Optimizing for the Black Box: How to Reverse-Engineer LLM Perceptions
You can't see the algorithm, but you can see the output. How to use prompt engineering to 'audit' your own brand's digital footprint.
SEO professionals are used to transparency. We have Google Search Console, keyword volumes, and backlink checkers. We know (mostly) how the algorithm works.
GEO is different. We are optimizing for a Black Box. We don't know the exact weights of the neural network. But we can reverse-engineer them.
The "Probe" Method: A Step-by-Step Guide
Just as a sonar ping maps the ocean floor, specific prompts can map how an LLM perceives your brand.
Don't just ask: "What is [My Brand]?" (That's too easy).
Ask "Probe Questions" to reveal the underlying knowledge graph.
| Probe Type | Example Prompt | What It Reveals |
|---|---|---|
| Category | "List the top 5 [Industry] tools. Explain why." | Visibility: Are you in the consideration set? |
| Association | "What are the downsides of [My Brand] vs [Competitor]?" | Weaknesses: Uncover "negative hallucinations" or outdated pricing. |
| Persona | "I am a [Job Title]. Which tool should I avoid and why?" | Sentiment: Does the AI think you are 'enterprise-only' or 'too expensive'? |
Phase 1: The Category Probe
"List the top 5 enterprise CRM platforms for healthcare. explain why each was chosen in one sentence."
Goal: See if you appear in the consideration set for your specific niche. If you aren't on this list, you don't exist.
Phase 2: The Association Probe
"What are the main criticisms or downsides of using [My Brand] compared to [Competitor]?"
Goal: Uncover "negative hallucinations." Sometimes LLMs memorize outdated pricing or features from 3 years ago. You need to find these phantom weaknesses and correct them with fresh content.
⚠️ Warning: Often, AI models will confidently state that you lack a feature you released 6 months ago. This is because your documentation schema is poor, not because the model is "dumb".
Phase 3: Action Plan
Once you map these perceptions, you can't just edit a meta tag. You have to influence the training data.
- Identify the Gap: "ChatGPT thinks we don't support SSO."
- Create the Artifact: dedicated page titled "Enterprise SSO Support" with clear JSON-LD schema.
- Distribute: Post on a high-authority medium (e.g. Medium, Dev.to) to force a re-crawl.
We can't open the black box. But we can definitely change what we feed it.
Share this article