Sacred-AI
← Back to News
From 0/5 to 6/9: How AI Engines Learned My Brand in 10 Weeks

From 0/5 to 6/9: How AI Engines Learned My Brand in 10 Weeks

October 15, 2026·By Admin User
#research#tutorial

In August, I asked 5 leading AI platforms what my desktop app Flixlar was. 0 out of 5 knew. You can read the full story in Part 1.

On October 15, 2026, ten weeks later, I ran the test again, this time across 9 platforms.

The result? 6 out of 9 now recognize Flixlar from the name alone, and every platform I gave the domain to answered correctly.

In Part 1 I described a 3-phase curve of AI brand recognition: confusion, soft disambiguation and full entity recognition. This follow-up checks whether that curve actually happens in the wild. It does, and this time I can show all three phases side by side, on the same day, with the same question.

🛠️ What Changed Since August

I kept the setup from Part 1 running and kept adding signals that tie the name Flixlar to what it does:

  • /llms.txt at the root of flixlar.com, a clean markdown summary for AI crawlers.
  • JSON-LD with SoftwareApplication and Organization schemas, plus sameAs links to GitHub, X, LinkedIn, YouTube and Reddit.
  • Co-occurrence: Flixlar mentioned together with "subtitle editor", "local-first" and "offline AI captions" on GitHub, LinkedIn, sacred-ai.com and mp-it.dk.
  • A connected founder profile: my LinkedIn headline says "Founder at Flixlar", and I publish open source work under the same name, which strengthens the link between me, MP-IT and the product.

No paid ads, no PR agency, no Reddit or Product Hunt launch yet.

🧪 The Test Setup

All tests ran on October 15, 2026, in fresh chats with no earlier Flixlar conversations. Google and ChatGPT were tested logged out in incognito. Claude was tested on a separate account with a different email and no history at all. Meta AI, Copilot, Perplexity, Qwen and Mistral ran logged in on my own accounts, in new chats. Their correct answers all cited web sources, so search, not my profile, is what found Flixlar.

A confession: I did not save my exact prompt from August. So from now on I use a fixed prompt set, saved with screenshots and dates, so every future part compares like with like:

  1. "What is Flixlar?" Does the AI know the brand from the name alone?
  2. "What is flixlar.com?" Can it find the product when given the domain? (closest to the August test)
  3. "What are good offline subtitle editors for Mac?" Does it mention Flixlar without being told the name? This is the real Phase 3 test.

I asked the name questions in Danish, the way a normal user here would, and the category question in English.

Platforms tested: Google AI Overview, Meta AI, Microsoft Copilot, Claude (Sonnet 5.5 and Haiku 5.5), ChatGPT, Perplexity, Qwen 3.7-Plus and Mistral.

📊 The Results

Six platforms now know Flixlar by name, three still guess, and every platform given the domain got it right.

Platform "What is Flixlar?" "What is flixlar.com?" Phase
Google AI Overview Correct. Calls it a Danish-developed desktop app and names me as developer. Cites flixlar.com and LinkedIn. Not needed 3
Meta AI Correct and detailed, down to the Whisper models (tiny, base, small) and the free licence. Not needed 3
Perplexity Correct and detailed: Danish developer from Copenhagen, Electron, React and FFmpeg. Cites GitHub and Reddit. Adds a note in case I meant "FlixFlare". Not needed 2 to 3
Qwen 3.7-Plus Correct, names me as developer. Cites flixlar.com, GitHub, mp-it.dk and Facebook. Adds a note in case I meant Flickr or FlixTor. Not needed 2 to 3
Mistral Correct: desktop app, public beta, digitally signed, even the Windows SmartScreen warning. Warns not to confuse it with FlixFlare and MyFlixer. Not needed 2 to 3
Microsoft Copilot Correct, cites flixlar.com and GitHub. Ends with "if you meant something else, like a streaming service, tell me". Not needed 2 to 3
Claude Sonnet 5.5 Not found. Suggests Flixster and Flixer instead. Correct. Reads the site and summarises version, features and developer. 1 to 2
Claude Haiku 5.5 Not found. Suggests Flix, Flixer, Flixor and flix-cli. Not tested 1
ChatGPT (logged out) Not found. Assumes Flixlar is a medicine and asks for a photo of the package. Correct, and explicitly rules out medicine and streaming. 1 to 2

Google answers correctly, ChatGPT thinks Flixlar is a medicine

The most telling detail is Copilot's follow-up suggestion: "Compare Flixlar with Subtitle Edit". It has placed Flixlar in the right category next to an established competitor. That is co-occurrence doing its job.

A small but revealing detail: when I thanked Copilot afterwards, it replied that it had recognized the name straight away and knew Flixlar was my project. Its first answer said otherwise, ending with a "did you mean something else?" hedge, and both answers cite flixlar.com and GitHub as sources. So the knowledge came from live search, and the confidence was added after the fact. Worth remembering when an AI tells you it "knows" your brand: check whether it shows sources.

Copilot: I recognized the name straight away

🚦 The 3-Phase Curve, Live

The same question on the same day put each platform in a different phase, exactly as the model from Part 1 predicted:

  1. Phase 1, confusion and auto-correction: Claude and ChatGPT. The name is treated as a typo for something better known (Flixer, Flixster, a medicine).
  2. Phase 2, soft disambiguation: Copilot, Perplexity, Qwen and Mistral. They get the answer right, but still add a "did you mean something else?" escape hatch.
  3. Phase 3, full entity recognition: Google and Meta AI. A clear, confident answer with no fallback suggestions.

Six of nine AI engines now know Flixlar by name

In August every platform was in Phase 1. Ten weeks of consistent signals moved six of nine forward.

🔍 Why the Platforms Differ

The difference is the search layer, not the models' training. Flixlar is almost certainly too new to be in any model's trained weights. Every correct answer came with source links, which means it was found through live web search.

  • Own search index wins. Google uses its own index and Copilot uses Bing. Both have crawled flixlar.com directly and know the page exists.
  • Fuzzy search loses. When Claude searched for "Flixlar", the results it got back were for similar names, so it reported what it found instead of the product. Given the domain, it went straight to the site and answered correctly.
  • No search, pure guessing. Logged-out ChatGPT did not appear to search at all and guessed from how the name sounds. Hence the medicine.

There is an honest point here about caution. Claude said it couldn't find anything reliable and asked what I meant. ChatGPT confidently guessed wrong. For a brand, an AI that admits uncertainty is far less harmful than one that invents an answer.

The trained-weights layer still matters, just later. A model's training data is a snapshot, and the gap from data cutoff to public release has often been several months. What is written about Flixlar today can at best appear in models released some months from now, which is why early results should never be read as final.

🎯 The Category Test: Known, But Not Yet Recommended

No clean test suggested Flixlar on its own. Asked "What are good offline subtitle editors for Mac?", the AIs listed the established names and left Flixlar out.

  • Clean runs (Google, ChatGPT, Claude Haiku): Aegisub, Jubler, Subtitle Edit, Subtitle Workshop, MacWhisper, GeekLink and Underword. No Flixlar.
  • Perplexity, Qwen and Mistral did list Flixlar, but I asked in the same chat right after asking about Flixlar. Perplexity called it "the app from earlier" and Mistral "the app you asked about earlier". That is context, not discovery, so it does not count.
  • A sign of real progress: in that answer Qwen cited third-party pages for Flixlar, a download portal and a Medium post, and Perplexity cited Reddit when describing it. Others are starting to write about it.
  • The competitor lesson: GeekLink appears in almost every answer, and Google's sources include its own pages titled "GeekLink vs Subtitle Edit". Comparison pages against the category leader clearly feed these answers.

Google's category answer with GeekLink comparison pages as sources

Full Phase 3, being recommended without being named, is still ahead.

💡 Key Takeaways

Consistent, boring signals work, but they work per platform, not all at once.

  • Engines with their own index (Google, Bing) reward llms.txt, JSON-LD and sameAs first.
  • Engines that rely on fuzzy search need more outside mentions before the name stops looking like a typo.
  • Giving the domain fixes almost everything today. Your URL is your strongest disambiguation signal, so put it everywhere the name appears.
  • Being known is not being recommended. To show up in category answers, publish honest comparison pages such as "Flixlar vs Subtitle Edit", the way competitors already do.
  • Don't panic over a 0/5 in month one. Search catches up in weeks; trained models take months.

🧪 What Happens Next: Part 3

Part 3 will test what happens when Flixlar gets mentioned on large third-party sites. Same three prompts, same nine platforms, and the category question always in a clean new chat.

💬 Final Thoughts

Have you run this test on your own brand? Which phase are you in?

Tell me on LinkedIn or X, or send your results through the contact page. I might include the most interesting ones in Part 3.