Monde — How to show up in AI answers

A product I co-founded and built through launch.

Why it exists

People increasingly ask AI assistants instead of searching. Brands have spent twenty years learning to see and shape how they appear in Google — and have no equivalent for ChatGPT, Perplexity, or Gemini. When an AI answer recommends a competitor or gets your product wrong, you don’t even know it happened. Traditional SEO tools don’t cover this channel at all. Monde makes it visible.

How it works

Monde watches the contexts where buying decisions actually form — product comparisons, recommendations, category questions — and turns them into something a marketing team can track and move. I built it as a self-service product with integrated payments.

1. Multi-engine queries. It sends structured queries to ChatGPT, Perplexity, and Gemini, collecting how each engine represents the brand across the comparisons, category questions, and recommendations that shape buying.
2. Competitive benchmarking. Every brand is scored against the competitors that show up in the same answers — visibility, sentiment, and positioning — because share-of-answer only means something relative to who you’re up against.
3. Trend tracking. Shifts in how the engines represent a brand are tracked over time, so a team can connect a move — up or down — to what they shipped or published.
4. Evaluation before model changes. The baseline runs 20 synthetic cases with exact labels and 20 frozen historical outputs through the real extraction and analysis pipeline. Code checks measure recall, precision, schema validity, and position accuracy; a cross-provider LLM judge scores completeness, precision, correctness, and summary faithfulness.
5. Observability and cost control. Every LLM request passes through a shared wrapper that records prompts, outputs, provider and model, tokens, calculated cost, latency, and failures in Langfuse and Postgres. Central model configuration and rollout gates make migrations explicit decisions rather than silent swaps.

What I proved

The production pipeline orchestrated 40,000+ model calls across multiple providers. Three consecutive runs on unchanged code established the baseline’s noise band: nine of 13 code-check metrics had zero spread, and every judge mean stayed within 0.15. The latest run held extraction recall at 95%, precision at 85%, brand-position accuracy at 90%, and schema validity at 100%.

I tested the judge against five deliberately poisoned outputs before trusting its scores; it caught all five, scoring each targeted failure at one or two out of five. I also built separate 30-case drift and candidate-gold datasets, stratified by model channel and known failure modes. Tightening the labeling rules changed 16 of 30 proposed labels, so the next model migration remains gated on completing human review and re-baselining rather than treating the evaluator’s first answer as ground truth.

That is the part of the work that matters beyond this product: turning an AI prototype into an operated system with explicit quality measures, observability, cost controls, and release governance.

Status

The product is live at trymonde.com.