---
id: e2
slug: the-86-percent-anchor-geo-and-decision-twins
title: "The 86% Anchor: Why You Can Predict a Company’s Next Move by Asking Its AI"
date: November 2026
tags: [GEO & Visibility, Decision Twins, Philosophy of Agency, Empirical]
canonical: https://amyote.com/#essay-the-86-percent-anchor-geo-and-decision-twins
---

# The 86% Anchor: Why You Can Predict a Company’s Next Move by Asking Its AI

> We spent months trying to build AI Decision Twins that predict humans. Then we realized the causal arrow had flipped: humans are aligning with the priors of their AI agents.

**Date:** November 2026 · **Reading time:** 4 min read  
**Tags:** GEO & Visibility, Decision Twins, Philosophy of Agency, Empirical  

For two years, consultancies have promised that if you feed an executive’s speeches, biographies and past trades into a language model, out pops a Decision Twin: a digital oracle that knows what the boss will do next. It is a lovely idea, rather like a horoscope with a GPU. In our PPDE project we put it through a strict temporal-cutoff audit, which is a polite way of saying we stopped it from cheating. [1]

We climbed a ladder of context: C0 (the person’s name and nothing else), C2 (raw RAG over biographical documents) and C3 (a structured decision model). The test bed was 57 historical Warren Buffett cases and a sealed prospective forecast of executives speaking at the Milken Institute Asia Summit on 7 October 2026. Naturally, we expected the most elaborate dossier to win.

> **Pivot Question:** If a detailed biographical dossier doesn’t make an LLM much better at predicting an independent human — what happens when the human starts consulting the LLM before deciding?

First, the cold shower from our PPDE benchmarks: on the 57 Buffett test cases, adding complex structured profiles produced no statistically significant gain over baseline, while a trivial heuristic (repeat-last-quarter) beat the LLM on portfolio moves. In our sealed Milken prospective run, structured twins (T) ranked last across all four evaluations. Across the academic literature (including Funhouse Mirrors, r ≈ 0.20), raw LLMs make poor crystal balls for independent human quirks. [2]

Then came the inversion. A pre-registered randomized controlled trial by Anthropic (N=50 professional analysts using Claude Code / Opus 4.8 on ambiguous business judgment calls) showed a large behavioral shift: when professionals delegate open-ended analysis to an AI agent, they follow the agent’s covertly assigned framing in 86% of cases (vs. 50% chance), abandon their own prior opinion 60% of the time when the agent disagrees, and reproduce the agent’s arguments nearly verbatim in 77% of their written justifications, while crediting the AI in only 15% of rationales. [3]

The practical upshot: you no longer need a mystical psychological twin to predict a decision-maker. If you know your buyer, investor, or competitor is "AI-pilled" on Claude or ChatGPT, ask that model what it recommends in their scenario, and you hold an ~86% cognitive anchor on what the human will argue in the boardroom tomorrow.

## How this rewires Generative Engine Optimization (GEO) and strategy

1. Simulation is no longer a fantasy about humans; it is a measurement of the human’s cognitive anchor. When our @agentcy/geo-sim engine runs 576 counterfactual traces across investor personas in Claude and ChatGPT, it measures the exact priors that shape real buyer shortlists.
2. Optimize for Argument Absorption (Hs), not vanity mentions. Because 77% of professionals borrow their agent’s causal reasoning to defend decisions to their peers, winning in AI search requires feeding models dense, verifiable comparison tables and high-fit opening paragraphs (Question–Paragraph Fit), not Schema.org placebo tags.
3. In niche B2B markets, third-party citations carry a 39.7× multiplier. Where base models lack strong training priors about a brand, appearing on cited external industry pages drives inclusion far more than tweaking on-site FAQ blocks.

## Margin Notes & Caveats

- **[1] Temporal Cutoff Hygiene:** Any evaluation of decision twins without a strict knowledge cutoff leaks future outcomes from pre-training weights. In PPDE, pre-protocol runs are quarantined in reports/_invalid and excluded from valid benchmarks.
- **[2] Prospective Truncation Lesson:** In our Oct 7 Milken blind scoring (Jane Sun, Lily Liu, John Kamara), all 12 sealed forecasts suffered from output truncation (16–31 words), yielding 0/80 exact matches — a humbling reminder to run completeness validators before sealing prospective commits.
- **[3] Retrieval vs. Page Tweaks:** Recent empirical work (Chen & Koudas 2026; Martinez GEO Survey 2026) shows retrieval explains ~32% of AI citation variance while on-page markup explains only 2.8%. Aggressive stylistic rewriting can actually impair dense vector retrieval.

## Appendix: Empirical Evidence

### Why did we pivot from predicting isolated humans (PPDE) to simulating the AI anchors humans rely on (GEO-Sim)?

**Takeaway:** Predicting an unaided human via persona prompts hits a low ceiling (r ≈ 0.20, no gain over C0 name baseline), whereas measuring the AI advisor captures an 86% decision alignment effect.

*Comparison of empirical findings across PPDE internal pilots, external survey literature, and agent-assisted judgment trials.*

| Paradigm / Study | Target Measured | Baseline / Control | Empirical Outcome | Strategic Verdict |
| --- | --- | --- | --- | --- |
| PPDE Buffett Pilot (N=57)* | Historical executive choices | Generic strategy / Repeat-last-quarter* | No significant gain from C3 profile; heuristic won portfolio task | Kill naive persona-prompting |
| PPDE Milken Sealed Run (N=12) | Prospective conference statements | C0 (Name only) vs C3 (Twin) | 0/80 exact matches (truncated outputs); Twin ranked last | Enforce pre-seal length gates |
| Anthropic Judgment RCT (N=50) | Expert analyst decisions with Agent | 50% neutral baseline / Pre-task view | 86% follow agent framing; 60% flip prior view; 77% copy arguments | Model output = 86% human anchor |
| Chen & Koudas / Martinez GEO (2026) | Brand citation in AI search | On-site Schema.org / FAQ blocks* | Retrieval = 32% variance; 3rd-party page mention = 39.7× multiplier | Focus on 3rd-party & Paragraph-1 fit |

## Sources & Methods

- PPDE evaluation ladder: compares C0 (name only), C1 (summary), C2 (unstructured RAG), and C3 (structured decision twin) under a frozen temporal cutoff.
- GEO-Sim matrix (Enso pilot spec): 24 investor personas across 4 cohorts × 2 AI systems × 4 counterfactual worlds (W0 baseline, W1 on-site fact vault, W2 analytical hub, W3 external citations) × 3 replicates = 576 traces.
