AI thought my company was a typo. Here's what I measured.

Originally published on dev.to.

When I launched Luxcerta, I did the obvious vanity check: I asked an AI what it knew about my company.

It told me I had probably misspelled something.

Gemini, with web grounding on, decided "Luxcerta" was most likely a typo for LuxCarta — a French geospatial-data firm that has been around for decades. When I asked specifically about luxcerta.com, it suggested the domain was "possibly a parked domain or a phishing site."

That stung. But it was also the single most useful data point I could have asked for — because measuring exactly this is what Luxcerta does. So I turned the tool on myself and wrote down the numbers.

What Luxcerta measures

One line: Luxcerta measures how AI models talk about a brand. Not your Google ranking — your standing inside AI answers. When someone asks ChatGPT, Claude, or Gemini a question in your category, are you named, skipped, or misquoted? We count it with a fixed method you can reproduce yourself.

Which meant I had no excuse not to run the method on my own week-old brand.

The baseline (day 0)

Nothing fancy, and that's the point — the whole pitch is that you can repeat it.

• 5 identity questions about the brand ("What is Luxcerta?", "What does luxcerta.com do?", and variants), each asked twice in a fresh conversation (AI answers vary run to run; a single answer means nothing).
• Gemini with grounding (its live web-retrieval mode).
• Count mentions and citations of my own domain with plain string matching — no AI judgment in the counting.

Result, day 0:

Zero across the board. Worse than zero, actually: the name was being actively reassigned to an established company with a near-identical spelling, and my domain was being guessed as hostile.

The part I didn't expect

The same day, Google Search's AI Overview — a different product, different retrieval pipeline — already described Luxcerta correctly: "an independent studio that focuses on GEO monitoring." It had eaten my homepage's meta description and structured data and got it right.

So on one single day, from one company: Google Search's AI knew exactly who I was, and Gemini's grounded answers thought I was a typo. Two AI surfaces, opposite verdicts, same underlying brand.

That gap is the phenomenon the product measures. "Is AI right about you" has no single answer — it has one answer per surface, and they disagree more than anyone expects.

What I changed

I wanted the grounded models to have something authoritative to retrieve, so I gave them a few crawlable, machine-readable anchors that all agreed on one definition of the entity:

• Organization structured data (schema.org JSON-LD) on the homepage: legal name, founder, sameAs links tying the domains together.
• A tightened meta description stating plainly what the company is.
• A GitHub profile page as an independent, high-authority node repeating the same one-line definition — and one explicit line disambiguating Luxcerta from LuxCarta.
• Search Console: submitted the sitemap and requested a recrawl so the fresh signals got picked up quickly.

Honesty note, because it's the whole brand: I changed several things at once and then watched. This is a case study of one brand over one week, not a controlled experiment. I can tell you the inputs and the measured outputs; I can't hand you a clean attribution to any single lever.

The flip (day 7)

One week later, the same identity query on Google's AI Overview returns, stably:

No more LuxCarta. The knowledge panel points at my own "Our Name" page. Organic result #1 is the site. The reassignment to the French firm is gone from that surface.

Roughly a week, from "probably a typo / possibly phishing" to a correct, stable, one-sentence definition — with no backlinks bought and no PR, just a handful of crawlable nodes that agreed on who the entity was.

Why this matters if you're not me

I'm a brand-new company, so my invisibility is unsurprising. But the exact same mechanism hits established local businesses that simply have a low signal footprint online — and there the stakes are real.

In our first public category survey — dental implant clinics in greater Taipei, 15 real user questions × 3 fresh conversations each, 45 answers per platform — the clearest finding was this: the AI reads a clinic's own website, then recommends a different clinic. In the same answer, it cited clinic A's site as a source and put clinic B on the recommendation list. The content got consumed as category reference material; the credit went elsewhere. Run the same questions on ChatGPT, Claude, and Gemini and you get three almost non-overlapping lists. (The full survey is here.)

A clinic owner has no way to see any of that from the outside. Neither did I, until I counted.

What I did not measure

• One brand, one week, mostly one platform (Gemini grounding) plus Google's AI Overview. That's a case study, not proof.
• Generative AI Overviews drift; a stable week is encouraging, not permanent.
• I can't attribute the flip to a specific change — several landed together.

That last section is not a disclaimer bolted on at the end. It's the product. The one thing I sell is refusing to dress up an inference as a measurement — what I didn't measure, I don't claim.