RAG architecture choices determine whose sources get cited

Researchers publishing on arXiv have identified a structural flaw in how generative search engines decide which sources to include in their answers. The flaw is not in the retrieval step. It is in the allocation step: how much of the model's attention budget gets assigned to each retrieved document before an answer is generated. The finding has direct consequences for whose evidence ends up in an LLM's response and whose disappears.

The paper, "The Laws of Context Allocation," identifies two compounding problems. First, the standard method for measuring whether a model actually used a retrieved source is unreliable. Conventional relevance proxies collapse on "hard negatives," documents that look relevant but carry no weight in the final generation. This means that for years, RAG systems have been optimised against a metric that cannot distinguish genuine reliance from superficial pattern-matching. Brands that appeared in retrieval results may never have influenced the answer at all.

Second, the dominant architectural choice in RAG systems, widening the context window to fit more documents simultaneously, actively degrades performance. The paper calls this the "architectural trap" of monolithic context widening: as more documents compete for the model's finite attention, relevance decays across the board. The sources that get cited are not necessarily the best sources. They are the sources that survive the attention dilution.

The sequential generation advantage

The alternative the paper proposes is iterative allocation across multiple sequential generations, each working through a smaller document portfolio, rather than one generation attempting to process everything at once. The recall gains are substantial: 16.7 to 20.5 absolute percentage points across models scaling up to 32 billion parameters. That is not a marginal improvement on an existing technique. It is a structural reframing of how context should be managed.

For B2B brands, particularly those in sectors where LLM-generated summaries increasingly mediate expert queries, this reframing matters in a very specific way. A document's chance of being cited in an LLM answer is not determined solely by its relevance to the query. It is also determined by how many other documents it is competing against in the same context window, and how the system has been built to allocate attention across them. A technically superior white paper from, say, a multilateral development institution loses citation probability not because it is less authoritative but because it was bundled into an oversized context batch where attention decay is severe.

This is a redistribution mechanism that content quality alone cannot fix.

What the diagnostic illusion obscures

The paper's causal leave-one-out probe replaces the standard relevance proxy with a direct test: remove the document, regenerate the answer, and measure the change. This isolates genuine generative reliance rather than apparent relevance. The implication is uncomfortable for any brand that has been tracking its presence in RAG retrieval results as a proxy for citation influence. Appearing in retrieval is not the same as shaping the answer.

Financial services firms, industrial groups, and UN-system institutions that publish substantial bodies of technical documentation face a specific version of this problem. Their content is often retrieved because it is authoritative and indexed; it is not necessarily cited because the RAG architecture assigned it enough attention budget to matter. The two conditions, retrieval and citation, are separable, and optimising for the first without understanding the second produces a false sense of visibility.

The paper's findings also suggest that platforms deploying sequential generation architectures will produce systematically different citation patterns than those using monolithic context widening. As AI search matures and different platforms adopt different allocation strategies, the same document may be cited frequently by one system and ignored by another, not because of content differences but because of architectural ones. Brand visibility in LLM answers will increasingly be a function of which systems your target audience uses and how those systems allocate context budget internally.

That is a harder problem to manage than improving content quality. It requires monitoring citation outcomes across platforms, not just retrieval rates, and adjusting content strategy based on where citation gaps emerge. Brands that conflate being retrievable with being cited are measuring the wrong thing.