Harsehaj's blogEO: The SEO Agent That 6x'd Browserbase Traffic

⬅️ Back to Videos

🎥 Harsehaj’s blogEO: The SEO Agent That 6x’d Browserbase Traffic

Alex Lieberman with Harsehaj (Browserbase). Duration: 23 min

Timestamps

  • 0:00 Introducing blogEO
  • 0:20 Before: the big problem, no measurement
  • 4:00 Step 1: run the AI audit
  • 9:18 Step 2: AI recommends surgical edits
  • 16:51 Step 3: avoiding AI slop
  • 19:09 Step 4: running the full blog machine
  • 21:10 Results: 5.8x impressions, page two to page one

A Browserbase intern rebuilt the company blog pipeline and nearly 6x’d search traffic. The surprise is not the AI. It is the restraint: her agent is architecturally unable to publish anything, and every meaningful edit waits on a human in Slack.

  1. She started from an embarrassing baseline, and said so out loud. No measurement, five scattered sources nobody joined by hand, 70 percent of traffic from 14 posts and half from just 5. Her top Google posts were invisible to LLMs entirely. The lesson I keep coming back to: audit before you generate anything.
  2. The agent is designed so it cannot hurt her. Their internal agent BB can never write content server-side, so the worst outcome is a bad suggestion, never a bad publish. Guardrails and skills cannot guarantee good behavior, she says, so the safety lives in the architecture instead of the prompt.
  3. Hygiene goes first, and robots do robot work unsupervised. Broken links and empty SEO fields auto-fix with zero human approval. The first run backfilled 74 posts missing SEO fields entirely. Humans spend their clicks only on judgment calls.
  4. Opportunity gets scored in clicks, not vibes. Three metrics: recovered clicks lost versus the 28-day baseline, CTR delta against the page-one average, and rank upside from page two or three to page one. One post at position 9.4 carried a 7,400-click opportunity. Real click losses outrank all estimates automatically.
  5. Cheap checks run on everything, expensive checks only on the top 15 plus a 3-post buffer. Full-doc drift verification against live docs costs real tokens, so near-miss posts get verified while the long tail gets hygiene. Token discipline is a design feature, not an afterthought.
  6. Small edits moved the needle faster than she expected. Title and description drive click-through almost alone, and cross-linking lifts intent signals. Her fix for the measurement problem is my favorite detail: a 28-day cooldown after every publish with readings at +28 and +56 days, because otherwise you would re-edit the same post weekly and learn nothing.
  7. She drew the AI-writing line where a writer would. Most Browserbase posts are handwritten, so positioning drift gets flagged to a person with no approve button attached. Her words: that is a you thing to do. She writes a daily newsletter herself, and that restraint reads as professional pride rather than caution.
  8. New posts come from three skills, not one prompt. A writing skill holds voice, cadence, and the TLDR checklist. An SEO strategy skill holds personas, clusters, and red-light topics like never bashing competitors. A generate-blog orchestrator picks between them. Anything truly novel gets sent back to humans for handwriting.
  9. Four hard gates stand between a draft and the blog. Strategy fit, TLDR structure, code provenance with every snippet verified against real docs through the fetch API, and a cannibalization check so two posts never chase one query. Losers get funneled into the audit queue. No hallucinated code has a path through.
  10. Publishing happens in Slack with names attached. Generation cards go out Mondays, audit cards Tuesdays, each with approve, edit, skip, or discard. Approving publishes immediately and enrolls the post in a growth curve from zero. The record shows who approved it, because that responsibility is the whole point. And her honest coda: she also hand-wrote 23-plus posts, so the 5.8x impressions and 9.8x page-one queries belong to the system plus the writer, not the agent alone.

Related TMFNK Content

Crepi il lupo! 🐺