My AI agent does my GEO
while I walk rice terraces in Vietnam.
I turned Instinct, my personal AI assistant, into my GEO agent. It audits my portfolio, writes the plan, ships the fixes as GitHub pull requests, and I merge them from my phone somewhere between two rice paddies.
The office this week. I sent this from the deck one morning while my agent was busy with my site.
What an AI GEO agent is.
An AI GEO agent is an AI assistant with access to your site's code that audits how AI engines read your content, then fixes it through changes you review. Mine is Instinct, working in my repo under its own GitHub account.
I'm a product designer. I went freelance in July 2026 and wanted AI engines to describe my work correctly when someone asks about EdTech design or AI products. I also wanted to see Vietnam. So I split the work: I travel, Instinct does the GEO, and we meet in pull requests.
How I set it up.
I created a GitHub account for my agent, Itamar-Dev01, and added it as a collaborator on my portfolio repo. That's the whole setup. It's the same thing I'd do for a freelance developer, and I kept the same rules:
- It never pushes to main. Every change is a branch and a pull request.
- I merge. Nothing goes live without me pressing the button.
- Before anything creative goes on the site, I approve the text itself.
- No invented metrics. If there's no data behind a number, the number doesn't go in.
I talk to it on WhatsApp, in Hebrew. It writes the site in English. Somehow this works.
What the agent actually did.
- 01Audit. It read the whole repo and the live site, then told me what an AI engine actually sees when it visits: a single-page React app whose real content showed up only after JavaScript ran. The words crawlers got came from a hidden SEO layer, which is exactly the kind of thing that ages badly.
- 02Plan. A short GEO plan: real HTML on every route, one citable sentence per project, FAQs written as the questions people really ask, structured data, and a proper domain. No keyword stuffing, no made-up numbers.
- 03First article. It rewrote my Pixel Poke case study as a citable article: definition first, then how it works, the design decisions, what shipped and what I have not measured yet. Then RecordAI, same format. I read both drafts before a single line hit the site.
- 04Prerendering. It added build-time prerendering, so all 22 routes now ship real HTML with their own headings and body text, and it removed the hidden SEO layer. Same site, same look, but now a crawler reads what a person reads.
- 05Domain. It suggested a short name-only domain. I bought itamarkatan.xyz and connected it in Vercel myself. It moved every canonical URL, Open Graph tag, sitemap, robots and llms.txt entry and all the structured data over to the new domain.
- 06Case studies. One definition line, a "Questions people ask" block and JSON-LD on each case study, without touching the body copy. It also removed a former client's name from a project page and blurred real names in the screenshots, because some things should not be citable.
- 07Housekeeping. It even fixed the credit link in my open-source design tool so it points to my site and not an old preview URL.
Who does what.
Reads the code and the live site, finds problems, researches facts, writes drafts, builds and tests locally, opens pull requests, takes screenshots at desktop and mobile size, and checks the live site after a deploy.
Decides what goes on the homepage, approves every creative draft, says no to things that aren't me, and merges.
Makes up numbers. When we don't have data, the page says so.
The funny parts.
The first thing the agent shipped got blocked. Vercel refused to deploy commits written by an account that isn't on my Vercel team, which is fair, I'd do the same. The fix was very low-tech: I merge with a merge commit, the commit is mine, and Vercel is happy.
My whole review process now happens on a phone, in between rice terraces. That turned out to be a good test. The agent sends me screenshots at desktop and phone size, and if a page makes sense on a phone screen in the sun, it makes sense.
What I learned.
GEO is mostly honesty in a format machines can read. A clear first sentence, real HTML, specific facts, and answers to the questions people actually ask. The agent is fast at all of it. What it needs from me is judgment: what's true, what's mine to share and what I sound like.
The best way to work with an AI agent as a designer is the way you'd work with a good developer. Clear rules, small reviewable changes, and you keep the final say.
Day 0: the engines don't recommend me. Yet.
I promised results only when I could show the method. So here is both. Every day, my agent asks AI engines a fixed set of questions and writes down whether they mention me, my site, or Pixel Poke. Same questions, logged out, no account personalization. This is the starting picture, from September 23-24, 2026.
- 31 fixed questions. A frozen set of 25 (5 about me by name, 10 category, 5 comparison, 5 problem-style; 13 in English, 12 in Hebrew), plus 6 target questions for the queries I want to win: "AI builder", "UX/UI AI" and "EdTech UX/UI" designers in Israel, in English and Hebrew.
- Same conditions every run. Logged out, fresh session, no personalization. Perplexity and Claude join when sign-ins are sorted; today they cover 16 of 25 and 0 of 25.
- Simple scoring. For every answer: does it mention me, itamarkatan.xyz, or Pixel Poke - yes, partial, or no - and was one of my own pages cited?
- Trends, not single cells. Answers vary between runs of the same question, so I compare weeks, not days.
| Question group | ChatGPT | Gemini | Notes |
|---|---|---|---|
| About me by name (5) | ChatGPT: 4 of 5 know me | Gemini: 4 of 5 know me | Every source is third-party (LinkedIn, Behance, the Chrome store). itamarkatan.xyz was cited zero times. |
| Category, comparison and problem questions, no name (20) | ChatGPT: 0 mentions | Gemini: 1 mention | Gemini named me among leading EdTech UX designers in Israel - citing someone else's site, not mine. |
| Target questions: AI builder / UX UI AI / EdTech UX UI in Israel (6 questions, EN + HE) | ChatGPT: 0 of 6 | Gemini: 0 of 6 | Who gets named instead: three freelance designers for "AI builder", and one studio for both "UX/UI AI" and EdTech. The names stay in my private log. |
| Pixel Poke in generic tool questions | ChatGPT: 0 mentions | Gemini: 0 mentions | Not on the radar for "edit CSS on a live site" questions. |
Two bugs fell out of day 0. ChatGPT once confused me with a different Itamar Katan (of Base44), and once claimed it couldn't open itamarkatan.xyz at all - while the site loads fine, including for OpenAI's crawler. Both are the kind of thing you only catch by asking every day.
So the honest starting point: 0 of 12 target answers mention me. The fixes my agent ships - the schema, the freshness dates, the about pages - exist to move exactly these numbers. When they move, I'll update this table and say what changed.
Working with an AI agent on GEO.
Can an AI agent do GEO for you?
Yes, most of the work. In my case an AI agent audited the site, wrote the plan, implemented prerendering, rewrote case studies as citable articles and added structured data, all as pull requests. What it cannot do is decide what is true about your work or how you want to sound. That part stays with you.
What is Instinct?
Instinct is a personal AI assistant that connects to your apps and accounts and does real work for you. You talk to it by text or voice. I use it as my GEO agent: I message it, it works in my GitHub repo through its own account, and I review and merge.
What is GEO?
GEO (generative engine optimization) is making your content easy for AI answer engines like ChatGPT, Perplexity and Google's AI answers to read, trust and cite. In practice it means real HTML, clear definitions, honest facts, FAQs and structured data.
Is it safe to give an AI agent access to your code?
Give it the same deal you would give a new contractor. Its own account, access only to the repos you choose, no pushing to the main branch, and every change as a pull request you review. I can revoke its access anytime.
Do you need to know how to code?
Not to direct it. You do need to be able to look at a preview and say "this is not me."
How long did it take?
About two days of calendar time, most of which I spent traveling. My part was a few minutes at a time: approving drafts, answering questions and pressing merge.