SOP-005: Lost in the Middle
Fresh Updated January 2026| Document Control | |
|---|---|
| SOP ID | SOP-005 |
| Version | 1.0 |
| Status | Active |
| Source | Chroma Research (2025) |
Overview
The "Lost in the Middle" phenomenon describes how LLMs disproportionately attend to information at the beginning and end of their context window, while systematically ignoring content in the middle.
The Position Effect
Key Research Data
Position vs Performance
| Position | Retrieval Accuracy | Relative Performance |
|---|---|---|
| First 20% | ~100% | Baseline |
| Middle 40-60% | ~70% | -30% |
| Last 20% | ~80-90% | -10 to -20% |
Query Position Matters
| Needle Position | Accuracy | Notes |
|---|---|---|
| Unique token early | 100% | Best case |
| Unique token middle | 20-50% | Severe degradation |
| Unique token late | 60-80% | Partial recovery |
Critical
Buried content loses 80% citability at scale.
Why This Happens
Attention is unevenly distributed - even on simple retrieval tasks, middle positions are systematically ignored.
Benchmark vs Reality Gap
| Benchmark | Result | Real-World Performance |
|---|---|---|
| NIAH (Needle in Haystack) | 80-100% | N/A - synthetic |
| NoLiMa (Semantic matching) | 40-70% | More realistic |
| LongMemEval (Conversational) | 50-80% | Most realistic |
WARNING
Long-context benchmarks are misleading. Real use fails silently because models appear to work on synthetic tests.
Content Strategy Implications
Do: Front-Load Critical Information
Content Positioning Guidelines
| Content Type | Recommended Position | Why |
|---|---|---|
| Primary claims | First 20% | Maximum attention |
| Key differentiators | First 20% | Citation probability |
| Supporting evidence | Middle 60% | Less critical |
| Call to action | Last 20% | Recency helps |
| Summary/repeat | Last 20% | Reinforcement |
Practical Implementation
For Blog Posts / Articles
[First 20%]
- State main thesis immediately
- Include key claims upfront
- Front-load unique data points
[Middle 60%]
- Supporting arguments
- Examples and evidence
- Background information
[Last 20%]
- Summarize key points
- Repeat most important claims
- Call to actionFor Product Pages
[First 20%]
- Product name + primary benefit
- Key differentiator
- Most important feature
[Middle 60%]
- Feature details
- Technical specifications
- Comparisons
[Last 20%]
- Repeat primary benefit
- Social proof summary
- Clear CTAVerification Checklist
- [ ] Audit existing content for buried key points
- [ ] Move critical claims to first 20% of content
- [ ] Repeat key points in last 20%
- [ ] Test content retrieval with position-specific queries
- [ ] Structure new content with position awareness
A/B Testing Approach
See Also
Citations
"Models perform best when target info at start/end; middle positions ~30% worse" — Chroma Context Rot Literature
"Position accuracy: 100% when unique token early → <20% when late" — Chroma Haystack Position Experiment