What is the lost-in-the-middle effect?

A finding by Liu et al. (2023, published in TACL 2024) that LLM accuracy follows a U-shaped curve over context position: models use information at the beginning and end of the prompt well, but performance drops sharply when the relevant passage sits in the middle of a long context.

Can I control where my content lands in an engine's context?

Not directly — but retrieval rank correlates with context position, since engines typically insert higher-ranked passages earlier. Winning a top retrieval slot therefore compounds: you are both more likely to be seen and more likely to be used.

What Is Position Bias in LLMs?

Last updated: July 5, 2026 · Menra Team

Position bias is the tendency of large language models to over-weight information that appears early (and to a lesser degree, late) in their input context, while under-using content buried in the middle. For AI search, it means retrieval rank quietly becomes attention rank: passages inserted at the top of the model's context disproportionately shape — and get cited in — the final answer.

What does the research show?

The canonical study is "Lost in the Middle" by Liu et al. (2023, TACL 2024), which measured question-answering accuracy as the position of the relevant document moved through the context. The result was a pronounced U-shaped curve: strong performance when the answer sat first or last among retrieved documents, with accuracy degrading substantially mid-context — in some configurations falling below what the model achieved with no documents at all. Related work on LLM-as-judge setups found a parallel effect: models systematically favor whichever answer option is presented first, which is why serious evaluation pipelines swap option order and average.

Why position bias matters for GEO

Answer engines retrieve a set of passages — often 5 to 20 — and concatenate them into the model's context roughly in relevance order. Position bias then amplifies the ranking:

Top-retrieved passages are both present and attended to, dominating synthesis and citations
Mid-pack passages may technically be "in the answer's sources" yet contribute almost nothing
Being retrieved is not being used — a distinction visible when engines list sources they barely drew from

The practical consequence: the gap between retrieval position 1 and position 8 is much larger than it looks. GEO effort should aim at winning the top retrieval slots for priority prompts, not merely appearing somewhere in the source list — a difference that answer-position tracking makes measurable.

Example

A security vendor appears in Perplexity's source panel for "what is zero trust architecture" but is never quoted in the answer text. Inspection shows its passage ranks seventh of nine retrieved chunks. After restructuring the page so the definition sits in a self-contained, answer-first paragraph, the passage climbs the retrieval ranking — and the brand's language starts appearing verbatim in the synthesized answer.

Frequently asked questions

What is the lost-in-the-middle effect?: A finding by Liu et al. (2023, published in TACL 2024) that LLM accuracy follows a U-shaped curve over context position: models use information at the beginning and end of the prompt well, but performance drops sharply when the relevant passage sits in the middle of a long context.
Can I control where my content lands in an engine's context?: Not directly — but retrieval rank correlates with context position, since engines typically insert higher-ranked passages earlier. Winning a top retrieval slot therefore compounds: you are both more likely to be seen and more likely to be used.

Keep exploring

See how AI engines talk about your brand — track mentions across ChatGPT, Perplexity, Claude, Gemini and 5 more. Start with Menra