The problem
Most agencies audit manually. A practitioner opens five browser tabs, runs a handful of free tools, and spends a day assembling a document that captures maybe half the picture. The result is surface-level: keyword gaps, a few technical flags, a Lighthouse score. It rarely surfaces the strategic shape of the problem — where traffic is actually coming from, how the competitive landscape stacks up, what the site is doing to convert (or fail to convert) the visitors it already has.
For a local healthcare practice, the stakes of that shallow read are real. When roughly 43% of existing organic traffic turns out to be a non-converting national blog audience rather than local patients, a shallow audit misses the main lever. When the mobile homepage takes over nine seconds to render its largest element on a 4G connection while competitors earn multiples of the site's traffic value, the urgency is measurable, not just implied. A generalist checklist does not surface that.
This engagement came in through a warm partner channel with a partner agency already in conversation with the prospect. Speed mattered. The pitch had to be credible enough to clear an internal review and defensible enough to stand behind every line item at contract time.
The approach
The core decision was to treat the audit not as a document but as a pipeline with four stages: research, synthesis, packaging, and handoff. Each stage produces an artifact the next stage consumes — audit findings feed the proposal, the proposal feeds the demo, the demo feeds the project board. Nothing is rebuilt from scratch at each handoff.
The audit itself follows a structured multi-dimension framework covering eight areas: SEO, UI, UX, content, brand, target demographics, competitive analysis, and performance. An orchestrator agent spawns parallel sub-agents across those dimensions, validates their findings, and iterates until the plan is complete. This is a reasoning loop, not a static checklist — it treats the whole site as a system and surfaces where dimensions interact. A slow mobile LCP and a weak Google Business Profile are not separate problems; they are compounding drag on the same new-patient acquisition funnel.
One key judgment call shaped the packaging phase: rather than a single proposal price, the engagement was staged as two tiered versions — a lean build with two optional add-ons, and a full build with additional growth levers. This let an internal partner group select the right ceiling before the client ever saw the numbers. Items belonging in an ongoing program (AI visibility, mobile speed) were positioned as retainer levers, so the proposal read as a coherent progression rather than an unbundled menu.
The build
The audit ran in a dedicated project directory using Claude Code's orchestrator-plus-sub-agents pattern. The orchestrator received a single prompt with the target URL, the eight audit dimensions, and an explicit instruction to plan, launch sub-agents, validate, and iterate. Sub-agents worked across dimensions in parallel; findings were written back to the repo as structured files, preserving context for every downstream step. One audit run consumed approximately 800,000 tokens and completed in roughly 40 minutes of wall-clock time.
From those raw findings, the packaging layer produced:
- Two proposal versions (lean and full), each with tiered pricing and add-on logic built in
- A live demo deployed to Cloudflare Pages with real local review data, audit findings formatted as a "gift," and three calls to action — accessible to the internal partner before any client meeting
- A SOW with six itemized deliverables, acceptance criteria, explicit out-of-scope language, HIPAA-safe tracking wording, a no-guarantee results clause, and 50/50 payment terms
- A GitHub repository bootstrapped via the gh CLI with 18 issues (one epic, sprint deliverables, and a phase-2 backlog), four milestones, label taxonomy, and a Kanban board ready for kickoff
Sprint work could begin immediately on receipt of deposit — no setup sprint required.
Outcomes
The audit surfaced findings a prior manual pass had missed — including a stale central thesis (metadata gaps already fixed by the incumbent agency) that was identified and discarded before the pitch was built. The fresh run caught the mobile performance regression, the Map Pack absence despite a claimed and optimized Google Business Profile, the traffic composition problem, and the competitive traffic-value gap — all specific enough to make the scope defensible line by line.
The proposal cleared internal partner review with one revision: plain-language framing over technical depth, now a reusable standard for healthcare clients. Both versions and the demo were live within the same working session the audit completed.
The GitHub repo moved from zero to build-ready in a single automated pass. No setup work would be required once the contract was signed.
The most durable output is the pattern. The same pipeline — audit orchestrator, synthesis, tiered proposal, live demo, repo bootstrap — runs on any prospect URL, at a cost low enough to treat as standard pre-pitch research: you walk into every partner meeting with a complete picture already in hand.
The pattern has since been proven across categories rather than a single vertical. Its deepest run to date was a national news-streaming brand: seven research tracks plus DOM and Core-Web-Vitals evidence (zero structured data confirmed in the live DOM, measured rendering failures), a client deck, a tiered proposal, a brand-true homepage demo taken through three review rounds, and an app-store audit with an action plan — the full engine, end to end, on a business that looks nothing like a dental practice.
What's next
The workflow the earlier draft pointed at — a documented, reusable sprint container — now exists. The orchestrator prompt, sub-agent dimension structure, proposal packaging logic, and repo bootstrap are captured as a staged monorepo pipeline with the proposal template baked in, so any engagement runs from a URL with no custom setup. The next step is breadth: more client categories through the same engine, and tighter coupling between the audit's findings and the live demo so the proof a prospect sees is generated straight from their own diagnosis.