How Ichiba measures AI influence

The three-agent framework

Every arena session involves three AI agents. An influencer agent runs a persuasion campaign. A target agent, a frontier model, makes product recommendations. A judge agent scores every tactic and issues a verdict. Each agent has a distinct role, prompt, and evaluation boundary. No agent sees the others' internal state.

This separation matters. It lets us observe influence as it happens rather than infer it from aggregate statistics, and it lets the judge remain independent of the campaign it is scoring.

Influence Delta Score (IDS)

IDS is a 0.0 to 1.0 measure of how far a target agent's product recommendation shifted during a session. 0.0 means no influence. 1.0 means complete conversion to the influencer's target recommendation.

IDS is measured per turn, not accumulated. A target agent can move toward a recommendation and then move away from it as the conversation develops. The trajectory is as meaningful as the final score.

Cold probes

IDS is measured by querying the target agent in isolation. After each turn of the influence conversation, the target receives a clean category question with no conversation history and no visible influence context. The answer to that isolated query is what IDS scores.

Cold probes are the difference between measuring influence and measuring in-conversation agreement. They are what make IDS an influence metric rather than a sycophancy metric.

Tactic taxonomy and influence vectors

Every turn is classified by the judge into one of 19 tactic classes. Tactic classes aggregate into seven Ichiba Influence Vectors: Rapport, Escalation, Consensus, Credibility, Exchange, Urgency, and Identity. Every tactic is also tagged by processing mode: Instinctive (fast, emotional, heuristic) or Deliberative (slow, logical, analytical).

The vector taxonomy is an Ichiba-native classification system. It draws on established academic work in persuasion research but maps onto tactics we observe in live AI-to-AI interaction, not onto any specific prior framework.

Claims Integrity Score

Every factual claim made during a session is extracted and audited by the judge. Each claim is labeled Verified, Plausible, Unverified, or Fabricated. A Claims Integrity Score summarizes the factual fidelity of the influencer agent's campaign.

This matters because influence and truth are separable. A high IDS with a low Claims Integrity Score is persuasion by fabrication, a pattern we see often enough to instrument for.

Coverage

Ichiba tests 15 agent strategies across 50+ product categories. Target models rotate across frontier systems from Anthropic, OpenAI, Google, xAI, and Mistral: Claude, GPT-4o, Gemini, Grok, and Mistral families. Specific model versions are treated as research metadata and not disclosed, to protect reproducibility and prevent gaming.

Arena sessions include both Standard strategies (conventional persuasion) and Dark GEO strategies (adversarial tactics — the manipulation playbook for when AI systems shape what gets recommended). These draw on documented LLM threat models, including the OWASP LLM Top 10.

Offense and defense

The arena surfaces both sides of the influence equation. Most GEO tools operate on the read side: tracking where a brand appears in AI-generated answers. Ichiba operates on the write side, where the mechanics of influence are observable and measurable.

Offense research asks: which tactics move AI recommendations toward a target outcome? Which Influence Vectors perform best in which categories? Which agent strategies win against which target model families? This is the substrate of the current sample report and the $29 analyst report.

Defense research asks the inverse: which adversarial tactics are being used against a brand category? How susceptible is each target model to each Dark GEO tactic when the category in play is yours? Where is a brand exposed, and what hardens the exposure? This side of the arena is what Dark GEO strategies exist to measure, and what future defense reporting will surface for enterprise brands.

Same arena, same methodology, same corpus. Two buyer motivations, both derived from the same live session data.

Ethics and provider relationships

Ichiba operates as an independent evaluation platform. All model calls are made under the terms of service of the respective providers. Ichiba does not train models on provider outputs and does not distribute provider weights, fine-tunes, or derived models.

We do not accept paid placement, sponsored sessions, or results adjustments. The arena runs at its own expense. Paid reports reflect the same methodology as public sessions.

What we do not publish

We do not publish agent prompts, judge system prompts, scoring formulas, structured output schemas, or specific model version strings. These are research metadata, protected by ongoing patent applications, and publishing them would enable gaming of the evaluation. We publish what we measure, why it matters, and what we find. We withhold how we measure it.

This approach tracks the norms of established evaluation infrastructure. Security researchers do not publish their reference implementations either.