Technical AEO

The 'Lost in the Middle' Problem: Where to Put Your Best Facts

Sep 8, 20268 min read

Research proves that LLMs exhibit primacy and recency bias: they use information from the beginning and end of the context window more than information in the middle. Your most important quantitative claims must be positioned at the start or end of your semantic chunks to consistently win the [1] citation.

InfographicLost-in-the-Middle: LLM Attention by Position

LLM Attention Weight by Token Position

Start (tokens 1–200)
91
Early-mid (tokens 200–600)
72
Middle (tokens 600–1200)
31
Late-mid (tokens 1200–2000)
24
End (tokens 2000+)
68

Where to Place Your Best Facts

Primary factual claimSentence 1 of first paragraphCritical
Key statistic or data pointWithin first 200 tokensCritical
Definition of key termFirst H2 section openingHigh
Supporting evidenceAfter primary claim (tokens 100–400)Medium
Related questions (FAQ)Last section (end recovery zone)Medium
Background contextMiddle — avoid or removeLow

Source: Stanford "Lost in the Middle" research + RankAsAnswer application · 2024

The “Lost in the Middle” research

The “Lost in the Middle” paper (Liu et al., 2023) demonstrated a consistent pattern across multiple LLM architectures: when models are given a long context window containing multiple retrieved documents, their ability to use information from the middle of that context degrades significantly compared to information at the beginning (primacy) or end (recency).

In RAG systems with 5–10 retrieved chunks, documents in positions 2–8 received significantly lower utilization rates than documents in positions 1 or 9+. The performance degradation at middle positions was not small: it ranged from 15–40% lower utilization depending on context length and model architecture.

The practical implication for GEO: even if your chunk is retrieved by the vector database, its position in the assembled context window determines how much the model uses it. A chunk at position 5 of 8 may be retrieved but largely ignored.

You cannot control context window position

You cannot directly control where in the assembled context window your chunk appears — that is determined by retrieval ranking algorithms. What you can control is the within-chunk position of your key facts, and the overall retrieval score of your chunk (which influences whether it ranks in position 1–2 or position 5–8).

Primacy and recency bias explained

Primacy bias refers to the disproportionate influence of early-appearing tokens on the attention mechanism output. When transformer models process a sequence, early tokens accumulate more attention interactions across the full sequence than middle tokens. This gives early tokens structurally more influence on the final output representation.

Recency bias refers to the tendency for later tokens to have higher residual attention weights due to proximity effects in causal attention. Final tokens in a sequence often have elevated influence on the output because they are the most recently processed representations.

Middle tokens — tokens in the central portion of a long context — experience neither the full primacy advantage nor the recency advantage. They are processed last by early layers (lower primacy) and first by final layers (lower recency). The result is consistently lower utilization rates for middle-position content.

Context window utilization by position (8-chunk context)

PositionUtilization rateCitation probability
1 (first)91%High
278%Good
361%Moderate
452%Low
549%Low
654%Low
771%Moderate
8 (last)85%High

Chunk level vs context level positioning

The Lost in the Middle effect operates at two levels: the context window level (which chunk ranks first vs fifth in the assembled context) and the chunk level (which sentence comes first vs middle in your 512-token chunk).

At the context window level, higher retrieval scores translate to earlier positions. The highest-density, most semantically aligned chunks rank at positions 1–2. This is controlled by information density, structural richness, and entity clarity.

At the chunk level, the primacy effect still applies within the chunk. Even if your chunk is at position 1, a fact buried as the fourth sentence of the chunk receives less attention weight than the same fact in the first sentence. Both levels require primacy optimization.

The fact positioning strategy

Apply primacy positioning at every level of your content structure. For each H2 section: the first sentence states the single most important claim. For each paragraph: the first sentence states the paragraph's primary claim. For each list: the most important item is first. For tables: the most-cited row appears first.

Recency positioning applies at the section level: the final sentence of each H2 section should restate or reinforce the most important claim with different phrasing. This creates both a primacy capture (first sentence) and a recency capture (last sentence) for the key fact, maximizing citation probability across both attention mechanisms.

The opening sentence rule

The opening sentence of every section must contain the most important quantitative or entity-specific fact in that section. It must follow the Answer-First pattern: Subject → Claim → Quantitative anchor. No setup. No context. No question-framing.

If you are writing a section about customer acquisition costs in SaaS, do not open with "Understanding customer acquisition cost is critical for any SaaS business." Open with "The median CAC for B2B SaaS companies reached $1,450 per customer in 2026, a 23% increase from 2024, according to OpenView Partners' annual SaaS benchmark report." The second sentence contains the fact. The first contains zero facts and wastes the primacy position.

The closing sentence rule

The final sentence of each H2 section should not be a summary or a transition. It should be a second, complementary quantitative claim that reinforces the section's primary topic. This positions a high-value fact at the recency position within the chunk, ensuring that even if primacy attention is partially captured by the opening claim, the recency position captures it again.

Avoid section-closing sentences like "As we can see, CAC optimization is a key priority." Replace with "Companies that reduce CAC by 20% through content-led acquisition consistently achieve payback periods under 12 months at scale."

Practical rewrite guide

For each section of your content: identify the single most important claim. Move it to sentence 1. Identify a complementary quantitative claim. Move it to the last sentence. Delete or rewrite every middle sentence that does not contain a specific fact. The resulting section will be shorter, denser, and significantly better positioned for both primacy and recency capture.

The section rewrite formula

Sentence 1: Primary quantitative claim (primacy capture). Sentences 2–N: Supporting facts, entities, and context. Final sentence: Secondary quantitative claim or strong corroborating fact (recency capture). This structure consistently outperforms narrative-structured sections in RAG citation tests.
Was this article helpful?
Back to all articles