GhostWriter

AI decides what your brand is. GhostWriter writes the source it reads.

Comment

Inspiration

AI is the new search. ChatGPT, Perplexity, Gemini, Claude, Google AI Overviews — they now decide who gets cited when consumers ask for product recommendations. Yet every existing GEO/AEO tool today only monitors how AI describes a brand and recommends that the team write a blog post. The result is indirect, slow, unmeasurable — and brands have no idea whether anything actually moved.

We wanted to ship the first autonomous agent that closes the loop: not monitor and suggest, but monitor, act, measure, and monetize. The shift from clicks-to-citations is the biggest reorganization of the discovery layer in 20 years; brands need a tool that operates at the speed AI engines re-crawl, not at the speed humans write blog posts.

What it does

GhostWriter is a single-trigger autonomous GEO agent. One click and it:

  1. Monitors how AI engines describe a brand across user-supplied queries (via Nimble SERP)
  2. Detects citation gaps where the brand is absent or out-cited (gpt-4o reasoning)
  3. Publishes grounded markdown citeables to cited.md (via Senso content-engine/publish)
  4. Expands coverage — proposes 5 strategic queries the brand should also rank for and auto-publishes citeables for them
  5. Measures the narrative-control lift over time (ClickHouse Cloud time series)
  6. Verifies the fix landed with 4 parallel signals (article live, brand in article, brand in fresh SERP, URL indexed)
  7. Monetizes every read via an x402 paywall — $0.01 USDC lands in the brand's wallet for every agent query

Every step is wrapped in a Datadog LLM Observability span. The full loop runs end-to-end in ~10 seconds.

How we built it

Stack: Next.js 16 App Router on Vercel, TypeScript end-to-end, Tailwind v4, recharts.

Agent loop (lib/agent/loop.ts): a single orchestrator wrapped in a top-level Datadog workflow span. Each phase is a child tool or llm span recorded both via llmobs.wrap() (real Datadog tracing) and synchronously into a trace[] returned to the UI for the live "proof panel". The loop:

ensureTable → monitor(5x in parallel) → detectGaps → for each gap: generateGroundTruth (gpt-4o) → publish (Senso) → simulateAIResponse (gpt-4o-mini ×2) → insertRows (ClickHouse)

Sponsor integrations are each a swappable module under lib/integrations/, so a broken one degrades gracefully without killing the run.

x402 paywall is wired with withX402(handler, recipientAddress, { price: '$0.01', network: 'base-sepolia' }). The buyer side is simulated for the demo flow — in production an agent embedding x402-fetch would auto-sign and retry.

The deck is reveal.js with a custom GhostWriter theme (animated gradient brand mark, per-section tone palette, live status pulse).

Challenges we ran into

  • Senso's publish endpoint wasn't in the public docs. We discovered it by reading the @senso-ai/cli source — POST /org/content-engine/publish with a geo_question_id lookup-or-created from /org/prompts. Once found, we wired lazy prompt caching in senso.ts.
  • Coinbase Wallet defaults to a Smart Wallet (passkey-based, no exportable private key) — incompatible with x402's signer. We generated a fresh testnet-only EOA via viem's generatePrivateKey() and funded it from the Coinbase + Circle faucets.
  • Page cold-start was 31 seconds because we server-fetched ClickHouse at render time. Moved to a static page with client-side hydration; first paint dropped to <500ms.
  • Nimble SERPs vary between calls, so gap counts weren't stable across runs. Widened the brand-match window from top-5 to top-10 and rephrased the verification panel to surface 4 independent signals (article live, brand in article, brand in SERP, URL indexed) instead of one binary "indexed yes/no".
  • pnpm 11 strict build-script policy kept blocking dd-trace's native module postinstalls. Resolved by listing them explicitly under allowBuilds: false in pnpm-workspace.yaml.

Accomplishments we're proud of

  • 6 sponsor tools wired end-to-end in a single autonomous loop (Nimble, Senso, ClickHouse, Datadog, OpenAI, x402).
  • First-class verification UX: 4 parallel signals per citeable, two of which are always green immediately, with the SERP-crawl signal honestly labeled "expected in 24-48h".
  • Real x402 paywall returning a real HTTP 402 with the full v1 protocol payload on Base Sepolia — not a mock.
  • AI engine impact preview: gpt-4o roleplays ChatGPT before vs after the citeable hits the source list. The citation literally flips on screen.

What we learned

The agentic web needs a new substrate — agent-readable, brand-verified, machine-monetizable. cited.md + x402 + Senso together form the rough shape of that substrate. The interesting product surface isn't "another GEO dashboard"; it's the agent that operates that substrate on the brand's behalf.

Also: when sponsor docs are incomplete, read the CLI source. Faster than waiting on support.

What's next for GhostWriter

  • Scheduled recurring runs (cron-style) to catch new gaps and re-publish stale citeables
  • Real ChatGPT / Perplexity API integration for actual (non-simulated) verification
  • Multi-brand orgs + role permissions
  • Sentiment tracking on citations (a citation isn't always positive)
  • Open the x402 endpoint as a marketplace for partner agents — every paid read is both revenue and a credibility

Built With

Share this project:

Updates