Engineering Research.
We publish engineering write-ups from our work on autonomous search optimization: the systems we build, the failures we document, and the field data behind both. Entries are dated to the work they describe and compiled from the engineering record.
University partnerships
We partner with programs that value forward-thinking, hands-on training in applied search systems. If that describes yours, see how Zyan works with education; we would welcome the chance to advance this research together.
Field data and client sites
The field data behind this work spans more than twelve months and was collected entirely from non-client properties. We do not train models or run automated systems on a client site without consent, and every property referenced here is anonymized by editorial policy.
Index
- 26 min read
The Strategy Compiler: A Closed-Loop Context Architecture for Autonomous SEO Agents
How search, analytics, technical, local, site, and research evidence compile into a living strategy that constrains specialized agents, verifies local-first execution, and uses measured outcomes to govern its next state.
Publication - 9 min read
Founding Data: What 12x Impressions Growth Actually Looks Like
Six months of Google Search Console data from two properties running our autonomous SEO pipeline: a 17x impressions curve, a flat one, and the reasons we publish both.
Conclusion - 12 min read
Local-First: Running Autonomous SEO Work on Machines We Already Own
A durable job queue that routes AI work to desktop hardware on subscription pricing first and scale-to-zero cloud second, together with the measured economics that forced the design.
Milestone - 11 min read
Anatomy of a Silent Failure: Five Days of a Dead Isolate
A scheduled step hung past its runtime kill ceiling and wrote no ledger rows, not even deferrals, for five days. We present the forensic analysis and the observability doctrine that emerged from it.
Publication - 12 min read
Reserve, Settle, Floor: Metering AI Spend as a First-Class Contract
Every expensive bug in an autonomous platform is eventually a money bug. We document the reservation ledger, the under-charge floor, and the timeout arithmetic that keeps agent spend honest.
Publication - 10 min read
Guardrails That Failed by Being Too Strong
Our file-allowlist guardrail held every blocked post on a file the agent legitimately needed. The resulting guarded-open design combines advisory scope, hard fences, and always-run verification.
Publication - 11 min read
Honesty Engineering: Teaching an Autonomous Platform to Tell the Truth
Sampled metrics presented as sitewide, two meters on one budget, fabricated URLs, and phantom drift banners: a defect taxonomy for systems that report on their own work.
Publication - 10 min read
Context Is an Engineering Surface: How AI Agents Read a Website
Compiled context packs, read-order indexes, and the incident that proved "the file exists in the workspace" is not the same claim as "the agent will read it."
Publication - 11 min read
Porting a 46,000-Line SEO Engine Into a Multi-Tenant Platform in Five Days
Census-first planning, a fourteen-PR chain, and adversarial review panels that caught two SSRF vulnerabilities and a cross-tenant traversal before any of it reached production.
Milestone