Feature deep dive · Alerts
The engine said something wrong about you. Now what?
A wrong price, a invented certification, a warning buyers should avoid you — each sits in answers eroding deals until someone notices. The radar notices the same week, with the quote and the kit to answer it.
- 5 drift signals
- Correction kit per incident
- Weekly digest + cooldown
- Drift, not live
Short answer
What counts as an incident?
What triggers the radar?
Five severity-ranked signals on every processed answer: an engine warning buyers away from you, negative sentiment that usually means a hallucinated claim, a stance drop from first choice or recommended to alternative or mentioned-only, a competitor capturing first choice on a decision-stage query where you are absent, and a 5.0-point weekly visibility fall between two complete weeks.
Why is negative sentiment treated as a hallucination signal?
Because a negative description of a vendor is frequently factually wrong rather than merely unflattering — a price never charged, a feature removed years ago, a compliance claim never held. The radar pairs sentiment with the forensic gap reason, so genuine criticism routes to the queue while invented claims route to a correction kit.
Do alerts arrive instantly?
No — and response plans built on that assumption fail. Detection runs per answer and weekly between complete snapshots, delivered as an in-app severity-sorted radar plus a batched digest with a twenty-four-hour per-brand cooldown and idempotency keys. Weekly drift detection catches what matters without paging anyone over a one-point wobble.
The five signals
Ranked, so severity wins
One run produces at most one alert — the most severe incident wins — so a caution never competes with five minor slips for attention.
| Severity | Signal | What triggers it |
|---|---|---|
| Critical | Cautioned-against stance | An engine now warns buyers away from you. |
| Critical | Negative sentiment | A claim about your product reads negative — the hallucination signal. |
| Warning | Stance drop | First choice or recommended this cycle becomes alternative or mentioned-only. |
| Warning | Decision-stage conquest | A competitor takes first choice on a buying-stage query where you are absent. |
| Warning | 5.0-point weekly drop | Weighted visibility falls five points or more between two complete weeks. |
Eight-week view · illustrative
A 5.0-point fall between two complete weeks raises a warning. Partial weeks can never trigger — the threshold compares complete snapshots only.
Delivery
Where alerts reach you
In-app radar
Severity-sorted incidents with evidence, kit and resolution control. Always on.
Email digest
One batched digest, severity-ordered, twenty-four-hour cooldown, idempotency key.
Optional webhook
Slack delivery after digest success, plus DFY request notifications.
- 5
- Signals watched
- 1
- Alert max / run
- 24h
- Digest cooldown
- 5.0
- Macro threshold
Per processed answer
Most severe wins
Per brand, idempotent
Points, complete weeks
The correction kit
Five deliverables, generated from the incident
Not advice — copy, markup and a plan ending in a re-run.
01
Cleaned objection
The claim isolated and characterized, stripped of hedging and repetition.
02
Direct refutation
A forty-five-word answer that opens by naming the claim as inaccurate, then states the documented fact.
03
FAQPage JSON-LD
Markup answering whether the claim is accurate and what you are documented to offer — liftable verbatim.
04
llms.txt patch
A verified-facts section with the engine quote and a last-verified date, ready to append.
05
Five-step plan
Publish, mark up, answer, re-run the cycle, escalate if the claim persists across two cycles.
Then re-run the cycle
Publish, mark up, answer — then let next Monday confirm the correction landed. If the claim persists across two cycles, escalate: the kit becomes evidence for provider feedback, not just site work.
Radar questions
Asked by brand and legal teams
Quoted verbatim via FAQPage markup.
What triggers a defamation or hallucination alert?
Five severity-ranked signals: a cautioned-against stance, negative sentiment, a stance drop from first choice or recommended to alternative or mentioned-only, a competitor taking first choice on a decision-stage query, and a weighted visibility fall of five points or more between two complete weeks.
What is in a Factual Correction Kit?
Five deliverables generated from the detected incident: the cleaned objection, a forty-five-word refutation that opens by naming the claim as inaccurate, a FAQPage JSON-LD block, a verified-facts section you can append to llms.txt, and a five-step remediation plan ending in re-running the weekly cycle.
Are alerts real-time?
No. Detection runs on every processed answer and weekly between complete snapshots, delivered as an in-app radar plus a batched weekly digest with a twenty-four-hour per-brand cooldown. This is drift detection rather than live streaming, stated plainly so response plans are honest.
Catch it this week
Wrong claims compound until answered
Five signals, evidence attached, kits ready — refreshed every Monday with a digest that respects your inbox. Fourteen days free.