AI visibility diagnostics · pre-launch

Three things can keep your brand out of AI answers. We tell you which one it is.

Then we check whether your fix worked.

PLATE 00 · Illustrative — example values, not customer data.

The three failures

A brand missing from AI answers has one of three problems. Each needs a different fix.

It can't find youwe call this discoverability: The AI never turns up your brand when someone asks about your category.
It finds you, but picks someone elsewe call this compellingness: The AI reads your pages. It just recommends a competitor instead.
It depends who's askingwe call this positioning: You win with one kind of buyer and lose with another, in the same category.

Why name the problem first? Take the best-known number in this field: rewriting your content can lift your visibility in an AI's answer by up to 40%. True — but that was measured on five pages already sitting in front of the model (Aggarwal et al., KDD 2024). It says nothing about getting in front of the model in the first place. So if the AI never finds you, no rewrite will help. Naming the problem first tells you where the money should go.

How the diagnosis works

Every diagnosis is two measurement runs, on two separate days. We name a problem only when both runs agree, and only when the result sits outside the noise band — the range where a reading could just be the measurement wobbling. Anything less and you get an honest “inconclusive”: the full evidence, and a second measurement we pay for. We never round a shaky result up to a verdict. That rule isn't caution for its own sake. Researchers asked AI search engines the same question twice, five minutes apart. The answer's overall decision flipped on 9–27% of questions — even for the engines where the randomness setting could be turned off (Kirsten et al., ACL 2026). Over a full day it flipped slightly more often. Measure once and you're reading noise.

Fig. 1

A verdict takes two agreeing runs, outside the noise

Two runs, on separate days · rate from 0 to 1, with its band

What you actually see

We show you the pages the AI actually read, and where your brand stands on each one — what we observed, not what we assume it noticed. Every rate comes with the band around it, so you can see how sure it is. There is no composite score anywhere in the product, ever. We also measure each AI system on its own, because they really do differ. One peer-reviewed audit compared two of them and found they shared only 26% of the sites they cited (Li & Sinnamon, 2024). That's why we never pool them into one reading.

Fig. 2

The pages the AI actually read

Pages from one run · each mark shows what we saw on that page

PAGES THE AI ACTUALLY READ

the AI recommended from here
you appear on this page
you don't appear here
you appear on this page

~~guessing what it noticed~~ — not used. We only report what we saw.

Fig. 3

Every rate keeps its band

Each rate from 0 to 1, with its band · three rates, never combined

How this is different from a prompt tracker

Prompt trackers	Ternith
They hand you a share-of-voice score and a dashboard to watch over time.	We name which of the three problems is causing your gap, and show the evidence — so your fix budget goes at the real cause.
They measure once and treat it as precise.	We measure twice, on two separate days, and only call it when both runs agree.
They give you a number that moves up and down.	We show the pages the AI actually read and where you stand on each — with rates, and the band around each one.
They leave you watching a dashboard.	We tell you the cause, then measure again to see if your fix worked.
They round an ambiguous result up to a verdict.	When our two runs disagree, we say “inconclusive”, show you everything, and pay for another measurement.
They give you one score that hides which of the three problems you actually have.	We separate the three, so you fix the right one.

What we set out to do

Ternith is pre-launch. Here's what we set out to do: stop the guessing about why a brand is missing from AI answers. Instead of a score to watch, name the actual cause — one of three problems — with evidence you can check yourself. Then your fix budget goes at the real thing, and we measure whether it worked. We're building the instrument now, and opening a waitlist while we do.

Fig. 4

The loop closes: diagnose, fix, measure again

The same two runs, before and after a fix · rate with its band

DIAGNOSIS

Failure named:

found, but not chosen

recommended · 0.24 ± 0.06

point your fix budget at this cause

RE-MEASURE · SAME TWO-RUN METHOD

before

after

did it move beyond its bands?

only if still inconclusive — and that next measurement is funded

Honest limitations

We measure the AI companies' developer interface, with live search switched on. That's a close stand-in for the app your customers use — not the identical thing. We'd rather say so.
Ternith is pre-launch and hasn't been tested with customers yet. Joining the waitlist is the only thing you can do here today.
“Audit-grade” describes the evidence trail behind each diagnosis: two separate runs, agreement outside the noise band, the pages the AI read, and every rate with its band. It is not a compliance certification — no standard or accreditation is implied.

Sources — The research cited on this page. Peer-reviewed venues only.

Questions

What is Ternith?

Ternith tells you why AI assistants like ChatGPT, Claude and Gemini leave your brand out of their answers. There are three possible reasons: they never find you, they find you but recommend someone else, or you win with one kind of buyer and lose with another. We work out which one is yours, show you the evidence, and then measure whether your fix worked. You get a diagnosis, not a score.

How is this different from a prompt tracker or AI-visibility monitor?

Those tools give you a share-of-voice score and a dashboard to watch. We name the specific problem behind your gap, show you the pages the AI actually read, and give you no composite score at all.

What are the three problems?

Discoverability: the AI never turns up your brand when someone asks about your category. Compellingness: it reads your pages but recommends a competitor instead. Positioning: you win with one kind of buyer and lose with another.

Does Ternith give a visibility score?

No. There is no composite score anywhere in the product. You get rates, each with the band around it. And when the evidence doesn't support naming a problem, you get an honest “inconclusive”.

How do you decide a diagnosis is real?

Every diagnosis is two measurement runs on two separate days. We name a problem only if both runs agree and the result sits outside the noise band. Otherwise you get the full evidence, an honest “inconclusive”, and a second measurement we pay for.

What do you actually show me?

The pages the AI read when it answered about your category, and where your brand stands on each one. That's what we observed — not a guess about what the AI noticed. Every rate comes with the band around it.

Do you measure exactly what I'd see in ChatGPT, Claude or Gemini?

No. We measure the AI companies' developer interface with live search switched on. It's a close stand-in for the consumer app, not the identical thing. We'd rather say so than pretend otherwise.

Can I use Ternith now, and is “audit-grade” a certification?

Ternith is pre-launch, so joining the waitlist is the only thing you can do today. “Audit-grade” describes the evidence trail behind each diagnosis — it is not a compliance certification, and no standard or accreditation is implied.

Join the waitlist

We'll email you once, when Ternith opens. No score to watch in the meantime — that's the point.