Help · How Adintel decides
How anomalies are found
Each campaign judged against its own normal, the line a departure has to cross, what makes one serious, and what the explanation may and may not do.
Reviewed 22 October 2026
What is a campaign's normal?
Its own recent history, figure by figure: the average and the usual spread. Spend, conversions and conversion rate are judged against the last 28 days, a stable normal, so a fall is not hidden by a weak last week. Clicks, impressions and CTR are judged against the last 7 days, so a deliberate recent change, such as new ads or a new budget, becomes the new normal quickly. Each campaign has its own normal. The account as a whole is judged too, on its conversions, value, ROAS and cost per conversion, by the week (below).
When is a move an anomaly?
When the day's figure sits 1.5 standard deviations or more from its normal. How far past that line decides its severity: Low from 1.5, Medium from 2, High from 2.5, Critical from 3. Nothing is judged where a campaign's figure barely varies (any change would look huge), where there is too little history yet, or on a day when the figure cannot be worked out, such as a conversion rate with no clicks.
Why are value, ROAS and cost per conversion judged on a week?
Because a single day of sales is too uneven to judge: on a store making a few sales a day, one day's ROAS swings so widely that only good days could ever stand out. So these are judged on the 7 days ending on the newest day whose conversions have come in, against the 7-day figure ending on each of the 28 days before. A week's ROAS is judged only where its spend would buy at least three conversions at the campaign's own cost per conversion over the 28 days before (the account's, for the account as a whole), and cost per conversion only where the week recorded at least three. Value and ROAS are judged only where the four weeks before recorded value on about three days a week or more.
Why are conversions judged a day later than clicks?
Conversions keep arriving for a while after the click. So conversions and everything built on them are judged on the latest day that closed at least 24 hours earlier, or up to three days earlier on an account where Adintel has measured that conversions arrive late. Clicks and spend are judged on yesterday. Where you have marked real outcomes on Conversions, conversions here means those actions.
What is a run?
When the next day also crosses the line in the same direction, the departure continues, and the card says how many days it has run and since when. A run that qualifies is put forward to your briefing on its first day, again when it grows more severe, and on its fifth day running, or its eighth for a figure judged on a week (how the briefing chooses its items).
What does the language model do?
It writes the explanation, after the rules have found the anomaly. It is given the figures, the evidence worked out by rule (where the change concentrated, what moved with it, and what changed in the account over those days and the 14 days before them) and the business context. It writes one explanation for each departure put forward to your briefing; where several come in as one, it explains the strongest. It cannot create an anomaly, remove one or change how serious it is. If it fails, the card shows the figures and evidence alone.
Still need a hand? Email support@adintel.io and a person replies.