemini, Anthropic

2024

"model collapse" first described, now extended to search

Answer diversity shrinks as the loop repeats

Starting from real results, the simulation turned AI answers into articles and fed them back as the next reference pages. Each cycle, recommendations converged.

Real web

diverse sources

→

Loop ×N

fewer distinct

→

Converged

same answers

The self-consuming loop

AI search answers

→

Published as AI articles

→

Re-cited as references

Errors can compound, and GEO tactics — crafting content to be cited — let the sources be gamed.

Where convergence helps

For fact-checking — a sports champion, a store's hours — agreement across tools is useful and efficient.

Where it harms

"If everyone relies on what the LLM produces, we're going to lose diversity and randomness and exploration."

A caveat: the work is a simulation, not proof that real-world AI search has collapsed — and it is company-led and non-peer-reviewed. The major providers say they continuously refine indexing, ranking and model behavior.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.