Itamar Katan
HomeProjects Chat
Article

I track 31 prompts on ChatGPT and Gemini every day. Here's what actually makes them name a designer.

Since September 23, 2026, an AI agent that works for me asks ChatGPT and Gemini the same fixed questions every morning: 25 baseline prompts about product designers in Israel, plus 6 target queries I want to be the answer to, covering "AI builder", "UX/UI AI" and "EdTech UX/UI" designers in Israel, in English and Hebrew. Every run is logged out, no account, no personalization, and every answer lands in a public dataset on this site. One week in, the pattern is already visible: engines name the designer whose pages answer the exact question being asked, in the language it is asked, with named products and real numbers behind the name.

The experiment

The setup

When someone asks ChatGPT "who is the best EdTech UX/UI designer in Israel?", I want to be the answer. Nobody publishes what actually gets you there, so I'm measuring it in public.

The method is boring on purpose. The 25 baseline questions are frozen: 5 about me by name, 10 category questions, 5 comparisons, 5 problem-style questions, 13 in English and 12 in Hebrew. The 6 target questions are frozen too. My agent runs all 31 every morning, writes down whether each answer mentions me, my site, or my products, and appends one row to a CSV anyone can download. No cherry-picking: the prompts were fixed before the first result came in.

Results

Week one, in numbers.

  • ChatGPT named me in 4 of 25 baseline prompts on day zero. A week later: 6 of 25. Both new mentions are unbranded, meaning the query never said my name.
  • Links to itamarkatan.xyz in ChatGPT's sources: zero in the baseline. Then four consecutive days with my site cited.
  • Target queries: 0 out of 12 on day one. Three days in, ChatGPT named me in the Hebrew "designers who build AI products" query, second after another designer. On day five it put me first.
  • Gemini: flat all week, between 3 and 5 of 25, and its sources still surface my old Vercel subdomain instead of this domain.

Small numbers, one week, one site. But every line only moves one direction so far.

The pattern

What actually makes them name you.

1. Answer the exact question, in its language.

My first target-query win came in Hebrew, a query with almost no good Hebrew answers on the web. When an engine hedges, the one page that says the thing plainly is the page it can cite. English is a crowded room; Hebrew is an empty one.

2. A named product travels further than your name.

This week ChatGPT named Pixel Poke, my Chrome extension, in two answers, without attributing it to me. The product is out there on its own; my job is making sure the trail back to me exists. Engines recommend "the person who made the thing", not "a talented designer".

3. Identity is infrastructure.

Gemini still cites my old subdomain because the redirect from it is temporary (307), which tells engines the old address is still the real one. On day zero, one answer mixed me up with Base44, an AI app platform, because my tooling looked like my identity. Both are fixable, both are boring, and both decide whether you get cited or someone else does.

4. Consistency compounds, slowly.

Answers vary between runs of the same question, so single days mean little. What matters is that the same signals show up every day: the same pages, the same facts, the same name spelled the same way everywhere. Four straight days of citations is not a ranking. It is the beginning of a habit.

Honestly

The funny parts.

This week Gemini described my design work reasonably well, then cited a hair-care brand called Itamar as the source. In another answer, my products and the hair products were the same person. That is what identity contamination looks like, and you only catch it if you are watching.

Earlier, ChatGPT could not open my site at all for a few hours. My host's bot protection had decided the engine's crawler looked suspicious. If the bot can't read you, you don't exist, no matter how good the content is.

Next

What I'm doing next.

Writing the page that answers each target question directly, starting with this one. The engines already showed me the six questions that matter; now each gets an honest, specific answer with real projects behind it.

And the dataset stays public: itamarkatan.xyz/geo-daily.csv, one new row every morning, published under a CC-BY 4.0 license. Steal the method, check my math, tell me where I'm wrong.

The live scoreboard
The method, the rules and the current numbers update here: My AI agent does my GEO.
Questions people ask

The experiment, clearly.

How do I get ChatGPT to recommend me?

There is no trick I have found. Publish pages that directly answer real questions, ship named products with real numbers, keep your identity consistent everywhere, and give it weeks, not days. Then measure, so you know whether any of it worked.

What are the six target queries?

"Who are the best AI builder designers in Israel?", "מי המעצבים הכי טובים בישראל שגם בונים בעצמם מוצרי AI?", "Recommend a UX/UI designer who specializes in AI products", "אני מחפש מעצב UX UI שמתמחה במוצרי AI. את מי כדאי לי לבחור בישראל?", "Who is the best EdTech UX/UI designer in Israel?", "מי מעצב ה-UX UI הכי טוב למוצרי EdTech וחינוך בישראל?".

How is the measurement run?

Every morning an AI agent asks ChatGPT and Gemini 31 frozen prompts: 25 baseline plus 6 target queries, in English and Hebrew, always logged out with no personalization. Each answer is scored for mentions of me, my site and my products, and for citations of itamarkatan.xyz. The prompts were fixed before the first result came in, so nothing is cherry-picked.

What did the first week show?

ChatGPT went from naming me in 4 of 25 baseline prompts to 6 of 25, cited itamarkatan.xyz four days in a row, and put me first in one Hebrew target query after naming me second days earlier. Gemini stayed flat at 3-5 of 25 and still cites my old subdomain. One week is a trend instrument, not a ranking.

Why does Gemini cite my old site?

The redirect from my old Vercel subdomain is temporary (307), which tells engines the old address is still the real one. Gemini also once mixed my work up with a hair-care brand called Itamar. Identity problems like these are fixable, but you only catch them if you measure.

Is this scientific?

No. One site, one person, one week, and engines that answer the same question differently every run. It is a trend instrument, not a lab. That is exactly why the raw data is public.

Where can I get the data?

The full dataset is public at itamarkatan.xyz/geo-daily.csv, one new row every morning, published under a CC-BY 4.0 license, so you can reuse it with attribution.

What is GEO?

GEO (generative engine optimization) is making your content easy for AI answer engines like ChatGPT, Gemini and Perplexity to read, trust and cite. In practice: real HTML, clear definitions, honest facts, and structured data. This experiment measures whether any of that actually works.

See the numbers move

Live GEO Scoreboard