The Playbook
We operate seven products with agent loops: agents that write, probe, monitor, and repair AEO daily. We won't tell you which products — so instead, here's exactly how the machine works.
We don't name the products in this portfolio, and we don't publish specifics precise enough to identify them — no domains, no prices, no categories narrow enough to reverse-engineer. That's a deliberate trade-off: it protects properties we still operate day to day.
It also means you shouldn't take outcome claims from this page — there's no way for you to verify a number attached to a product we won't name. That's exactly why outcome numbers don't live here. They live on /lab, where the domain is named (ours), the query set is frozen and public, and every claim is inspectable. This page is mechanisms and cadences only.
The shape of the portfolio
Seven products, one operator, agent loops instead of a team. No engineering hires, no content team, no support team — the loops below do that work, with a human approval gate on anything that touches money, email, or infrastructure.
The portfolio spans two shapes: content-heavy reference sites (think a B2B pricing database, a remote-work guide site) and transactional AI tools (an AI image editor, a file-conversion tool, an AI video tool). Different shapes need different loops — a reference site lives or dies on whether assistants cite its pages; a transactional tool lives or dies on whether the checkout works and support tickets get answered. The loops below run across both, tuned per shape.
AttractOS is what happens when we point that same machine at a domain we're willing to name: this one. More on that below.
The loops that run daily and weekly
Six mechanisms, each running on its own cadence, each catching a specific failure mode:
Citation probing
A fixed set of queries, frozen at a baseline date, gets probed weekly across three live assistant-search lanes: OpenAI's web-search API, Perplexity's Sonar API, and Gemini with grounding. Each probe logs who got cited, appended to a dated, append-only record — never overwritten, so a bad week stays visible instead of getting quietly edited away.
Content generation with a quality gate
LLM-generated pages don't publish on faith. They pass a fabrication check first — because our own evaluation, run across every major model, found that all of them fabricate roughly a third of platform-specific facts on certain content shapes (how-to and platform-fact pages especially). For those shapes, grounded verification against primary sources is mandatory before anything ships. No verification, no publish.
Crawler and health monitoring
Canaries verify that pages actually render, that sitemaps get submitted, and that robots.txt and llms.txt haven't silently broken. Then a watchdog watches the monitors — if a canary goes quiet, the watchdog escalates. Then a dead-man's switch watches the watchdog — if the watchdog itself goes silent, an out-of-band alert fires anyway. Three layers, because monitors that only watch the product and never watch themselves eventually go blind.
Content pruning
Underperforming pages get proposed for removal on a capped monthly budget, not left to rot indefinitely. Growth includes deletion — a page that never earns a citation or a visit is a maintenance cost and a dilution risk, not a free asset.
Attribution
Every content piece carries a source tag, so a scan or a purchase can be traced back to the specific page that produced it. Without this, "content is working" is a feeling. With it, it's a query.
Approval gates
Content-path changes — a new page, a copy edit, a pruning proposal — auto-merge on green CI. Anything touching money, email, or infrastructure requires explicit operator approval before it executes. The gate is drawn by blast radius, not by how the change was authored.
Scar tissue
None of the above was designed up front. All of it came from something breaking on a real property first. Four examples, anonymized but real:
Stale pages, invisible cache
A property silently served stale pages for over a day because its edge cache ignored query-string cache-busts — a deploy would go out, and visitors would still see the old version. The fix: explicit cache purges became a required step in every content deploy, not an optional afterthought.
Secrets wiped around deploys
On one property, the secret store got wiped around a deploy — twice. The fix: a checked-in secrets file plus a post-deploy verification step that fails loudly and blocks if any secret is missing, instead of failing silently in production.
"Traffic growth" that was bots
A traffic chart on one property showed real growth. It wasn't — it was bots hitting a set of programmatically inflated pages. The fix: bot-signature filtering runs before any traffic number gets believed, let alone reported.
Fabricated facts in generated how-tos
Early generated how-to pages contained plausible-sounding platform facts that were simply wrong — the models were confident and incorrect in roughly equal measure. The fix is the grounded-verification gate described above: for fact-heavy shapes, nothing ships unless it's checked against a primary source first.
Why AttractOS is this machine, productized
AttractOS isn't a separate product built to sell a story about the loops above — it's the same probe engine, pointed at your domain instead of one of ours. The free scan and the deep audit run the identical three-lane citation probe described above, against your site.
And the /lab ledger is the same loop running in public, on our own domain, where we can't hide behind anonymity — frozen queries, dated logs, published failures included. It's the one property in the portfolio we're willing to name, specifically so you have somewhere to check our work.
Where the receipts live
Two pages exist so you don't have to take our word for anything above:
- • /lab — the inspectable experiment. Frozen queries, weekly probe logs, interventions, and verdicts, all on a domain we name.
- • /lab/evidence-standard — the claim format. What we mean by "verified," "correlated," and "unproven" whenever we use those words.
Want these loops running on your domain?
Full autopilot is still in development. Today, a design partner engagement means our probe engine and our operator eyes on your citation gaps — not an unsupervised agent making changes to your site.