Find out how invisible you are to AI.
Everyone else asks ChatGPT once and writes down yes or no. We ask three times, report a real confidence interval, and keep every answer so you can check us.
No instant demo. Real measurements take time and we would rather be right.
We ran it on a product we own.
99 judged answers about architecture diagram tools. Ours is the sad little orange one.
- ✓Real figures from our own instrument, not a mock-up.
- ✓Counted over decided replicates only, for every tool equally.
- ✓Every one of those 62 draw.io mentions is an answer you can open and read.
Yes, we published our own bad number. It is real, and a truncated answer is thrown out for every tool equally — including ours.
Three asks, not one.
Models are not deterministic. Ask once and you have written down a coin flip.
- ✓Three replicates per prompt, per model, every week.
- ✓A 95% confidence interval, so you can tell a real change from noise.
- ✓A cut-off answer is marked insufficient and left out — not counted as a zero.
Two replicates is a wide band, and the screen says so. Most tools would have shown you “67%” and moved on.
Things it does not do.
Lumizam measures. It does not write your blog or fix your pages. Better you read this now than find out in week two.
- ✕No instant scan A real measurement is three replicates across three models, and it costs us money to run. You will wait a bit.
- ✕No recommendations You get what is true today, not a ranked to-do list. That comes once we can prove a fix actually moved something.
- ✕No content generation Nothing here writes articles or opens pull requests against your site.
- ✕API, not the app What a model returns through its API is close to, but not the same as, what a person sees in the chat product. We would rather say so.
See who is getting named instead.
A mention rate on its own tells you nothing. The same question, asked of your rivals, tells you whether the category is winnable or already taken.
- ✓Every competitor counted on the same answers, in the same week.
- ✓Which sources the models cite when they recommend someone else.
- ✓A control group: if everyone drops at once, that is the model changing, not you.
Illustrative shape, not a measurement. Your own version is built from your answers once we have run them.
Measure first. Act second.
The instrument has to be trustworthy before anything built on top of it is worth having. This is the order we are building in, and roughly where we are.
- ✓Measurement Weekly rates across three models, share of voice, crawler logs, Search Console and a health audit. Running now.
- →Diagnosis Pages losing clicks, titles worth rewriting, two pages competing for one query, and a ranked list of what to fix first.
- ·Fixes you approve Small, surgical changes delivered as a pull request you read and merge. Nothing is published without you.
- ·Did it work? The point of the whole thing: hold back a control group, re-measure after a change, and report the effect — including when the answer is “no difference”.
What you can hold us to.
See your own numbers.
Tell us the domain you want measured. We run it when beta opens and send you the result — good or bad.