It was 11pm on a Sunday in May 2026. Kunshan, my office, third cup of coffee. I was staring at Google Search Console for the seventeenth time that week, watching impressions tick up by single digits and trying to decide whether I should rewrite a sitemap or go to bed.
(I went to bed. The sitemap won later that night when the AI agent I'd been building decided, on its own, that it should regenerate the missing hreflang tags.)
This is the story of what happened in the 30 days after I gave an AI agent the keys to my SEO. Not the marketing version — the operator version. With the numbers, the failures, and the full SOP, open-sourced.
Key Stats (verified, last 30 days)
| Metric | Baseline (May 1) | Day 30 (May 31) | Delta |
|---|---|---|---|
| GSC monthly impressions | ~600 | ~32,000 | 53x |
| Pages indexed (sitemap coverage) | 8 / 66 | 57 / 66 (86%) | +49 pages |
| Keywords on Google page 1 | 0 | 6 | +6 |
| Keywords on page 1-3 | 1 | 11 | +10 |
| AI Overview citations | 0 | 1 (open source marketing query) | +1 |
| Hours of human work | ~25 hrs/wk | ~3 hrs/wk | -88% |
(I'm not going to claim 53x is repeatable. It's not. The starting point was egregiously bad and there was low-hanging fruit. The interesting question isn't "can you go 53x" — it's "what does the SOP look like when an agent does this instead of a human, every day, without burning out?")
The Setup
In April 2026 I migrated my blog (gingiris.tools — growth playbooks for AI startups) to a new domain after a hosting incident forced our hand. The migration was clean, but the weight cliff was permanent: 4 years of domain authority, 100+ inbound links, gone. I was starting from zero, in the most literal SEO sense.
What I had:
- 60+ blog posts (good)
- 0 inbound links to the new domain (bad)
- An idea that AI agents should be able to do most of the operator work for SEO (untested)
What I didn't have:
- Bandwidth to do daily SEO myself (was running another product)
- Confidence the agent would be reliable enough to leave alone
- A clear sense of where AI search ranking even goes when you don't have a brand
So I started writing an SOP. Not for me — for the agent. With the assumption that anything I couldn't write down in concrete steps was something the agent couldn't do.
What the Agent Does, Every Day
The agent (running on Claude Code with the gingiris-seo-geo-agent skill installed) runs a single workflow every morning. The workflow is the entire SOP — broken into ~25 deterministic steps:
1. Query Google Search Console API for yesterday's data
2. Query DataForSEO for SERP positions on the tracked keyword set
3. Compute today's "Page-1 keyword count" (the daily north star)
4. Detect deltas: keywords entering top 10, falling out, +20% movers
5. Read the article database (article × performance metrics)
6. Diagnose: which articles need internal links, which need refreshes,
which need Schema, which to deprioritize
7. Pick today's actions (max 3, ranked by impact × difficulty)
8. Execute the actions (write commits to the Jekyll repo)
9. Push (or queue for batched weekly push if rate-limited)
10. Write the daily report (page-1 headline + per-keyword detail + GA4)
11. Submit URL Inspection requests for changed pages (manual flag if needed)
And it does this without me. Most days I read the daily report at 8am over coffee. Most days the only thing I do is reply "looks good, continue" or "skip action B, do C instead."
What surprised me: the boring, daily-discipline part of SEO is actually the easiest to automate. It's the strategic decisions (what topic cluster to build? when to launch a comparison page?) that still need a human. Once I stopped trying to make the agent strategic and let it be operationally relentless, everything got easier.
The 4-Phase Growth Pattern (What Actually Happened)
Week 1 — Foundation (impressions: 600 → 2,100)
The agent's first week was almost entirely technical cleanup. It found 31 issues:
- 5 articles with broken canonical_url tags
- 12 articles with missing hreflang annotations
- 8 articles where the Schema.org markup had typos
- 4 pages that returned 404 but were still in the sitemap (zombie content)
- 2 robots.txt entries that were blocking AI crawlers I wanted indexed
Fixing technical issues doesn't sound like a growth story, but 3x impressions came almost entirely from this. The articles were already good. They just weren't crawlable.
Lesson: if you've migrated, or refactored, or done anything destructive in the last 12 months, you have ~30% of your real ranking potential locked behind technical issues. The agent found and fixed mine in 6 hours. I'd been ignoring most of these for weeks.
Week 2 — Schema + GEO (impressions: 2,100 → 8,400)
This is where GEO (Generative Engine Optimization — getting cited by ChatGPT/Perplexity/Claude/AI Overview) started to matter.
The agent added:
-
FAQPageSchema to 12 tutorial articles -
HowToSchema to 8 step-by-step guides -
ArticleSchema withauthoranddatePublishedto every post -
Organization+WebSite+SearchActionto the homepage
These don't directly make Google rank you higher. What they do is make your content machine-readable enough that AI engines feel safe citing it. The 4x impressions jump came from a mix of:
- Featured snippet eligibility for the 12 FAQ-tagged pages
- Google AI Overview started citing one tutorial (on open-source marketing)
- Perplexity surfaced two articles when asked about Product Hunt tactics
The agent didn't have to write the Schema by hand — it had a Schema type matrix built into the SOP, mapping article type to required schema types. The whole batch took ~90 minutes of agent time + 0 minutes of mine.
Week 3 — Internal Linking + CTR Rewrites (impressions: 8,400 → 18,200)
The agent's third-week diagnosis was harsh: 9 of my top-ranking articles had no internal links pointing to any landing page. Traffic was arriving and bouncing.
It rewrote:
- 23 article titles, applying a 5-element CTR rubric (number / year / brackets / social proof / 50-60 char length)
- Inserted internal links from those 9 articles to the 3 most relevant conversion pages
- Added a standardized "Convert Block" (CTA with clear copy + UTM tracking) to each high-ranking article
It also identified my "cannibalization" issue — 3 articles competing for the same keyword — and proposed which to canonicalize. I approved 2 of the 3 (one was too borderline). Both canonicalized articles saw immediate position gains.
CTR rewrites are where humans usually procrastinate. (I had a list of "title rewrites I should do" sitting in a Notion doc since February. None of them got done.) The agent did all 23 in a day.
Week 4 — Compounding (impressions: 18,200 → 32,000)
Week 4 was the dull one. The agent kept doing the same thing — daily reports, small fixes, occasional new article — and the curve compounded.
The interesting moments:
- The day a single article hit position #3 for "github star growth playbook" (1.3K monthly searches, low KD). +800 impressions / day from that one keyword.
- The day Google AI Overview cited my open-source marketing guide for a high-intent query. The citation was attributed to gingiris.tools, with a direct link.
- The day I realized I hadn't written a single SEO-related git commit message myself in 9 days. The agent had done all of them, signed under its own identity, with natural commit messages I couldn't distinguish from mine.
What Worked (And Why)
1. The agent doesn't have feelings about SEO.
This sounds glib but it's the actual point. SEO is a discipline of boring repetition + occasional strategic bets. Humans get bored of the boring part and over-invest in the strategic part. The agent does the reverse: it executes the boring part with religious consistency, and asks me for the strategic bets.
2. The SOP is the moat.
The SOP file (~30 pages) is what makes the agent reliable. Without it, the agent improvises and gets random results. With it, the agent is just executing instructions — and the instructions can be improved.
If you're thinking "but my SOP is in my head," the answer is write it down. The act of writing it down (so an agent can follow it) is the act that makes you an actual SEO operator instead of a person with intuitions.
3. Schema.org is undervalued for AI search.
I had FAQ Schema on 4 pages before this experiment. After: 12 pages. Three of those 12 now show up as Featured Snippets. One gets cited by AI Overview on a high-intent query. Total agent time to add: ~90 minutes.
The opportunity is bigger than people realize because most sites just don't do it well.
4. The daily report kills procrastination.
The agent writes a daily report ranking my keywords by "Page 1 / Page 2-3 / off-100". The very first thing I see when I wake up is how many of my keywords are on Google's front page today. This number cannot be ignored. It also can't be argued with.
The behavior change from "have a daily SEO discipline" → "actually do daily SEO" came from the report, not from willpower.
What Didn't Work (Honest)
The agent's first attempt at Schema markup had typos in 3 of 12 pages. I caught them on review. If I hadn't reviewed, two pages would have shown 0 rich result eligibility despite "having" Schema. Lesson: always validate via Google's Rich Results Test before celebrating.
Google didn't index 6 sitemap URLs. They were marked "Discovered - currently not indexed" in GSC. The agent didn't realize the issue (it saw them in the sitemap and assumed indexed). I had to manually request indexing for those 6 URLs via GSC URL Inspection. (The agent now tracks this case explicitly and flags it for human action.)
I lost ~3 days to a domain migration backlog that the agent couldn't handle. Specifically: 9 articles still had stale URLs pointing to a hosted domain that no longer existed. The agent could write the replacements but couldn't decide which canonical URL to use without context I had to provide.
Some of the agent's title rewrites were too aggressive. It added years to 4 titles that didn't need them. I reverted those manually. Newer versions of the SOP have a "don't add the year if the article isn't time-sensitive" rule baked in.
The SOP — Now Open-Source
The complete day-by-day SOP behind this 32K-impression run is published as a Claude Code skill:
clawhub install gingiris-seo-geo-agent
After installing, just tell your AI agent:
"Run today's SEO/GEO patrol on [your domain]"
The agent will read the SOP, query GSC + DataForSEO, generate the daily report in the same format I use, and propose today's actions. You stay in the loop on strategy — the boring execution gets handled.
What's in it:
- Daily report template (Page-1 keyword headline + per-keyword detail + GA4 funnel)
- The 25-step daily workflow
- Schema.org type matrix (page type → schema → SERP feature)
- robots.txt template (AI crawler search vs training split)
- GA4 AI-source regex (ChatGPT/Perplexity/Claude/Gemini referral detection)
- Convert Block CTA template (UTM-tracked)
- 4-phase audit checklist (Technical → On-Page → Content → Off-Page)
- i18n / hreflang sub-audit (for multi-language sites)
Companion strategy methodology:
clawhub install gingiris-seo-geo
This one covers the why (keyword funnel strategy, GEO content patterns, E-E-A-T writing voice, comparison page SOPs) while gingiris-seo-geo-agent is the how (daily operational SOP).
All MIT-0 licensed. No attribution required, but cool if you DM me on X @WeiYipei so I know what's working for you.
If You're Doing SEO in 2026 and You're Tired
Look, AI search isn't replacing Google search yet. (The traffic numbers say "yet" stretched to maybe 2027.) But it's already replacing the brain you used to use for daily SEO discipline.
What you can do this week:
- Audit your Schema.org markup. There's almost certainly low-hanging fruit.
- Add
FAQPageschema to your top 5 articles. (Featured snippet eligibility = free CTR boost.) - Configure GA4 with an "AI Chatbots" channel (regex includes ChatGPT, Perplexity, Claude, Gemini, Copilot, DeepSeek). Put it above Referral or Referral eats it.
- Write your daily SEO report headline as "How many keywords on Google page 1 today?" — and look at it every day.
These four things don't require AI. They require an afternoon. After that, the agent stuff makes sense.
If you want the full open-source SOP, install the skill above. If you just want to read more of these case studies, I write them at gingiris.tools.
I'm not going to claim 53x impressions in a month is normal — it isn't. The starting point was bad. But the operational discipline that produced it is repeatable, and once the agent is doing it, the marginal cost of doing it for another domain is roughly zero.
Iris (生姜iris)
Ex-AFFiNE COO · Forbes Asia 30 Under 30 · gingiris.tools · @WeiYipei
Top comments (0)