ADR-0007: Dogfood jsglobtech.com as the live demo asset

• Status: Accepted
• Date: 2026-07-30 (recorded; from ../source/JSGlobtech_GEO_Agency_Playbook.pdf §6, authored 2026-07-27)
• Deciders: Junaid Siddiqui (owner)
• Related: ../sales/demo-framework.md, ../architecture/geo-stack.md, ../integrations/mcp-server.md, ADR-0001, ADR-0002, ADR-0003

Context

Selling GEO and an agentic action layer to a non-technical buyer is hard because both are invisible. Slides describing schema markup do not persuade a clinic director, and with no client case studies yet there is nothing to point at.

Options:

1. Slides and case studies — nothing to show; no clients yet.
2. A throwaway demo site — controllable, but obviously staged and proves nothing about production capability.
3. A client pilot — best proof, but requires a client first, which is the problem being solved.
4. The agency's own site as the reference implementation — available now, and being cited by a real LLM is self-evidently not staged.

Decision

Build the full GEO and agentic stack on jsglobtech.com itself and demo against it live, in two steps (../sales/demo-framework.md):

1. Citation search — run a real query on ChatGPT or Perplexity, show JSGlobtech being cited, attribute it to the live /.well-known/llms.txt and structured data.
2. MCP lead capture — in a connected chat session, ask for a consultation, show the host's confirmation prompt, then show the lead arriving in the destination Sheet.

The consequence is architectural, not just marketing: the GEO artifacts (ADR-0002) and the MCP endpoint (ADR-0003) are sales infrastructure. public/llms.txt's "Location Signals for AI Search" section is written to match the demo query, and the MCP endpoint must stay live and working for the demo to exist.

Consequences

• Positive:
  • Proof rather than assertion, with no client and no case study required.
  • Forces the stack to be genuinely production-quality, because it is demoed at full volume to prospects.
  • "We did it to ourselves first" is a credible answer to "does this actually work?".
  • Every improvement to the site improves the sales asset.
• Trade-offs / costs:
  • The demo depends on non-deterministic LLM output. A citation may not appear on a given run, and it varies by model, version, and whether browsing is on. Rehearse immediately before every call and keep screenshots.
  • Breaking the site breaks the pitch. The MCP endpoint, both llms.txt files, and CRM_WEBHOOK_URL in production are now sales-critical — an outage costs deals, and there is no monitoring in place.
  • Demo leads pollute the real lead Sheet unless cleaned up.
  • The proof is for a software agency. A clinic prospect will ask whether it works for a clinic; the visibility layer transfers directly, the action layer needs reshaping per vertical (ADR-0006).
  • Requires per-call setup: the MCP endpoint or OpenAPI action must be pre-connected in the host client.
• Follow-ups:
  • Add uptime monitoring on /api/mcp/sse and both llms.txt paths — these are revenue-critical now.
  • Keep a fallback screenshot or recording of a successful run of each demo step.
  • Replace with a real client case study per vertical as soon as one exists.
  • Adopt a recognisable demo-lead naming convention so rows are easy to purge.