Every AI visibility number starts with reading the answer. Caliper does that reading inside Omni Impact, with one specialist model for each job: naming the brands in an answer, sorting the sites it cites, and crediting each cited page with the brands it backs. All three are live as of November 5, 2026.
Why we built our own.
Most of the work behind an AI visibility score happens after the answer comes back. For every answer, we need to know which brands it names, whether the client is one of them, which brand it leads with, what kind of site each citation points to, and which brands each cited page backs.
A general AI model can read an answer well. Running one on every answer is another matter. Each call costs money. The same answer can come back read differently from one run to the next. And the general model we used missed details clients care about most, like a single school inside a university. Fixed site lists have a harder limit: they can’t sort a site they have never seen, and AI answers cite thousands of them.
So we built Caliper: specialist models, one job each, that run on our own servers and read the same answer the same way every time.
Meet Caliper.
- Caliper Brand Live
- Reads an AI answer and lists every brand it names. It marks which belong to the client, including sub-brands such as a school, a department or a product line, which are recommended, the tone toward each, and which brand the answer leads with. It powers competitor lists and recommendation and top-pick rates.
- Caliper Source Live
- Sorts each site an AI answer cites by the kind of source it is: the client's own site, a competitor or other business, a directory or ranking list, a publisher, a social or forum site, or a reference source. It powers source breakdowns and suggestions for where to get listed.
- Caliper Page Live
- Decides which brands each cited page actually backs within an answer. A page gets credit only for the brands its cited text supports, so you can see the exact pages that speak for you and the ones that speak for competitors.
The proof.
Every number here comes from answers the models never saw while they were being built, and each test was scored once. For a university client, an AI model from a different lab labeled the answers blind, without seeing any of our systems’ output. Results in other industries are graded against our own answer keys.
Naming brands
The biggest gain is on the client itself. A single school or institute inside a university is easy for a general model to miss. Caliper Brand catches far more of them, and it stays ahead across industries on the client’s own brand and on which brands an answer recommends. On whether the client was named at all, it missed a fifth as often as the general model.
Sorting cited sites
Caliper Source knows what kind of site a citation points to even when the site appears on no list, and it recognizes competitors’ own sites.
Crediting cited pages
An AI answer often names several brands and cites several pages, and each page usually speaks for only some of them. Caliper Page reads the text each page is cited for and credits the page only with the brands that text supports.
Same reading, every time.
Caliper gives the same answer the same reading every time. When a client’s numbers move on the monitoring dashboard, it’s because the answers changed, not because of how they were read.
Because the brand step no longer calls an outside AI model, it carries no per-call model fee, and answer text stays on our servers for that step.
Where we’re still pushing.
Caliper doesn’t win every row yet. Here is where it is level with the general model or behind it.
- Tone is level with the general model. Both get it right 84% of the time for the university client. Across industries Caliper Brand is one point behind, 83% against 84%. A larger model for tone is in the works.
- It misses more small local names. Outside higher education it finds 86% of brand names against the general model’s 95%. More of the names it does report are right, 92% against 86%.
- Caliper Page covers three answer engines so far. The rest follow as each gets its own test set.
What’s next.
- Caliper Page on more answer engines, as each gets its own test set.
- More Caliper models for other jobs in our pipeline. Each one ships only if it beats what it replaces on answers it has never seen.
Caliper runs inside the Omni Impact platform today.
How we measured.
Percentages are the share of items read correctly. Scores out of 100 are F1 × 100, a standard score that rewards finding what is there and penalizes reporting what is not.
For the university client, an AI model from a different lab labeled each answer blind, as an outside labeler. Our own answer keys come from blind AI labeling against a written rulebook, with second reviews on disputed items and human spot checks.
