RAG leaves geographic gaps in LLM answers on public firms

Retrieval-augmented generation was supposed to be the great equaliser. Feed the model the right document and it will answer correctly, regardless of whether the firm in question is headquartered in New York or Nairobi. A study posted to arXiv demolishes that assumption with systematic evidence across roughly 2,000 public companies drawn from global equity indices.

The paper, "When RAG Fails to Equalize: Geo-bias in Factual Question Answering over Public Companies," tests six LLMs under four retrieval conditions: no context, perfect context, misleading context, and distraction context. The finding that should concern every communications director at a non-Anglo-Saxon company is this: even when models receive accurate, relevant retrieved documents, the accuracy gap between well-covered and poorly-covered firms does not close. It narrows. It does not disappear.

The mechanism behind the gap

The authors demonstrate that retrieval gains correlate with baseline accuracy. A model that already knows a lot about a firm extracts more value from a retrieved document about that same firm. A model that knows little about a firm misreads, ignores, or is confused by the same document. Internal representations shape how retrieved information is processed. RAG does not bypass parametric knowledge; it interacts with it.

This is a structurally important result. The common defence of RAG-based systems is that grounding in source documents removes the need for the model to "know" something. The arXiv paper shows this is only partially true. The model's prior familiarity with an entity acts as a scaffold for retrieval. Without that scaffold, the document lands on uncertain ground.

The geographic pattern is predictable. Companies in US and European indices score consistently higher than firms in emerging or frontier markets, both without context and with perfect context supplied. Larger models improve overall scores but do not remove these structural disparities. Scale is not a solution to systematic underrepresentation.

What this means for brand visibility in LLM answers

For a multinational industrial group with significant operations in Southeast Asia or sub-Saharan Africa, the implication is direct: if the LLM's parametric knowledge of your local subsidiary or joint-venture partner is thin, even a well-optimised retrieval pipeline will produce weaker, less accurate answers about that entity. The problem is not solely one of corpus quality; it is one of model architecture interacting with uneven training data.

Multilateral institutions face a version of this problem that is particularly acute. The UN system, the World Bank group, and regional development banks operate across precisely the geographies where model knowledge is thinnest. When a policymaker or researcher uses an AI assistant to query facts about a programme in Francophone West Africa or a bond issuance from a Latin American development bank, the retrieval layer may supply the right document and still produce a degraded answer, because the model's ability to parse and apply that document is compromised by weak prior familiarity with the institutional context.

Financial services firms with emerging-market portfolios have a related exposure. Analyst tools built on LLMs will, if the arXiv finding holds at production scale, produce systematically less reliable outputs on companies in markets where model training data is sparse. This is not a temporary calibration problem. It is a structural feature of how current architectures work.

The misleading context finding compounds the risk. When models are fed incorrect documents, they copy the errors rather than flagging uncertainty. For geographies where models have low baseline confidence, a poisoned or outdated document in the retrieval corpus becomes a direct vector for factual error in the model's output. A firm that has had a change of leadership, a restructuring, or a rating change, and whose Wikipedia entry or press coverage has not caught up, faces a real probability that an LLM will confidently assert the outdated information.

The practical consequence for content strategy

The path forward is not to abandon RAG. It is to recognise that retrieval quality is necessary but not sufficient for factual parity across geographies. Brands operating in underrepresented markets need to build the parametric scaffold, not just the retrieval layer. That means sustained, structured, machine-readable content published in sources that AI training pipelines actually ingest: major financial data providers, regulatory filings in standardised formats, wire services with broad syndication, and Wikidata entries with verified attributes.

The arXiv research makes a quiet but pointed argument: the companies and institutions that win in LLM-mediated information retrieval are those the model already half-knows. For organisations whose geographies and business contexts sit outside that charmed circle, the work of building AI visibility is less about prompt engineering and more about the long, unglamorous task of becoming legible to a system that was trained, overwhelmingly, on someone else's world.