The 5-phase method for making a company findable and recommendable by ChatGPT, Claude, Gemini, and Perplexity—with the complete zero-customer case study.
# The method's critical path F0 Diagnóstico → benchmark (baseline) F1 Catálogo → content (~50 fact sheets) F2 Publicação → exposure (AEO/GEO) F3 Agente → conversion (chat + leads) F4 Loop → evidence (measure the delta) # No phase is completed "by eye": # every gate is an executed command.
Most companies simply don't exist for AI—and that's measurable before and after. AIV 2026 packages the method that reverses this, along with the complete execution, documented with the zero client (INEMA).
Diagnosis → catalog → publication → agent → loop. Each phase has steps, verifiable exit criteria, a budget, and real pitfalls encountered along the way.
The zero-customer started with 0% visibility across 52 neutral prompt measurements. War-game plan, record of what was built, and decision report—all in the repo.
A chat that guides visitors through the site, qualifies interest, and captures leads—with proven security gates (adversarial suite, rate limit, closed schema).
Rules that apply to all of them: never invent client data (without confirmation → PENDING), measure before changing anything, test locally before publishing, and never put an agent in production without explicit approval.
~20 neutral questions in the niche, 3 runs on each AI. Records who is cited, with which sources, and with which errors. The baseline is the “before” for the case.
~50 versioned Markdown+YAML fact sheets: company information, products, FAQ derived from F0, portfolio. Extraction with real sources, client review in batches of 10.
Static generator → pages with a direct answer in the 1st paragraph, JSON-LD, and a sitemap, served on the client's domain, with AI bots allowed and tested by user-agent.
Widget in shadow DOM + serverless function + LLM. Navigation and capture via tool use with a closed schema; catalog as the only source—the agent doesn’t make up prices or URLs.
Monthly AI Share of Voice measurement against the baseline + a weekly loop based on real conversations. Questions without a fact sheet become new fact sheets.
The plan assumes barriers will arise: each phase states optimistic and pessimistic assumptions, failure modes with predesigned fixes, and a retry budget.
The method is inexpensive to run—the publishing infrastructure is static and free; the agent uses a low-cost LLM behind a serverless function.
Clone the repo and read the playbook from start to finish before doing anything.
# starting point git clone https://github.com/inematds/aiv2026 cat aiv2026/playbook-aeo-geo.md
Access to the domain/site (for the proxy and robots.txt), 1–2 h/week of review during catalog creation, and a decision-maker available to approve the agent's transcripts.
Static hosting (GitHub Pages works), a serverless Postgres (Supabase) for the agent, and an LLM key—via OpenRouter by default—that never touches the browser.
The executable playbook summary—from zero to continuous operations. The details, failure modes, and complete criteria are in the repo documents (the README includes this same expanded step-by-step guide).
Clone the repo, read the playbook from start to finish, and use the original prompt with the [placeholders] replaced by the new client's project. Plan the safeguards before writing a line of code.
git clone https://github.com/inematds/aiv2026 cat aiv2026/playbook-aeo-geo.md # the method cat aiv2026/case-inema/prompt-war-game-portugues.md # generate your plan with it
Gather ~20 neutral questions in the niche with the client, freeze the question set as v1, and run each prompt 3× on each AI. Without a baseline, no later result can be demonstrated.
# log in CSV: date, tool, prompt, execution, # client mentioned?, competitors, cited sources, AI errors resultado do cliente-zero: 0% de aparição em 52 medições
Markdown+YAML entries in batches of 10, always with a real source. The FAQ comes from the F0 question set itself—questions with proven demand, not made up.
# required front matter for each record slug, tipo, titulo, resumo, status, fonte, confianca, atualizado_em # hard rule: without confirmation → confianca: PENDENTE # the validator rejects publishing a PENDENTE record
Generate static pages with a direct answer in the 1st paragraph, JSON-LD, and a canonical URL on the client's domain. Allow AI bots in robots.txt and verify with curl.
# the gate isn't an opinion; it's a command: curl -A "OAI-SearchBot" https://dominio-do-cliente/conhecimento/ # 200 curl -A "ClaudeBot" https://dominio-do-cliente/conhecimento/ # 200 # + validated JSON-LD + sitemap submitted to Google Search Console # and Bing Webmaster Tools (one-time setup; after that, they re-crawl)
Lightweight widget on the site, serverless function in between, catalog as the single source of truth. Navigation and lead capture are tools with a closed schema—the model can’t make up a URL. The brain calls the OpenRouter by default: an OpenAI-format endpoint for hundreds of models, switching models means changing a string, and the cost per conversation is ready in the dashboard.
# gates before go-live (none are optional) suíte adversarial ≥20 ataques # 0 leaks, 0 actions outside the schema flood de requisições # rate limit returns 429 lead de teste # confirmed with a database query 10 transcrições # approved by the client
Install the 4 measurement signals (site, knowledge pages, agent, share of voice) and rerun the v1 question set every month. The delta against the baseline is the case's key number.
# share of voice stagnant after 3 months? # the lever stops being content and becomes authority # external: the sources AI cites instead of the client re-medição mensal → delta vs baseline → relatório de 10 linhas
Every content change goes to three destinations: the public site, the agent’s database, and search engines. Forgetting any one of them leaves a surface out of date—the agent answers from the database, not the files.
# edit the record in catalogo/ (without confirmation → PENDENTE), then: node scripts/gerar-site.mjs # 1. site: HTML+JSON-LD+sitemap → push node scripts/sync-catalogo-supabase.mjs # 2. agent database (catalogo_fichas) # 3. search engines: the sitemap is published with a new lastmod, and Google/Bing # re-crawl on their own; strategic page changed → request # manual reindexing in Google Search Console (URL Inspection)
Everything verified with a real command and documented in the repo: the original war-game plan, the record of what was built (with runbook), and the decisions and risks report.
of appearances in the 52 baseline measurements—the proof of the problem and the “before” for the case.
catalog entries became indexable public pages, with JSON-LD, sitemap, and AI bots tested one by one.
in the adversarial suite against the published agent: 0 instruction leaks, 0 actions outside the schema.
measurement tools installed for the loop: site, knowledge pages, agent conversations, and monthly AI Share of Voice.
The original prompt that generated the plan—reusable: replace the [placeholders] with the new client's project.
The complete war-game plan, with tables of 22 failure scenarios that transfer almost entirely to any client.
What has been built + the operational runbook: how to regenerate the site, sync the catalog, and remeasure the question set.
The zero-customer completed all 5 phases in 2026-07; the indexing window and monthly remeasurement are underway.