A Score You Can Defend Beats a Score That Looks Precise
A defensible AI visibility score is one where every number resolves to a reason: a check that passed or failed, an answer engine that did or did not name you, a judgment made the same way every time. Most tools cannot give you that. They hand you one grade, a 62 or a C+, and when a client asks "why," or your founder asks "why did it drop," you have nothing to point to.
That gap is not cosmetic. If you run an agency, the score is something you defend in a slide. If you are a PM, it is something you explain when it moves. If you are a growth lead, you have probably run an AI visibility audit twice and gotten two different numbers, with no change to your site in between, and quietly stopped trusting the tool. A score you cannot defend is a score you cannot act on.
Radar was rebuilt to fix exactly this. The point of the rebuild was not more features. It was stability, explainability, and closer alignment to what answer engines actually do, so the number you get is one you can stand behind. This is the methodology behind the decision-stage frame: to win the AI recommendation, you first need a score you can trust enough to act on.
What Makes an AI Visibility Score Trustworthy?
An AI visibility score is trustworthy when it changes because your site changed, not because the grader ran again. That single property is what most tools quietly fail.
The old way of scoring AI visibility had a strong measurement platform but a noisy and hard-to-defend scoring layer. Deterministic signals with definite right answers were sometimes judged by a language model. Subjective judgments ran once, at an unknown temperature, so the next run wobbled. Weighting was equal by default, with no per-context calibration. And there was no reliable audit trail tying a score back to the tools, models, and prompts that produced it.
The result is a grade that looks precise and behaves like a guess. You cannot tell whether a five-point drop means a real regression on your site or just a model that answered differently this morning. The rebuild routes signals by type, anchors the composite in live engine measurement, stabilizes the subjective layer, and pins every score to a method trail.
Picture the moment this fails. You send a client their monthly report and the score dropped four points. They ask what happened. You rerun the audit to investigate, and it comes back up three points, with nothing changed on their site. Now you are not explaining a result, you are explaining your tool, and the conversation about their AI strategy is over before it started. A score that moves on its own is not a metric. It is a liability you manage in every client call, and it is the exact problem the three-tier rebuild was designed to remove.
| Dimension | Before the rebuild | After the rebuild |
|---|---|---|
| Overall model | One blended grade where noisy signals could dominate | Three tiers separated under a method trail |
| Deterministic checks | Some right-answer checks judged by a model | Routed into parser-based scoring paths |
| Live measurement | Measured, but not cleanly anchored to engine outcomes | Anchors the composite in live engine results |
| Subjective judgment | Ran once at unknown temperature, causing rerun wobble | Repeated runs, fixed settings, confidence bands |
| Weighting | Equal by default, no per-context calibration | Calibration infrastructure built; weights not yet calibrated against outcomes |
| Provenance | No reliable audit trail for a score | Resolves to tool, model, prompt, weight version, run |
| Trend integrity | Recalibration could silently rewrite past scores | Scores stored with a method label, never recomputed |
The Three Layers of a Defensible Score
The core idea is simple: different kinds of signal deserve different kinds of scoring. A pass-or-fail check should be scored by code. A live outcome should be measured. A subjective call should be judged carefully and reported with its uncertainty. Blending all three into one number is what makes a score impossible to defend.
How Radar builds one score
Four steps, each scored the way that kind of signal should be
Deterministic checks
Parsers, right or wrong
Live measurement
What the engines say
Stabilized judgment
Subjective, with bands
Method trail
Every score is traceable
Each layer answers a different question. Tier 1 asks "is this technically correct," and a parser can answer it. Tier 2 asks "what do the answer engines actually do," and only a live query can answer it. Tier 3 asks "how good is this in a way that has no single right answer," and that requires judgment, handled so it does not introduce noise. The method trail then makes the whole thing legible after the fact.
As of October 2026, this 30-second Scout film covers the access side of Radar: the six free technical checks, the full paid 13-tool audit and the evidence behind a finding. Those free and paid access levels are separate from the scoring layers above.
Read the transcript
Hmm. What's behind a Radar score? Six free technical checks examine whether AI can reach and read your site. The full paid audit runs thirteen tools, including checks of what AI engines say. Open a finding. See the evidence behind it. Then decide what to fix first.
Deterministic Checks: Where There Is a Right Answer
A deterministic check has a definite right answer, so Radar scores it with a parser, never with model opinion. Can GPTBot reach this URL. Does robots.txt allow the AI crawlers. Is the JSON-LD valid. These are not matters of taste.
This sounds obvious, but letting a model grade a pass-or-fail check is the single most common way an AI visibility score becomes undefendable. The model is confident, fast, and occasionally wrong, and there is no way to audit why it called a valid schema invalid. By contrast, a parser either finds the tag or it does not, and you can reproduce the result on demand.
Some tools still generate a plain-language summary of these checks with a model, so the finding reads like a sentence instead of a flag. That is fine. The important line is that the narrative never moves the number. The score for a deterministic check comes from code, and the writeup is a separate, downstream convenience.
Live Measurement: Where the Answer Engines Decide
The second layer measures what ChatGPT, Claude, Gemini, and Perplexity actually say, because the only honest way to know what a generative engine says about you is to ask it.
This is the layer that anchors the composite. AI answers change, so a paid Radar audit issues fresh queries on every run so the measured signals (citations, share of voice, source influence, factual flags) reflect the current state of the model, not a stale capture.
Live measurement is also where the difference between monitoring and readiness shows up. Monitoring tells you that you appeared. Measurement at the decision stage, which is on our roadmap, would ask whether the answer recommended you when a buyer applied real criteria. Detection is stronger on engines that ground their answers in retrieved sources, and we are still hardening it on the engines that lean on model prose. That is a known edge we name rather than paper over.
Stabilized Judgment: Only Where Subjectivity Is Unavoidable
Some signals genuinely have no single right answer, like how cleanly a page answers the question a buyer would actually ask an AI. For those, Radar uses model judgment, but it stabilizes that judgment so it does not behave like a coin flip.
“If a score moves when the grader reruns instead of when the site changes, the score was never measuring the site.”
Radar scoring principle
Stabilization means the subjective check runs repeatedly at temperature zero, cross-checks agreement between models, and reports a confidence band rather than a single brittle number. When two independent models agree, confidence is high. When they diverge, the band widens and the result says so, instead of pretending to a precision it does not have.
This is the opposite of the old failure mode, where a single judgment call ran once at an unknown temperature and the whole grade rode on that one roll. By isolating the subjective layer and forcing it to show its uncertainty, Radar keeps model noise from leaking into a number you are about to defend in a meeting.
For an operator, the confidence band is the part you actually use. A high-confidence subjective score is something you can put weight on and act against this week. A wide band is a signal to wait for a rerun or to treat that dimension as directional rather than precise. Either way you know which is which, instead of treating a shaky judgment and a solid one as the same number. That distinction is the difference between a report you forward to a client without a second thought and one you quietly double-check first.
Provenance: Every Score Resolves to a Method Trail
Provenance is the quiet feature that makes everything else trustworthy: every score resolves to a record of how it was produced.
Each score ties back to the tool that produced it, the tier it belongs to, the model and prompt used, the weight-set version, the run ID, and the timestamp. That record is what lets you answer "why" with a specific reason instead of a shrug.
It also protects your history. Each saved score is stored with its method label and is never recomputed, so improving the model later cannot silently rewrite last month's numbers. That label has not yet changed when the rubric changed, so it cannot tell you which rubric produced a given score. The dated scoring changelog documents some rubric changes, with a note beside the scores they affected, but it is not yet complete. A trend you cannot trust across methodology changes is not a trend, so read any trend with that limit in mind.
Corrected 25 September 2026: an earlier version said a method version pinned to each score keeps trends honest across methodology changes. Each score does carry a method label and is never recomputed, but the label has not changed with rubric changes so far, and the scoring changelog documents some of those changes but is not yet complete.
This is what changes the client conversation. When a score moves, you do not guess. You open the trail, see that a specific check started failing or a specific engine stopped naming the brand, and you walk in with the cause and the fix already in hand. The number stops being something you defend and becomes something you use. That is the entire point of separating the layers and recording the method: not to look rigorous, but to give the person holding the score an answer they can stand behind.
A Score That Changes When Your Site Changes, Not When the Grader Reruns
The payoff of all this separation is a single, plain promise: the number moves when your site moves, and holds steady when it does not. That is what makes it defensible to a client or a founder.
What you can actually do with the score
- One opaque number, no reason behind it
- Wobbles on rerun with no site change
- Cannot explain a drop to a client
- Recalibration quietly rewrites your history
- Every number resolves to a check or a measurement
- Built to control rerun variance, not hide it
- A drop points to a specific, fixable cause
- Scores never recomputed; changelog covers some rubric changes
Now the honest part, because honesty is the whole point of a defensible score. Radar is not finished, and pretending otherwise would undercut the argument.
Two things are deliberately incomplete. First, the stabilized judgment layer is built and ships behind a flag that is off by default, and it does not carry weight in the visible composite score. We hold it out until calibration has enough labeled outcome coverage to earn that weight against real results. Shipping a larger number we cannot stand behind would be the exact failure this rebuild was meant to fix. Second, calibration infrastructure is in place, but it sharpens as coverage grows. It is not honest to say every segment is fully calibrated today, so we do not say it.
What is built, and what is still earning its weight
Three-tier separation and the provenance trail are in production
Stabilized judgment runs but does not yet move the visible composite
Calibration improves as labeled outcome coverage expands
That candor is not a weakness in the pitch. It is the pitch. Any tool can publish a blog post claiming its score is the most accurate. Almost none will tell you which part of their own score they do not yet trust enough to weight. The reason to believe Radar is defensible is that it tells you where the edges are, names the caveats, and pins the method so you can check the work. A score that hides its uncertainty is asking for faith. A score that shows it is asking to be verified. For a skeptical operator, the second one is worth more.
If you are choosing between tools, this is the lens that matters more than feature counts. The comparison of AI visibility tools in 2026 is really a comparison of which scores you can defend when someone pushes back.
“A score that hides its uncertainty asks for faith. A score that shows it asks to be verified.”
Radar scoring principle
How Radar Scores AI Visibility: Questions Operators Ask
How Radar Scores AI Visibility: Questions Operators Ask
Common questions about this topic, answered.
How does Radar score AI visibility?
Radar separates a score into three layers instead of blending them into one grade. Deterministic checks (crawl access, robots.txt, schema) are scored by parsers because they have a right answer. Live measurement queries ChatGPT, Claude, Gemini, and Perplexity to see what they actually say about you. Stabilized judgment handles the subjective signals with repeated runs and confidence bands. Every score then resolves to a method trail, so you can see how the number was produced.
What makes an AI visibility score reproducible?
A score is reproducible when it changes because your site changed, not because the grader ran again. Radar gets there by scoring deterministic checks with parsers (no model opinion), storing every score with its method label and never recomputing it, and stabilizing subjective signals with repeated runs at temperature zero plus a confidence band. The architecture is built to control rerun variance rather than pretend it does not exist.
Why do AI visibility scores change when you rerun them?
Most scores wobble on rerun because a language model graded something subjective once, at an unknown temperature, and the next run produced a slightly different answer. When that single noisy judgment feeds straight into the composite, the whole grade moves even though nothing on your site changed. Radar isolates the subjective layer, runs it repeatedly, and reports a confidence band so a real shift is distinguishable from model noise.
What is the difference between a deterministic check and a judged signal?
A deterministic check has a definite right answer that a parser can verify: can GPTBot reach this URL, does the robots.txt allow it, is the schema valid. A judged signal is subjective and has no single correct value, such as how clearly a page answers a buying question. Radar scores the first kind with code and the second kind with stabilized model judgment, and it never lets a model vote on a check that has a right answer.
Does Radar use AI to grade its own checks?
Not on the score path for deterministic checks. Checks with a right answer are scored by parsers, not by a model. Some tools still generate a written summary with a model so the result reads in plain language, but that narrative never moves the number. The separation matters because a model grading a pass or fail check is the most common way an AI visibility score becomes impossible to defend.
How is Radar different from a monitoring tool?
Monitoring tools track whether your brand appears in AI answers over time. As of September 2026, Radar audits technical readiness and samples AI answers about your brand, with the evidence and a suggested fix behind each finding and a score you can inspect. Decision-stage measurement, whether the engines recommend you, is the direction Radar is built toward, not part of the live score. The two are complementary layers, not competitors. Many teams use an audit to check the foundation and a monitoring tool to watch the trend over time.
Is Radar fully calibrated yet?
No, and that is on purpose. The calibration infrastructure is in place, but it sharpens as labeled outcome coverage grows, so it is not honest to claim every segment is fully calibrated today. The stabilized judgment layer runs now but does not yet carry weight in the visible composite score, because we hold it out until calibration can earn that weight against real outcomes. We would rather ship a smaller defensible score than a larger one we cannot stand behind.
Can I see how a Radar score was calculated?
Yes. Every score resolves to a provenance record: the tool, the tier, the model, the prompt, the weight-set version, the run ID, and the timestamp. Saved scores are never recomputed, so recalibrating the model later cannot silently rewrite last month’s numbers, and the scoring changelog dates some rubric changes, though it is not yet complete. That trail is what turns a number into an answer you can give a client or a founder.
The Bottom Line
A defensible AI visibility score is the one you can act on. Most tools optimize for a number that looks precise; Radar optimizes for a number you can explain. The separation into deterministic checks, live measurement, and stabilized judgment, with a method trail behind every score, is what lets you tell a client why the grade moved and point to the exact cause. It is also why the honest caveats stay in the open: a score that shows its edges is a score you can verify, and verification beats faith for anyone who has to defend the number.
For the full technical reference, the Radar methodology page walks through the audit dimensions, the scoring model, and how variance is handled. And if you want to see your own defensible score, the fastest path is to run the readiness check on your domain.
Ready to see a score you can defend?
- Run a free Radar check - Score your technical readiness; paid audits add the AI answer checks
- Read the methodology - How every dimension is measured and scored
- Talk to us - Bring a defensible AI visibility score to your team or clients
