Geo underperformers , named.
Location Performance Intelligence runs a weekly cron over your campaign location data, ranks every campaign × location pair by ROAS, and produces a list of split opportunities for the underperformers. Each opportunity comes with a Lyra-Agent-written reasoning, an estimated improvement, and a confidence score. Detection is automatic; the apply step is yours.
Every campaign, every location, every Sunday.
The sync service rebuilds location_performance_weekly each Sunday at 02:00 UTC, aggregating spend and conversion by campaign × location across the previous 7 days. The detection layer ranks tiers and emits opportunities to a queue you review the following morning.
Out: a split opportunity queue.
Below is a sample queue from the run above. Total opportunities flagged, distribution across the three tiers, and the top individual opportunities — each with the campaign × location pair, performance variance, estimated improvement, Lyra Agent reasoning, and confidence score.
Percentile ranking, Lyra Agent reasoning, your call.
The detection pipeline is rule-based: aggregate by location, rank by ROAS within each campaign, flag the bottom percentile. The Lyra Agent enters once a candidate is shortlisted to write the reasoning string and emit a confidence score. The apply step is human.
Top 20 / mid 60 / bottom 20 by ROAS.
Within each campaign, sort the campaign × location pairs by 7-day ROAS. Top quintile is overperforming; bottom quintile is the split candidate set. Mid 60% gets a watch flag if variance grows.
Location ROAS vs campaign baseline.
Performance variance = (location ROAS − campaign baseline ROAS) / campaign baseline ROAS. Negative variance flags locations dragging the campaign; positive variance flags locations carrying it.
How much a split is worth.
Project the recovery from splitting: location moves to its own campaign with its own bid strategy, CPCs drop to the auction baseline. Estimated improvement is in percentage points of ROAS, computed from variance × spend share.
Why this split, in human language.
Once a candidate is shortlisted by the rule layer, the Lyra Agent reads the structured opportunity (variance, spend share, conversion delta, locality context) and writes the reasoning string. The Agent never invents the variance; it narrates what the rules surfaced.
How sure are we this is worth it?
Confidence reflects: variance magnitude, spend share (volume signal), data freshness, and locality context (small countries get a confidence haircut). 0.7+ is high-conviction; below 0.5 is a watch flag, not a split.
Detected → approved → implemented.
Detection writes opportunities to a queue with status=detected. You move them to approved when you've reviewed; implemented when you (or the Lyra Agent in chat, if you ask) ship the split. All three states logged with timestamps and the user who acted.
The questions about how this actually runs.
When does the sweep run?
Weekly · Sunday 02:00 UTC. Each Sunday's run rebuilds the location_performance_weekly table and emits split opportunities to a queue you review the following morning. Newly connected accounts are picked up on the next Sunday sweep.
Why weekly and not daily?
Geographic ROAS needs a meaningful sample to be honest. Daily runs would emit noisy opportunities (single-day swings driven by inventory, not real performance gaps). Weekly aggregation produces stable signals worth acting on.
How are the tiers calculated?
Within each campaign, every location is ranked by 7-day ROAS into percentile tiers: top 20%, middle 60%, bottom 20%. Ranking is per-campaign, not account-wide — each campaign is its own ranking universe so you don't compare a brand campaign to a generic one.
Does the tool apply splits automatically?
No. Opportunities flow through a 3-stage status workflow — detected (auto, by the cron), approved (you, after review), implemented (you, when you ship the split in Lyra or Google Ads). All three states are logged with the user who acted. The Lyra Agent in chat can sequence the implementation if you ask.
What's the confidence score actually telling me?
How sure we are the split is worth doing. It blends variance magnitude, spend share, data freshness, and locality context. 0.70+ is high-conviction (act); 0.50–0.70 is mid (review and decide); below 0.50 is watchlist (revisit next week, don't act yet).
Run the audit.
Keep the findings.
Connect your account and let Lyra run its 18 tools for 14 days. If the projected waste recovery isn't worth at least 10× the $49, don't pay. You keep every insight either way.
Read-only connect · write access opt-in per tool · SOC 2 in progress · GDPR + CCPA compliant