Setting Up Smart Alerts

Most monitoring tools ask you to build alert rules: pick a metric, pick a threshold, pick a comparison. There is no rule builder here, and that is deliberate. Every completed scan is analysed against your previous one and the notable findings become alerts on their own. What you configure is which findings are worth telling you about, where they go, and how often — not the magnitude that counts as a change. That last one is ours, and Step 1 explains why we would rather own it than hand you a number to guess at.

That makes setup short. It also makes a handful of quiet non-delivery cases worth knowing about, because “I never got an alert” almost always turns out to be one of them rather than a bug. Step 6 lists them.

Step 1: Learn what actually generates an alert

Alerts are derived from a completed scan and the one before it, so the set is fixed and predictable. It splits in two, and the split is worth holding onto because it decides what a silent inbox actually proves.

State in this scan — these need nothing but the scan in front of them, so they fire on your very first one:

• Strong overall visibility — a success alert when the scan’s headline score is high.
• An engine that mentioned you on nothing — a warning, per engine. This is the one that catches a surface going quiet.
• An engine that describes you negatively overall — a warning.
• A possibly inaccurate claim about your brand — a warning, naming the engine and quoting the claim. This is the brand-safety path.
• A competitor out-mentioning you across your tracked queries — informational, naming them. Note this is a level: they appear more often than you in this scan, whether or not anything changed.

Movement since your previous scan — these compare the completing scan against the one before it, so they are silent on a first scan and after any gap the comparison cannot bridge:

• Your visibility score moved — a warning when it fell, a success when it rose, quoting both figures.
• A competitor gained share of voice — a warning, quoting the before and after. Distinct from the level alert above: this one is news.
• A source is being cited that wasn’t before — a success when it is your own domain, otherwise informational, naming the domains. It is a one-directional set difference, so it announces a source appearing and has no equivalent for one going quiet.
• Engines describe you less positively than before — a warning. Triggered on the mean sentiment across delivered answers, but reported as a count of negative answers, because that is the vocabulary the analyzer actually produces.
• An engine that used to mention you stopped. This is not a separate alert — it is the mentioned-you-on-nothing warning above, worded more sharply because a baseline exists. One finding, one alert, whether or not there was a previous scan to compare with. It is listed here because the wording changes, not the trigger.

Each of the movement alerts fires at a fixed magnitude — so many points of score, so many points of share of voice — and you cannot edit those numbers. That is on purpose rather than unfinished: a threshold you pick is a threshold you guessed at, and AI answers are non-deterministic enough that a badly-guessed one produces either a channel you learn to ignore or a silence you trust wrongly. The live figures are printed next to each toggle in Settings, so read them there rather than from this page — that copy is generated from the same constants the engine fires on and cannot drift from them, which a number typed into a tutorial can.

One more, in neither group:

• A high-volume tracked prompt you are absent from — a warning; and a newly added high-volume prompt that has not been scanned yet — informational. Both of these are Alerts-page only. They are derived from your tracked-prompt volume data, which only the Alerts page loads; email and the webhook build their alert list from the scan alone, so neither of these two ever triggers a notification. See case 5 in Step 5.

Severity is what drives delivery for the alerts that are delivered at all, so it is worth internalising: warning is the level that jumps the queue in Step 3. Reading the two lists back with that in mind: a new citation and a rising score are genuinely good news and deliberately do not interrupt you, which is why they show up on the Alerts page and not in your inbox.

Settings → Notifications carries a switch for five of these — score change, new citation, hallucination (which covers both the inaccurate-claim and the negative-sentiment warnings), competitor spike and sentiment shift. All five default on. Switching one off stops it reaching your channels; it does not stop it being derived, so the Alerts page keeps showing everything either way.

Three findings have no switch, and one of them matters. The strong-score success and the competitor-out-mentions level alert have none because neither emails anyone, so there is nothing to turn down. The engine-that-mentioned-you-on-nothing warning also has none, and it does email you. That is deliberate rather than an oversight: a surface going silent on your whole prompt set is the single finding this product exists to catch, and it is not something a customer should be able to switch off by accident while tidying up a noisy inbox.

Step 2: Open Notifications and pick an email cadence

Go to Settings → Notifications. The tab sits on the workspace Settings page alongside Members, Domains and Branding — but unlike those three, the settings themselves are per user. Each member configures their own delivery, and a teammate can neither see nor change yours. A per-scan notification follows one person’s settings — whoever ran the scan, or for an automated scan the project’s creator, falling back to a workspace owner. Weekly and monthly digests are the exception: every member who opts in gets one covering the whole workspace.

Turn on Email digest and choose one of four cadences:

• After each scan — a summary email every time a scan completes.
• Weekly (Mon) or Monthly (1st) — batched instead, assembled by a scheduled job rather than by the scan itself.
• Off — no digest at all.

Choose by scan cadence, not by preference. Automated scans run weekly on every paid plan — the cadence is uniform rather than a tier perk — so “after each scan” is a weekly email plus one for each scan anybody runs by hand. On the free plan there is no automated monitoring at all, so it is an email only when someone presses Run scan. Match the two or you will either drown or forget the product exists.

Digest and alert emails carry RFC 8058 one-click unsubscribe headers, so a recipient’s mail client can opt them out in one action and the opt-out is honoured server-side.

Step 3: Turn on alert emails so warnings jump the queue

Alert emails is a separate switch, and it is the important one.

With it on, a scan that produces any warning-level alert emails you regardless of your digest cadence. Someone on a monthly digest still hears about a hallucinated claim or a silent engine on the day it appears.

The two settings can both fire for the same scan; you still receive exactly one email, because the per-scan email is de-duplicated on the scan. Turning the digest off and alert emails on is a perfectly good configuration — arguably the best default for a busy team — and it means the only mail you receive is mail that needs a decision.

Step 4: Wire an outbound webhook into your own tooling

Outbound webhook URL receives a JSON POST on each completed scan, carrying the event name, brand and domain, the score, the summary, the per-engine results and the alert list.

This is the integration path for anything the product does not do itself: a ticket per warning alert, a row in a warehouse, a metric in your own dashboard. Because it is the full structured payload rather than rendered text, you can apply your own thresholds downstream — which is where threshold logic belongs, since it is specific to your business.

Deliveries are queued jobs rather than fire-and-forget calls. A failed POST is retried with backoff, and both the failure and the retry are visible on Activity in the Workspace group, where a stuck or dead job can be restarted by hand. Restarting re-drives the same job rather than creating a new one, which is what keeps a scan from being charged twice — but that is not de-duplication of delivery. Make your receiver idempotent. Outbound webhook POSTs carry no idempotency key at all, so both the automatic backoff and a hand restart re-post; a receiver that processed the payload and then answered a 500 will be sent it again. The payload identifies the scan, so keying on that is enough. (Email is the one channel with a provider-side idempotency key, and even there the window is finite — restarting a dead job days later re-sends.)

Step 5: Know the cases where nothing arrives

Between them these explain almost every “why didn’t I get alerted” question:

1. A competitor benchmark scan notifies nobody, on any channel. Not email, not the webhook. A digest reading “Competitor AI visibility 74/100” delivered to your team as if it were your own result would be worse than silence. Read competitor results in the dashboard.
2. A scan that fails sends nothing. Notifications are built from a completed scan. A failed or dead job surfaces on Activity, not in your inbox.
3. A suppressed email address is skipped. A hard bounce or a spam complaint recorded against an address adds it to a suppression list, and later mail to it is dropped rather than retried — re-sending to a dead mailbox damages your sending reputation and reaches nobody. The skip is logged, so it is discoverable rather than silent.
4. A blank webhook URL produces no delivery and no job — nothing to see anywhere. Email is not symmetrical with that, and it is the one people misdiagnose: the job is enqueued regardless of whether the deployment has an email provider configured, so if it is unset (or the provider is simply down) the send fails, retries with backoff and eventually dead-letters. Look on Activity rather than assuming a silent skip.
5. The two tracked-prompt alerts never leave the Alerts page. A high-volume prompt you are absent from is warning-level and still does not email you, because the volume data it is derived from is loaded only by that page — the email and webhook paths build their list from the scan alone. This is the one exception to Step 3’s “warnings jump the queue”, and the reason to open Alerts occasionally even on a well-configured account.
6. A per-scan notification follows one person’s settings, not the team’s. Notification settings are per user (Step 2), and a scan notifies the member it is attributed to: whoever ran it, or the project’s creator for an automated scan. If a colleague’s scan produced nothing in your inbox, that is why — shared delivery needs the same settings on each member whose scans should reach it.
7. A first scan produces no movement alerts, and neither does the first scan in a new market. The baseline is the project’s previous completed brand scan in the same region, which is a deliberately narrow definition: a competitor benchmark scan and a foreign-market companion scan both land with fresh timestamps, and differencing your UK numbers against your US ones would manufacture a large, confident, meaningless delta. With no eligible baseline the five movement triggers simply stay quiet; the state alerts in Step 1’s first list still fire as normal. The same thing happens, harmlessly, if the lookup fails outright — a missing baseline reads as “nothing to compare”, never as “nothing changed”.
8. The alert-type switches gate your per-scan delivery, not the batched digest. Turning Competitor spike off stops it reaching the email or webhook fired by a completing scan. The Weekly and Monthly digests are assembled by a scheduled job that summarises each project’s latest scan against its predecessor for the whole workspace, and it does not read your per-type switches — so a type you have switched off can still appear in a digest line. If you want it gone entirely, the digest cadence is the setting to change.

Step 6: Where to go next

• Real-time alerts — the feature page behind this.
• Alert strategy for AI monitoring — choosing what to be interrupted for.
• How often should I monitor AI visibility? — matching cadence to how fast your category actually moves.
• Brand Safety Monitoring Setup — the findings behind most of the warning-level alerts.
• Setting Up Tracked Prompts — high-volume prompts you are absent from are an alert source on the Alerts page, so the prompt list shapes what you find there.