Brand Safety Monitoring Setup

Brand safety here means one specific thing: an AI engine stating something about you that is outdated, unverifiable or simply wrong, to someone who asked a reasonable question and has no way to check. It is not a keyword watchlist and there is nothing to configure — the analysis runs on every scan, on every answer, whether or not you go looking for it.

What you configure is where the findings reach you. This tutorial covers both halves: reading the findings, and making sure the serious ones do not sit unread in a dashboard.

Step 1: Know what is measured on every answer

When an engine answers one of your prompts, the analyser records — alongside mention, rank and sentiment — two extra fields that exist purely for this:

• Claims about your brand. Specific factual assertions the answer makes: pricing, features, funding, ownership, comparisons. Statements that could be checked, quoted verbatim.
• An accuracy risk rating. How likely the answer is to contain an outdated, unverifiable or inaccurate statement about you: none, low, medium or high.

Both are per answer, so the same engine can be clean on one prompt and risky on the next, and two engines can disagree completely on the same question.

This is the product’s answer to a well-documented failure mode rather than a novel one. The survey literature on hallucination in large language models — see Huang et al. (2023) — distinguishes factuality failures, where a model asserts something untrue about the world, from faithfulness failures, where it departs from its source. What shows up in your brand’s answers is overwhelmingly the first kind, and it tends to be plausible, specific and confidently worded, which is exactly what makes it worth monitoring instead of assuming you would notice.

Step 2: Open the findings

Go to Alerts in the Monitor group of the sidebar. Below the alert list is a Brand Safety card, populated from your latest completed scan.

Findings are surfaced when the analyser saw a real risk, not for every answer. An answer rated low risk appears only if it also carried quotable claims; medium and high always appear. A clean scan says so explicitly rather than showing an empty table.

The card sorts highest risk first, so the top row is where to start.

Step 3: Read a finding properly

Each finding gives you four things: a risk badge, the engine that produced it, the prompt that triggered it, and the claims themselves as a bulleted list — quoted, so you are reading what the model actually said rather than a paraphrase of it.

Read the prompt as carefully as the claim. A wrong statement produced by a hostile or leading question is a different problem from the same statement produced by “what does your brand do”, which is the question a genuine prospect asks. The second is far more urgent even at the same risk rating.

Note what is deliberately absent from this card: there is no safety score, no resolution workflow and no export. The findings are a read of one scan. Tracking a finding to closure is your process, not a field in the product — and the honest reason is that “resolved” is not something a scan can observe. What a scan can observe is whether the claim reappears in the next one.

Step 4: Verify before you respond

The first move on any finding is to check whether the claim is true. A meaningful share of them are, and reacting to a correct-but-unflattering statement as if it were misinformation wastes the week.

Work through it in this order:

1. Is it true today? Pricing, positioning and feature claims are the most common findings and the most likely to be true-but-stale — right eighteen months ago, wrong now.
2. Where could it have come from? Check the Citations page for the same scan. If the engine cited sources for that prompt, one of them is frequently the origin, and that is your fix.
3. How many engines say it? One engine is noise; three engines converging on the same wrong statement means it is in material they all reach, which is a much bigger job and a much clearer one.
4. Is it reachable by a buyer? A claim surfaced only by an obscure prompt is lower priority than the same claim on your main category question.

Record the prompt, the engine and the date alongside the quoted claim. Answers are generated fresh each time and are not guaranteed to reproduce, so the scan record is the evidence.

Step 5: Make sure findings reach you

Findings you never read are not monitoring. Two mechanisms carry them out of the dashboard:

• Warning-level alerts. A medium- or high-risk finding becomes a warning alert naming the engine and quoting the claim. So does an engine that describes your brand negatively overall, and an engine that failed to mention you on any tracked query.
• Alert emails. With “Alert emails” on in Settings → Notifications, a scan that produces any warning-level alert emails you regardless of your digest cadence. That is the setting that matters here: someone on a monthly digest still hears about a hallucination the day it appears.

Outbound webhooks carry the same alert text if configured. One caveat that catches people: a competitor benchmark scan notifies nobody on any channel, by design. And it goes further than the notifications — the Alerts page, the Overview, the Brand Safety card, the shared report, the print view and the public API all read your brand’s own home-market scan, so a benchmark’s brand-safety findings are not sitting in the dashboard waiting to be noticed either. They are recorded in full on the scan; reading them means asking for that subject explicitly, with GET /api/v1/scans?project_id=…&subject_type=competitor&subject=Acme, where every answer comes back with its extracted claims and accuracy risk. Plan the benchmark around that, not around a card you will go looking for.

Setting Up Smart Alerts walks through the delivery settings in full.

Step 6: Fix at the source, then re-measure

You cannot edit a model’s answer, and there is no channel for filing a correction. What you can change is the material an answer is built from.

• Correct the third-party page. If a cited source carries the wrong figure, that is the highest- leverage fix available, and it is usually the fastest — a directory listing or review-site profile with stale data is often a form submission away.
• Publish an unambiguous statement of the fact on your own site. One page, plainly worded, easy to retrieve. Vague marketing copy is exactly the input that produces a confident wrong answer.
• Retire or update the stale source you control. An old pricing page or announcement that still ranks is a supply of the wrong answer.
• Escalate as a communications issue when it is one. A reputational claim spreading across several engines is not an SEO ticket.

Then re-scan and check the same prompt. Treating this as a monitored loop rather than a one-off is the same posture the NIST AI Risk Management Framework asks of organisations deploying AI systems: identify, measure, and keep measuring, because the system’s behaviour changes underneath you. Here the system is somebody else’s, which makes the measuring the only lever you own.

Step 7: Where to go next

• Brand safety monitoring — the feature page behind this view.
• Understanding hallucinations about your brand — why models produce confident, specific, wrong statements.
• Fixing AI brand safety issues — the remediation playbook in detail.
• Defending brand reputation in AI — the communications side of the same problem.