We built an AI visibility platform. The hardest problem was deciding when NOT to score

We launched ThisIsGEO.ai today, and one of the more interesting product problems we ran into had nothing to do with prompts or LLM APIs.

It was this:

What do you tell a customer when you don’t have enough evidence yet?

AI search is inherently messy. Run the same question twice and you may get different wording, different sources, or a different brand recommendation altogether. So, when we started building our scoring system, we had a choice:

Treat “we didn’t observe it” as a zero.

Or treat “we don’t know yet” as “we don’t know yet.”

We chose the second.

Here’s how we approached it:

• We separated GEO Readiness from AI Visibility. One measures whether the site gives AI systems the technical/content signals they need. The other measures what we can actually observe in AI responses.

• We distinguish between a mention, citation, and recommendation instead of lumping them together as “visibility.”

• Observations are dated, because an AI answer today isn’t necessarily an AI answer next month.

• When we don’t have enough evidence to make a defensible claim, we’d rather show pending / insufficient evidence than manufacture a precise-looking score.

That last one was surprisingly hard. Dashboards want numbers. Customers understandably want numbers - and a big green 87/100 feels much more satisfying than, “We need more observations before we can tell you.”

I’m starting to think false precision is going to become a major problem in AI analytics - so I’m curious how other builders are thinking about this:

When you’re measuring something nondeterministic, is an incomplete answer actually more trustworthy than a complete-looking score?

And where do you draw the line between making analytics simple and making them artificially certain?