How signals and recommendations are generated

Signals tell you what changed in a project, in order of importance. The full list is in Overview → Insights, and the top three appear on Paid media.

What it guarantees

• Deterministic. The same data gives the same list. AI plays no part in the calculation.
• Read-only. A signal asks you to check something. It does not change any campaign.
• Facts kept apart from hypotheses. The title and the values are measured. “Possible explanations” are hypotheses to verify and are labeled as such.
• No budget advice. No signal tells you to raise or cut a budget.
• No mixing. Leads, purchases and Google conversions are not added together. A campaign with an unknown objective gets no signals about results. A missing value does not become zero.

Which periods are compared

On the signals page, the last 7, 14 or 28 complete days (14 by default) are compared with the same number of days before them. Today is not included. On Paid media, the period and comparison chosen on the page are used, so that a signal does not contradict the figure next to it.

A platform is compared only if it has data on at least 60% of the days, in both periods. Otherwise the page says why it is not compared. A campaign under 3% of the platform's cost gets no signals, and each campaign gets at most one signal: the most serious one.

The rules

Confidence

Every change goes through a statistical test. Confidence is high at z ≥ 2.58, medium at z ≥ 1.96 and low below that threshold. The test tells you that a difference is real, not why it happened. Conversions from the last few days can still grow, which is why reporting delay appears among the possible explanations.

Priority

The order of the list comes from the score: severity + confidence + what is at stake. Severity: critical 40, tracking and data 30, watch 20, opportunity 15. Confidence: high 15, medium 8, low 2. What is at stake: the estimated extra cost, as a percentage of the platform's cost, up to 30.

“Why was this flagged?”

A signal tied to a product, a campaign, a platform, a dataset or a term opens on its own investigation page. The page recalculates everything from the stored data and shows:

1. the rule and its thresholds, and whether the figures, read again now, still meet the condition;
2. the measured facts for the signal's period and for the reference period;
3. the daily chart; a day that was not imported breaks the line, it does not become zero;
4. the related campaigns and their search terms, with the association level;
5. the conversions measured by the Admetriva tracker, separate from those reported by Google Ads;
6. the product data and Merchant issues;
7. data quality: imported days, gaps, products missing from the catalog;
8. possible causes, labeled as hypotheses, each with how to check it;
9. recommended checks, each with its reason and the exact link.

For a product, “Product 3407” means product ID 3407 in Merchant Center, as Google Ads reports it. It is not a SKU.

Signals and recommendations

Signals do not carry over from one day to the next: they have no “resolved” or “ignored” states. The recommendations on the same page come from the project's rules (for example terms or products with cost and no conversions, feed issues) and have the actions “Mark as applied”, “Dismiss” and “Copy”. They are proposals too: Admetriva does not apply them on its own.

Limits

• Two equal periods are compared, with no adjustment for seasonality or holidays.
• Individual ads and creatives are not yet analyzed by signals.