Let an agent draft SEO updates on a schedule, with a human editorial gate, and log what broke
An experiment log from a site run by a scheduled agent that pulls Search Console data and drafts updates for human approval. Traffic grew a little; three silent failures are written up.
Evidence: The author reports this. We have not checked it beyond reading the source.
The business problem
A one-person content site wants to keep publishing fresh, data-driven updates and improve search results without spending hours a week on it.
What was tried
A scheduled job pulls fresh Search Console, analytics and usage data each day and runs a light analysis. When something is worth saying, such as a ranking shift or a material number change, the agent drafts an update and puts it in a queue. It cannot publish by itself: a person approves or kills each item, and handles judgement-heavy work such as the newsletter, at about 30 minutes a week by the author's estimate. Approved items go to the main branch, the host deploys automatically, and a smoke test against the live page triggers an automatic rollback on unexpected output. The author set two kill criteria to review by 1 September 2026: click-through on definition queries reaching 1.5 percent, and automation reliability around 95 percent.
What was reported (mixed)
After three weeks: about 1,300 unique visitors a month and 2 email subscribers. By the 20 July update: about 2,500 visitors and 5 subscribers (three traced to signup placements added on 2 July). Definition queries rank but click-through stayed under the 1.5 percent target.
Limitations
Author unnamed and no AI model or framework is named; costs are not reported. Numbers are tiny and self-reported, and traffic growth is not tied to specific changes. Three failures are listed honestly: a section cut using a three-month average that hid recent growth, a signup form change that the email service rejected so signups failed silently for six days (found only by a real end-to-end submission), and a stale production build after a deploy outside the pipeline. The automation also had a two-day outage when the host machine slept. The experiment was not finished at the time of the update.
What you need
Search Console and analytics access, a scheduler, a Git host with auto-deploy, an email provider API, and an agent or model to draft. The source names no specific model.
Sources
- tokenmaxxing.com experiment log ↗ Firsthand write-up, published June 28, 2026
Source published: June 28, 2026. Last reviewed here: October 16, 2026. Spot a mistake? Tell us.
Tools in this workflow
- Google Search Console: Daily ranking and click dataTry Google Search Console
Related workflows
Use an OpenClaw agent for price alerts and weekly research digests, and expect about a dozen silent failures in a month
A user's 30-day test of a local OpenClaw agent for market alerts and weekly research digests for a newsletter, reporting about 15 hours of set-up, hours saved on monitoring, and recurring silent failures.
Monitor your brand and competitors on X by sending mentions to Grok for sorting, then review a daily digest
A managed-IT firm describes pulling mentions of a brand and competitors from the X API, having Grok sort each one into complaint, praise, comparison or noise with a rough sentiment score, and posting a digest for a person to review.
Run a small SaaS feedback loop with Claude: a weekly PostHog survey digest, weekly release summaries and browser-based product testing
A course-creation software company asks new free users to rate their first result, has a scheduled Claude job turn the answers into a weekly issue list, has Claude summarise each week's releases for the team, and has Claude test the product in a browser like a new customer.
Price Manus tasks in credits before you automate them: a sales deck, landing page, SEO audit and daily social report
A Manus tutorial lists what common small-business tasks cost in credits, including a sales-data deck, a landing page, an SEO audit and a daily social monitoring report, and shows which would exceed a standard plan.