SEO and lead capture case studies

Transport · United States · My own project

Context: I wanted a direct answer to a question clients kept asking me: does Google penalize AI-assisted content? Opinions were everywhere and evidence was nowhere, so I built the test myself instead of quoting someone else's blog post.
What I found: That the question was badly framed. Google doesn't penalize the tool; it penalizes content with no verification behind it. The failure mode isn't the model, it's publishing without anyone checking.
What I did: Seven agents with separate jobs: research, writing, evaluation, verification and publishing. A human decision gate at every step that touched a claim. Nothing published without a source the agent could point at.
Result: Zero to organic traffic in seven months, on a site with no backlink budget and no brand. The answer to the original question is no, it doesn't penalize it, when the work is done properly — and now I can say that from my own data.

Financial services · Mass messaging

Context: A financial institution sending mass SMS and WhatsApp messages. The system worked in testing and failed in production, without a clear error anywhere.
What I found: Race conditions and feedback loops. Two processes claiming the same message, and responses re-entering the flow that had generated them. Neither shows up in a log as an error: it shows up as duplicated messages and as messages that never arrive.
What I did: Audited both channels, isolated the reproducible cases, and redesigned the flow so that the same send could not be claimed twice. Documented the architecture for both channels so the next person to touch it wouldn't have to rediscover it.
Result: Documented architecture and reproducible failure modes, which is what turns an intermittent problem into a fixable one.

Transport · Routes platform

Context: A routes platform that needed content at volume. Volume, done wrong, is the fastest way to have your own pages compete against each other.
What I found: That the real risk wasn't writing 245 articles, it was writing 245 articles that cannibalize each other's keywords and split the signal.
What I did: Built a cannibalization framework before writing anything: one search intent per page, documented internal linking, and rules for when a new topic deserves a page and when it belongs inside an existing one.
Result: 245 articles published against a structure, competing with the market instead of with each other.

Food · Hispanic market in the United States

Context: Two sites with more than 50,000 subscribers between them. Search, content architecture and daily operation, continuously.
What I found: An AI-assisted audit surfaced two things a manual review had missed for years: 48 recipes carrying medical claims that fall under Google's health guidelines, and recipes that were technically impossible to cook because a translation error had broken the quantities.
What I did: Fixed the affected recipes one by one and swept the defect across the whole site — a translation error is never a single typo, it's a pattern that repeats wherever the same process ran.
Result: Both defect classes eliminated at the source, not just on the pages where they had been spotted.

Software procurement

Context: A client being charged for something they weren't using, on terms nobody had read since signing.
What I found: The invoice didn't match the contract. It usually doesn't, and almost nobody checks.
What I did: Read the contract, documented the discrepancy and argued it on the client's behalf.
Result: Dispute won. It isn't glamorous work, but the tools inventory in the diagnostic exists because of cases like this one.

What I found, what I did, what came out of it.