My Agents Built Paul Ryazanov’s Revenue Optimization Pack

TL;DR: I pointed my agent at a plain-English brief: take the audit, optimization, and reporting agents we have demonstrated publicly, tune them for e-commerce revenue optimization, and turn them into a skill pack that Paul Ryazanov‘s agency can deploy across its clients — then publish the whole system on his site so anyone can inspect it. The agent researched Paul’s business, wrote nine skills (from a GA4 + Google Ads + Clarity revenue aggregator to a CRO experiment planner), generated a QR code, built and published paulryazanov.com/agents with a free downloadable pack, and QA’d everything against our article guidelines. When the working-session notes corrected the positioning mid-flight, the agent retuned the whole system the same day — the revision log below shows it. This meta-article documents exactly how, following the Meta-Article Prompt.

The Task

Paul Ryazanov runs MageCloud, an e-commerce agency built on development roots — dozens of engineers and marketers across the UK, Denmark, Ukraine, and the USA who ship and maintain Magento, Shopify, and WooCommerce stores. He has earned $2M+ on Upwork across 274+ jobs at 100% Job Success, advised 120+ retailers, and spoken at PubCon, ConversionConference, and BrightonSEO. We have known each other through those conference circles for years, and my team already maintains his personal brand site.

Paul is in the middle of the hardest transition an agency can make: from hourly billing to result-driven revenue optimization, where the agency is judged on what the store earns. That model needs two things his team should not hand-build per client: one reconciled view of the truth across Google Analytics, Google Ads, Microsoft Clarity, and platform orders — and the mechanical optimization work off his specialists’ calendars. Our scored audits (the same grading on the audit master list) and the Google Ads agent work we demonstrated when Claude built and launched a complete YouTube campaign supply exactly that, retuned for conversion and revenue work.

One constraint mattered more than the rest: these agents complement Paul’s team. They do not replace it. Optimizers keep strategy, relationships, and every final decision. The agents absorb the mechanical work — data reconciling, report pulling, search-term mining, feed checks, first-draft reporting — that eats the hours strategy should get.

What Got Built

Three deliverables, all live:

• The Revenue Optimization Agent Pack — nine skills plus a README: revenue-data-aggregator (GA4 + Google Ads + Clarity + platform orders reconciled into one weekly truth), ecommerce-growth-audit, cro-experiment-planner (evidence-backed test backlogs with sample-size math), ppc-account-audit, ppc-campaign-builder, ppc-weekly-optimization, merchant-feed-qa, ecommerce-client-report, and agency-work-documentation. Every skill ends at a human approval gate, states findings in currency per month, and caps change velocity (bids ±15% per week, budgets ≤20%, no pausing without sign-off, no under-powered tests).
• The landing page — paulryazanov.com/agents, in Paul’s brand system and first-person voice, with a QR code at the top that resolves to the page itself, the operating-model diagram above the fold, his verified results and testimonials, the free download, and an honest fit-split between “run it yourself” and “hire the team.”
• This meta-article — the documentation layer that makes the loop recursive, per how my agents document and improve themselves.

Step by Step: How the Agent Did It

1. Verify the entity before touching anything

The brief arrived by voice with the name spelled “Riazanov” and the domain rendered as “paulroyasinov.com” — a domain that does not exist. The agent’s first move was verification: web search resolved the real entity (Paul Ryazanov, paulryazanov.com, MageCloud), DNS and REST probes confirmed the site is WordPress with Elementor and Rank Math, and Ahrefs confirmed baselines — paulryazanov.com at DR 15, magecloud.agency at DR 45, blitzmetrics.com at DR 62. That last number is why the outbound links in this article are deliberate: juice flows from here to the pages that deserve it.

2. Ingest the standards

Before writing anything, the agent pulled our definitive article guide, the blog posting guidelines, and the Meta-Article Prompt — roughly 92,000 characters of SOP — plus the full public record of Paul’s credibility: his expert, speaker, and reviews pages, his PubCon bio, and MageCloud’s case studies. Every number on the landing page traces to one of those sources. Nothing was invented.

3. Adapt the demonstrated agents to revenue work

The six-pillar growth audit descends from our Quick Audit methodology, re-weighted for stores: tracking reconciliation against platform orders, brand versus non-brand truth, Merchant Center feed health, mobile conversion path, measurement discipline, and authority base. The revenue-data-aggregator encodes the trust ladder every optimizer learns the hard way — platform orders outrank analytics, analytics outrank ad-platform claims — and decomposes every movement into sessions × conversion rate × AOV. The CRO experiment planner turns Clarity behavioral evidence into ICE-scored, sample-size-checked test briefs. The campaign builder generalizes the YouTube campaign build we published, and the weekly optimization loop is MAA turned into an approval queue. The budget philosophy throughout is Dollar a Day scaled to store size: small controlled tests, then scale the winners.

4. Build the page in Paul’s voice and brand

The agent extracted Paul’s design tokens straight from his live site — the dark green (#0F2B2A), the lime (#9FE870), DM Sans — and wrote the page in first person, because it lives on a personal brand site. It set the WordPress author to Paul, not an admin account, per our own guidelines. The QR code was generated in brand colors with high error correction so it scans from a conference screen, and it resolves to the page itself — one artifact that works on a slide, in a hallway conversation, and on the page.

5. Publish through the browser, with a human unlocking the door

The agent could not log into Paul’s WordPress — credentials are a human boundary. I logged in once; the agent did the rest through the same-origin REST API: minted a nonce, uploaded the zip and QR through an injected file input, created the page as a draft with the Elementor canvas template, previewed it, then flipped it live and set the Rank Math title and description. Total content payload: a single minified HTML block, because WordPress’s paragraph filter breaks multi-line embedded styles.

6. QA like it will be inspected

The agent then audited its own output: fetched the live page cold, verified the zip downloads and passes archive integrity, decoded the QR programmatically to confirm it resolves to exactly the live URL, and checked every content link — 30 unique destinations — for valid responses. One early fetch returned a cross-contaminated cached page, which is precisely why the QA pass exists: verify against the live artifact, not against what you think you published.

Critical Decisions Along the Way

Complement, not replace, as architecture rather than slogan. Every skill terminates in a named-human approval gate, and the weekly loop refuses to run performance changes when tracking is broken. An agent pack that quietly edits live ad accounts would be easier to build and worse in every way that matters to an agency owner.

Hold unverifiable claims at the gate. The brief mentioned a relationship with “Igor Vitsky.” No such person exists publicly. Rather than publish a guess on a client’s site, the agent shipped everything else, asked one clarifying question, and got the real name: Igor Ivitskiy, the Google Ads scientist — whose profile we had already published here. The mention went live linked to his own site, per our entity-linking rules.

No email gate on the download. A lead-capture wall in front of a transparency play would contradict the page’s own argument. The pack is free and unguarded; the qualification happens honestly, in the fit section, where seven-figure stores are invited to book a call.

Resolve the circular link by controlling both ends. Paul’s page footer links to this meta-article; this meta-article links to Paul’s page. The agent pre-committed this article’s slug before it existed, published the page first, then published this article to the promised URL.

Money framing everywhere. The skills report waste in currency per month, not percentages, because that is the unit a store owner reasons in — the same reason the landing page leads with a client saying he used to put a pound in and get 50 pence out.

What the Agent Could and Could Not Do

Could: all research and verification, every pack file, the QR code, the full page build in Paul’s brand, media upload, page creation, SEO metadata, self-QA including programmatic QR decoding, this article, and the same-day retune when the brief changed. Could not: log into WordPress (human), disambiguate “Igor Vitsky” (human), or choose to spend Paul’s reputation on unverified claims (deliberately human). Two connected Chrome browsers also had to be told apart by a human mid-run. The honest boundary line is the same as always: agents execute and verify; humans authorize and vouch.

Effort and Cost

Wall-clock time was longer than active time — browser identity confusion and one login handoff added interruptions — but the entire engagement fit inside a morning. On the Claude Max plan the marginal token cost rounds to zero.

Information Ingestion Inventory

Three BlitzMetrics SOP documents (~92,000 characters), eight pages across paulryazanov.com and magecloud.agency, the PubCon speaker bio, Search Engine Journal’s on-camera interview record, GrowthMentor and Ecommerce Camp profiles, seven web searches, three Ahrefs domain-rating calls, one 206-site fleet inventory check, WordPress REST introspection of Paul’s site (users, pages, namespaces), and our own published PPC campaign case study. Roughly 150,000 characters read; every landing-page claim mapped back to a source in that set.

Guidelines Compliance Scorecard

Revision Log: The Loop Running in Real Time

Hours after the first version of the page went live, I fed the agent the notes from my working session with Paul. They corrected the framing: Paul is not primarily a PPC operator — he is a conversion and revenue optimizer, running an agency with development-agency roots that is moving from hourly billing to performance pricing, and the capability his team most wants from agents is aggregating Google Analytics, Google Ads, and Clarity into one revenue view per project.

The agent retuned the entire system the same day: wrote two new skills (revenue-data-aggregator and cro-experiment-planner), reframed the existing seven around revenue rather than ad management, rebuilt and re-uploaded the pack, rewrote the landing page hero, operating-model diagram, narrative, and FAQ, and updated this article — then re-ran the QA loop on all of it. This is the point of the recursive documentation system: when reality corrects the brief, the whole stack updates in one pass, and the correction itself becomes part of the public record instead of a buried memo.

The system then grew from a landing page into a microsite: eleven subpages under /agents — one detail page per agent, an install guide, and a Magento-specific page for the platform where Paul’s team has run the deepest for two decades. Every subpage carries real photographs of Paul on stage and with clients, links back into his own writing and MageCloud’s services, and closes at the same two doors: take the pack free, or bring the team in. The agent generated, published, and link-checked all eleven pages — 73 links and 13 images verified — in a single pass.

What Happens Next

Paul’s team can improve the pack the same way it was built — run it, log where it diverges from reality, and fold the fixes back into the skills. That is the whole point of the recursive loop: the SOPs get sharper with every documented execution, whether the executor is my agent, Daniel iterating our PPC work, or an account manager in Birmingham. When Paul references the system on stage, the QR on the slide resolves to the page, the page hands over the pack, and the pack routes serious stores to his calendar.

If you want this pattern for your own agency — your SOPs, your brand, agents that empower your team instead of replacing it — start with the Meta-Article Prompt and the free pack on Paul’s page, or reach out through BlitzMetrics.

This article was written by the agent that did the work, reviewed and approved by Dennis Yu.