How We Got AI to Find INTENTION: A Taipei Creative Studio's GEO Experiment Log
The short version: if you want AI to recommend you, the most useful thing isn't writing more articles "for AI." It's making sure AI knows who you are, and then getting other people to keep mentioning you. That's actually the same old method I've always used to run INTENTION (識聚有限公司): keep saying clearly what you do, and let the people who agree come find you. The only difference this time is that some of the listeners are machines.
This piece records every step we've taken since September 2026, the numbers we measured, and what I predict comes next. The numbers get updated once a month.
1. Why do this now
Search looks different. In 2024, a research firm predicted that traditional search volume would drop 25% by 2026 because of AI chatbots [1]. This is the year that forecast comes due. Search hasn't disappeared, but clicks are falling: when an AI summary appears, only 8% of people click a traditional search result, compared with 15% when there's no summary [2].
Even when people search your brand name, the first thing they see may be AI. Figures from October 2026 show that 82.91% of brand-name searches on desktop in the US bring up an AI summary [3].
People in Taiwan are using it too. In 2025, 43.19% of people in Taiwan had used generative AI in the past three months [4]; Gemini passed 1 billion monthly active users in August 2026 [5]. A September 2026 Taiwan survey, as reported in the media, put ChatGPT usage at 75.2% and Gemini at 56.3% (the original report isn't public and the sampling method is unclear, so treat it as a reference only) [6].
So the question isn't "will AI search replace Google." The question is: when someone asks AI, what does it say about you.
2. Starting point: AI didn't know us, and got us wrong
On September 25, 2026, we asked three AI services 49 questions of two kinds. Some named us directly, like "What kind of company is INTENTION?" Some didn't, like "Can you recommend an MV production company in Taipei?"
Date | ChatGPT | Gemini | Perplexity | Unbranded questions |
|---|---|---|---|---|
2026-09-25 | 4% | 6% | 2% | 0% for all three |
Every time we got mentioned, it was because we had named ourselves in the question.
On October 7 we ran another test: no web access, so the AI had to answer 11 questions about us using only what it remembered. Not one of the 11 came back right. It said INTENTION is a software development consultancy, that NOUS ASIA is a business consulting firm in Hong Kong, and that INTENTION Made is a brand launched by a Japanese design team. People on our team were also mixed up with other people online who share their names. In fact, INTENTION (識聚有限公司, INTENTION CREATIVE CO., LTD.) is a creative strategy and visual production company in Taipei, NOUS ASIA (冉冉創意股份有限公司) is our production arm, and INTENTION Made is the set of tools we build and use every day ourselves.
This is more trouble than "not knowing us." When AI doesn't know something, it doesn't necessarily say so. It often fills the gap with something that sounds reasonable.
3. How I reasoned it through
AI trusts "what other people say about you," not what you say about yourself. A study of 75,000 brands found a correlation of 0.664 between how often a brand is mentioned online and how often it shows up in AI summaries. For backlinks it was only 0.218 [7]. In another study, the strongest correlation was with mentions on YouTube (about 0.737), while the number of pages on a website barely mattered (about 0.194) [8].
AI cites video most, and long video at that. In August 2026, the source cited most often across eight AI platforms was YouTube (212,500 times) [9]; 94% of the YouTube videos cited were long-form [10]. But it wears off fast: a video's chance of being cited drops by about 43% within a month [11].
So this isn't a one-time job. Being mentioned takes consistency. Wrong information has to be fixed first. And the data that lets AI match things up (who is who, who made what) needs to be in order first.
I only sort things into two piles: what we do and what we don't. Below is what we did, and what we deliberately didn't.
4. What we did (in order)
Let AI read the words (September 25). Our website is mostly images, and the homepage's raw text was only about 200 characters. Many AI tools don't run code when they read a page, so what they see is almost blank. We rewrote the titles, descriptions, and machine-readable structured data for 34 pages, and put an llms.txt file in the site's root directory. It explains in Chinese, English, Japanese, Korean, and Thai what we do, when it makes sense to recommend us, our key work, and how to reach us.
Fix the wrong turns AI was taking (September 25). We found ChatGPT citing a URL that doesn't exist, intention.ltd/artist, 6 times in total. We redirected it to the real directors page. The directors page had also been set to "don't let search engines index this," so we opened that up too.
Give every person an identity (October 7). For AI to answer "Who is Pinbo?", it first has to know that Pinbo, 黃品柏, and Pinbo Huang are the same person, and that he's a director at INTENTION. We built machine-readable identity data for 8 directors, then linked the director on each of 14 project pages to that one identity. We linked the company's founder too. Directors with complete data went from 0 to 8.
Connect third-party press (October 7). Focus Taiwan (Taiwan's Central News Agency English service) reported in May 2026 on Pinbo directing "NO PACER" for Thai singer JAYLERR. Coverage written by someone else is one of the sources AI trusts most, so we linked it from the project page and from llms.txt.
Say clearly who we are not (October 7). The App Store has plenty of apps called Talky, Hearby, Fully, and Slimly, and plenty called Intention. None of them are ours. We added a section on shared names to llms.txt.
Check which social platforms AI can read (October 7). We went through each platform's crawler rules one by one. Instagram and Threads block the crawlers of ChatGPT, Claude, Perplexity, and Gemini; public Facebook pages and YouTube don't. If you want AI to recommend you, YouTube and Facebook are more useful than Instagram. Instagram mainly helps with Google search and in-app search.
5. What we deliberately don't do
We don't block AI training crawlers. Some people suggest letting AI search you but not train on you. Our memory test came back 0 for 11, which means AI hasn't learned about us at all. Blocking it would only keep it from ever knowing us.
We don't write "the 10 best production companies in Taipei" lists that rank ourselves first. Google has already been pushing this kind of content down in 2026, and research has also found that when AI cites these lists, it mostly doesn't mention who wrote them.
We don't hide facts in places people can't see if they aren't on the visible page. The data we write for machines only restates what's already on the page.
We don't mass-produce pages "written for AI." Google's guidelines from October 1, 2026 say plainly that AI-generated content must be checked by a person before it's published [12]. On October 5 it listed "effort, originality, talent, and accuracy" as the four conditions for content quality [13].
We don't buy ads to get recommended. Ads can buy you a position. They can't buy you "being recommended."
6. What isn't done yet
Links in YouTube descriptions. The 11 official MVs we produced all name us in the description, with about 11.2 million views combined (2026-10-07), but only 1 includes a link to our website. We need the record labels' help on this one.
Pages that answer questions. We checked 65 questions against every passage of text on our website. The ones we answered worst were "how does a creative company get recommended by AI" and "what should creators learn." This piece is our answer to the first one.
A page for international production service. Our website doesn't yet have a page that directly answers foreign crews who want to shoot in Taiwan.
7. My predictions
The nearest one: when we retest on October 9, the share of unbranded questions where at least one of the three services mentions us reaches 5%. The prediction ledger puts it at 20%. We're not very hopeful, because most of the changes just went live and AI hasn't reread them yet.
The three-month target: by January 2027, the three services mention us in 20% of unbranded questions on average. This is a target, not a forecast. When it comes due we'll publish the result, and if we miss it we'll write down why.
The two-year call: "being described correctly by AI" will become a basic requirement for a brand, the way "having a website" was ten years ago. In October 2027 we'll run the same 11-question memory test again. The goal is more than half right.
8. For creative teams who want to do the same
First ask AI how it describes you. Being described wrong needs fixing before being unknown does.
Connect the people, the work, and the company. Who directed it, which company produced it, who covered it.
Ask your collaborators to put a link to your website in their YouTube descriptions. This is the most underrated step we've seen.
Measure once a month. Stop whatever shows no effect. Put more into what works.
We apply the same way of working to our own tools: INTENTION Made is where we turn the places our creative process gets stuck into tools, then release them for free. https://www.intention.ltd/made
Numbers (updated on the 9th of each month)
Date | ChatGPT | Gemini | Perplexity | Unbranded questions | Notes |
|---|---|---|---|---|---|
2026-09-25 | 4% | 6% | 2% | 0% | Starting point, 49 questions |
Next update: 2026-10-09.
Sources
CMSWire, reporting on a Gartner press release of 2024-02-19 (Gartner's site blocks automated reading, so this is a secondary source), 2024-03-20. https://www.cmswire.com/digital-experience/will-search-engine-traffic-really-drop-25-by-2026-as-gartner-predicts/
Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results, 2025-07-22. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
Ahrefs, AI Overviews Currently Appear for 83% of Branded Searches (US desktop), 2026-10-02. https://ahrefs.com/blog/ai-overviews-on-branded-searches/
TWNIC, 2025 Taiwan Internet Report (2,142 samples), 2025-12. https://report.twnic.tw/2025/assets/download/TWNIC_TaiwanInternetReport_2025_EN_summary.pdf
TechCrunch, Google's Gemini app surges to one billion users, 2026-08-11. https://techcrunch.com/2026/08/11/googles-gemini-app-surges-to-one-billion-users/
SayDigi, reporting on a survey by 大社會調查中心 (965 samples; the original report is not public, so this is a secondary source), 2026-09-20. https://www.saydigi.com/2026/09/1455725.html
Ahrefs, AI Overview brand correlation (75,000 brands), 2025-05-26. https://ahrefs.com/blog/ai-overview-brand-correlation/
Ahrefs, Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews, 2025-12-12. https://ahrefs.com/blog/ai-brand-visibility-correlations/
Meltwater, AI Search Visibility Report: August 2026 (about 7.3 million citations), 2026-09-11. https://www.meltwater.com/en/blog/ai-search-visibility-report-august-2026
OtterlyAI, YouTube AI Citation Study 2026, 2026-03. https://otterly.ai/blog/youtube-ai-citation-study-2026/
Scrunch, YouTube influencers have an AI audience, 2026. https://scrunch.com/blog/youtube-influencers-have-an-ai-audience-influencer-sponsorships/
Google Search Central, Guidance on using generative AI content, updated 2026-10-01. https://developers.google.com/search/docs/fundamentals/using-gen-ai-content
Google Search Central, Creating helpful, reliable, people-first content, updated 2026-10-05. https://developers.google.com/search/docs/fundamentals/creating-helpful-content
About the author
陳韋傑 (Aloha jie) is the founder of INTENTION (識聚有限公司, INTENTION CREATIVE CO., LTD.) and a director. INTENTION is a creative strategy and visual production company in Taipei, founded in 2022, working on album visuals, music videos, brand films and social content. Production is handled by its production arm NOUS ASIA (冉冉創意股份有限公司). INTENTION Made is where the team turns the friction in its own creative workflow into free tools.
留言