AI NativeBrand guardrails are rules checked automatically against every live scan that tell you whether AI answers get your brand right, beyond just whether you appear. See Brand Truth Studio for where facts and guardrails are created together.
What guardrails measure
Guardrails run against the answers from a scan and compute a per-guardrail compliance rate: the share of relevant answers that satisfy each rule. Four numbers appear at the top of the page: the overall compliance score, failing rules, at-risk rules, and passing rules. An amber banner appears when any rule is failing.
The four rule types
Required - a key positioning term must appear in answers where your product is named. Pass when the share of branded answers containing one of the rule's terms meets the threshold.
Prohibited - a damaging term must not appear in branded answers. Pass when branded answers containing the terms stay below the threshold complement.
Tone - branded answers must carry a non-negative sentiment score. This rule ignores the terms list and reads the scan's computed sentiment score.
Accuracy - factual claims in answers must match verified ground-truth facts. Pass when the share of answers rated accurate meets the threshold.
Evaluation runs deterministically against scan answers without calling an additional model.
Pass, at-risk, and fail
Each rule lands in one of three states after a scan:
- Pass - compliance meets or exceeds the threshold.
- At risk - compliance falls within 15 percentage points below the threshold.
- Fail - compliance falls more than 15 percentage points below the threshold.
Rules with no relevant answers (for instance an accuracy rule when no accuracy scores exist yet) show a "no data" state rather than a false pass or fail.
Severity
Each guardrail carries a severity of high or medium. Severity does not change the pass/fail calculation or the overall compliance score; it is a prioritisation label that surfaces which failing rules to act on first.
Adding and editing rules
Click "Add guardrail" to open a form for label, kind, terms (comma-separated, optional for tone and accuracy), pass-at threshold, and severity. Existing rules can be edited inline via the pencil icon on hover. Deletions are also available from hover. Edits take effect from the next scan.
Reading the failing answers
When a rule fails or is at risk, its card expands to show the answers that violated it: engine, prompt, and a short excerpt. This gives you the source material to act on rather than an aggregate number alone.
Questions
What is the compliance score?
The compliance score is a summary percentage: the average of the individual compliance rates across all your guardrails. A score of 80 or above turns green; 60 to 79 turns amber; below 60 turns red.
Does a new guardrail affect past scans?
No. Guardrails are evaluated at scan time. A newly added rule will show "no data" until the next live scan runs and produces answers to check it against.
Can I set different thresholds for different rules?
Yes. Each rule has its own pass-at threshold set when you create or edit it. The default is 0.70 (70%). You can raise or lower it per rule to match how strictly you want that specific requirement enforced.
What is the difference between required and accuracy?
Required checks whether certain terms appear in the answer text. Accuracy checks whether the claims in the answer match the verified facts in your brand's truth record. Required is keyword-based; accuracy uses the answer's computed accuracy state from the scan.
Does tone use my terms list?
No. The tone rule ignores whatever is in the terms field and instead reads the sentiment score computed for each answer during the scan. A score of zero or above counts as non-negative.
What does high severity mean for a failing rule?
Severity is a prioritisation label. It does not change the pass/fail calculation or the overall compliance score. It signals which failures are most important to fix first.
Can I add guardrails from the Brand Truth Studio page?
Yes. The Brand Truth Studio page includes a quick-add guardrail form and shows the current list. The full compliance view with per-answer breakdowns is only on the dedicated Brand Guardrails page.