Source diversity fails before accuracy does

Retrieval collapse is the failure mode where AI-generated pages take over the results a search or RAG system retrieves, eroding the diversity of sources behind an answer while the answer's factual accuracy stays flat. The order matters: diversity goes first, accuracy holds, and the standard quality checks see nothing wrong.

That claim comes from a controlled experiment, not from vibes. It is worth understanding because it changes what you should be measuring.

#The corpus condition

Start with what the web now contains. Graphite sampled 55.4k English-language article URLs from Common Crawl published between January 2020 and March 2026, classified each with three separate detectors (Pangram, Copyleaks, GPTZero) and averaged the results. Primarily AI-generated articles came in at 49.9% of the sample in Q1 2026, after peaking a quarter earlier at 50.9%.

The number that matters more than the level is the shape. The proportion has plateaued at roughly 50% since Q1 2025, rising fast for two years and then flattening. Two caveats the authors state plainly: detector false-positive rates ran between 1.36% and 1.84% on pre-ChatGPT control articles, and the study does not attempt to classify AI-assisted writing where a human substantially edits a generated draft. Call it "about half, give or take the definition of authorship."

#What happens to retrieval at that ratio

Researchers at NAVER ran the experiment that the corpus number implies. In Retrieval Collapses When AI Pollutes the Web, presented at the ACM Web Conference 2026, they took 1,000 MS MARCO queries over 10,000 original documents and progressively injected synthetic documents across 20 rounds, pushing the pool from clean to roughly two-thirds AI-generated.

The headline result is the gap between two curves. At 67% pool contamination, exposure contamination exceeded 80%, meaning synthetic documents took more than four in five of the top-10 retrieved slots. Over the same 20 rounds, answer accuracy stayed stable or improved slightly, moving from roughly 68% to 70%.

Read those together. The evidence base underneath the answers was almost entirely replaced, and the output quality metric did not flinch. The paper calls this a "homogenized yet deceptively healthy state," which is the most useful phrase in it.

#The uncomfortable part: the synthetic content was better

The reflex response to AI content flooding an index is that quality will sort it out. The experiment says otherwise, and the reason is specific.

The synthetic SEO-style documents were not slop. The SEO pool scored a Micro Correct Rate of 66.79%, against 51.69% for the pool of original documents. The generated content was measurably more factually accurate than the human-written corpus it displaced, on top of being cleaner and better structured for retrieval. It won on the merits as any ranker would score them.

This is why "just write better content" is not a defense against homogenization. Better on accuracy, structure, and readability is exactly the axis the synthetic content already wins. If your differentiation is quality as a ranker measures quality, you are competing directly with something that can be generated at volume and scores well.

#Rerankers filter quality, not sameness

There is a real defensive result in the paper, and it is narrower than it first appears.

The team ran a second scenario with adversarially crafted low-quality documents. There, the ranker choice mattered enormously. BM25 let roughly 19% to 24% of the abusive documents into the top-10, while the LLM-based ranker suppressed them and held exposure contamination near zero. If your worry is deliberate garbage, a modern reranker handles it.

But in the first scenario, the LLM ranker offered no such protection. Its exposure contamination reached 0.7998 by round 20, essentially identical to BM25's. The reranker correctly identified the synthetic content as good, and promoted it, because it was good.

That is the practical distinction. A reranker is a quality gate. It is not a diversity gate. It cannot tell you that ten excellent documents are ten restatements of the same underlying source, because nothing in its objective asks that question.

#What live engines are showing

The NAVER work is a simulation on a benchmark corpus, so it demonstrates a mechanism rather than measuring the open web. Separately, there is observational data pointing the same direction.

Evertune ran 330 distinct prompts repeatedly across 25 days in April and 25 days in May 2026, 99,000 prompts in total, and found Google AI Mode cited 59% fewer unique URLs in May than April, roughly 23,000 fewer, with Gemini down 33%. Total citation volume did not fall. Concentration rose instead: citations per unique URL climbed 8% on AI Mode and 16% on Gemini.

Be careful with this one. Google shipped several changes in that window, and the authors say directly that it is not clear which of them, if any, caused the contraction. A month-over-month window on a moving product is a weak basis for a trend. What it does establish is that a sharp contraction in source diversity is a thing that happens in production, on the timescale of weeks, without any corresponding drop in how much the engines cite.

#What to do differently

Two changes follow, one to content strategy and one to measurement.

On content: optimize for non-substitutability rather than quality alone. The question to ask about a page is whether its content could be regenerated from what already exists in the corpus. If it could, the page is a candidate for displacement by something cleaner that says the same thing. What survives that test is information only you hold: first-party measurements, a named and repeatable methodology, dated primary observation, specific numbers from your own systems. This is the mechanical reason original research earns AI citations while competent summary does not.

On measurement: track the denominator, not just your share. Most citation dashboards report your share of citations for a prompt set. That metric is ambiguous on its own, because it moves when the candidate set moves. A flat share while the unique-domain count is contracting means you are gaining ground. A flat share while it expands means you are losing. Log the count of distinct domains and distinct URLs cited per prompt alongside your own appearances, and you can tell those apart. That is the same reasoning behind treating AI visibility as a distribution rather than a number, and behind the finding that AI citations are graded on a curve against whatever else made the candidate set.

#What this does not prove

None of the above shows that AI-generated content is degrading answers you get today. The accuracy figures point the other way, which is exactly the paper's point. It also does not show that the open web sits at the 67% pool contamination the experiment pushed to; the corpus sampling says something closer to half of new articles, in one language, under one definition, measured by detectors with their own error rates.

What it does support is narrower and still worth acting on. Source diversity and answer accuracy fail on different schedules, accuracy is the lagging one, and a monitoring setup that watches only accuracy or only your own citation share will not see the first failure arrive.