Surfaced

Know where your brand stands in AI answer engines and exactly which creator content will move it.

Comment

Inspiration

Brands obsess over Google SEO but have no visibility into AI answer engines — ChatGPT, Gemini, Perplexity, and Claude now answer millions of buying decisions every day. An influencer agency asked us: "Is our client the brand being recommended, or just mentioned?" Turns out those are completely different things.

What it does

Surfaced probes four AI engines with real buyer questions, scores whether the brand is the lead recommendation (not just mentioned), and ranks creator-content plays that will close the gap — by platform, format, reach, and citability.

Three tabs:

  • Scorecard — lead share per engine, topic heatmap, trend delta over time
  • Answer Inspector — the actual answers, citations, and prominence signals
  • Game Plan — ranked creator plays with a 6-step reasoning chain per play

Add any brand in ~2 minutes: the system verifies it, auto-generates the probe config, and runs a simple → deep sweep.

How we built it

Surfaced is an autonomous GEO agent. Given a brand name, it dispatches web-grounded queries across ChatGPT, Gemini, Perplexity, and Claude in parallel, classifies citations from the open web, scores competitive prominence, and publishes a ranked content strategy to cited.md — all without human intervention.

  • OpenRouter as the single provider: one key, four engines. All citations come back as url_citation annotations via the openrouter:web_search server tool.
  • Tavily Search powers brand research in the onboarding step — instead of parsing raw HTML, we search for structured brand overviews before LLM config derivation.
  • ClickHouse Cloud stores every sweep as timestamped rows; delta is computed from the previous snapshot — not hardcoded.
  • Langfuse traces every probe call for cost and latency.
  • x402 payment gate on /api/probe — any autonomous agent can discover, pay for, and consume a GEO sweep. Server returns 402 + payment details; client pays and retries.
  • Prometheux produces explainable provenance chains for recommendations; the TypeScript recommender is the graceful fallback.

Challenges

Binary mention detection was misleading: Anker scored 100% surface rate even though UGREEN was the #1 pick in every answer. We replaced it with a prominence scorer — pure heuristics (first-mention race, superlative proximity, list-item position) — that asks the sharper question: is the brand the lead?

What we learned

Prominence > presence. A brand can appear in every AI answer and still be losing. The creator-content play that moves the needle is the one that gets the brand into the first sentence of the answer, not just into the answer.

Built With

  • claude
  • clickhouse-cloud
  • langfuse
  • node.js
  • openrouter-api
  • prometheux
  • react
  • tavily-search
  • typescript
  • vite
  • x402-payment-protocol
Share this project:

Updates