Behind the website.
The public website and its chat agent remain available. We retired automated website acquisition on September 7, 2026. The earlier design is preserved below as a reference, with its original limits and evidence rules.
Earlier architecture Historical reference · Retired September 7, 2026
The content engine
Tuesday and Thursday, 15:00 UTCTwice a week, a scheduled workflow reads the industry feeds, picks one topic that does not duplicate anything already on the blog, and drafts a post in the founder’s voice. An independent critic pass then scores the draft on six rubric items: factual, specific, relevant to operators, voice, answer-shaped, honest. Any score below the bar kills the piece before it exists anywhere public.
Two posts a week is a ceiling, not a target. Quality over quantity is the brand, and a slow cadence is what gives the gate room to say no.
Publish a weak draft. A run that ships nothing prints SKIPPED and ends green, because the gate rejecting a post is the system working, not failing.
It also refuses to invent: no statistic without a named real source, and no client result beyond the ones already published on this site.
Every engine-drafted article carries a disclosure line near the byline linking back to this page, and every run, including the skipped ones, journals itself to the operations database. Published posts land as reviewed commits in the site’s repository.
The community scout
Weekdays, 13:00 UTCEvery weekday morning, the scout gathers live forum threads where an operator is asking the kind of question Benian answers for a living, scores each one against the ideal customer profile, and drafts a reply for the few genuinely worth answering.
Send a generic pitch into a discussion. Connected LinkedIn answers need a relevant operator question, an independent review and a check of the existing replies.
Answers must add something useful without invented experience or client results. Reddit and Hacker News stay manual. Other forums need an authorized connection and rules that permit the proposed automation.
Unsent drafts and accepted replies have separate records. An automated reply needs the platform’s message ID and a readback from the discussion before delivery is treated as verified. Finding a thread or preparing a draft is not a conversation.
The chat agent
Live, on every pageThe widget in the corner answers from a knowledge base compiled out of the same data files that render this site, so it quotes the same published facts and the same labeled case-study figures a visitor would find by reading. Its tools are narrow: it can record a lead when a visitor asks to be contacted, and it can hand over the scoping-call booking link.
Returning visitors are remembered. The tail of an earlier conversation is recalled at the start of the next one, so nobody re-explains their business to a widget.
Guess. Anything outside the published facts gets a straight “that depends on scope” instead of an invented number, and it cannot offer discounts or guarantees the site does not publish.
Every exchange lands in the operations database, which is what makes cross-visit memory and the weekly digest possible. A captured lead is stored on our side first and then forwarded, so a dead webhook loses nothing.
The inbound database
Always on · digest MondaysEnquiries retain the available first and latest campaign information. Browser storage restrictions or missing tags can leave gaps, which remain unknown. Anonymous landing signals are counted separately from identified enquiries, calendar-confirmed bookings and recorded sales outcomes.
At the edge, the site logs GET requests matching 25 self-declared AI and search user-agent labels, with no IP addresses stored. A match is recorded before the downstream response; it does not verify bot identity or successful content retrieval. In the earlier architecture, each Monday a digest composed the week from those tables: leads, inbound signals, crawler coverage, the scout’s review queue, and agent-run health.
Editorialize. The weekly digest makes no model calls; it is a mirror, not an author.
Telemetry is fire-and-forget by construction: a failed write can never slow or break a page for a visitor.
Weekly reports are stored in the operations database, and crawler telemetry is published per user-agent label at /reports/ai-crawlers.
The evidence rule
Every number, every pageThe house rule the other four systems answer to: every figure published on this site carries one of four labels. Measured means we pulled it from production call logs or delivery records ourselves. Client-reported means the client stated it and we print it as their statement. Estimated means calculated from measured or reported inputs, never an observed cash result. Projection means arithmetic from current pace, labeled as such and never presented as achieved.
If a number cannot carry its label, it does not go on the site.
Round up. No unlabeled numbers, no invented industry benchmarks, no vendor stats laundered into results. The label states how we know the number, not how good it looks.
The labels render beside every figure on the case studies page, and the same dataset is served machine-readable at /evidence.json for anyone, or any engine, that wants to check.
Automation needs evidence at each step.
A working system, a published article, and a qualified enquiry are different results. Historic records are useful for learning; they do not establish current operation or commercial outcomes. Our client case studies identify how each reported result was measured.
Architecture record first published August 20, 2026. Retirement status updated September 7, 2026. Crawler measurement wording clarified September 8, 2026. The archived descriptions are not a statement of current schedules.