Every bid change, reasoned .
AI Bid Optimization Engine reads your campaigns, computes the rule-based signals (CPA delta, CTR, conversion rate) that justify a bid change, and asks the Lyra Agent to write the reasoning per row. You preview the full recommendation list, set bid caps, then apply — or stay in analyze-only mode and never write to Google Ads at all.
Five fields. Five hard rules.
AI Bid Optimization runs on a single account at a time. The five fields below shape the run — the mode you pick decides whether anything writes to Google Ads. Default is preview; nothing leaves Lyra unless you explicitly choose execute.
Bid changes, with the why , per row.
Each recommendation is a keyword line plus a reasoning paragraph from the Lyra Agent. Editable before execute. Below is a 3-row sample from a 31-row preview run on a Manual-CPC Brand · Search campaign.
Heuristics produce the number. The Lyra Agent writes the why .
The bid math is deterministic — same inputs, same recommendation, every run. The Agent only writes the reasoning string. We don't ask it to set bids, because asking an LLM to do arithmetic is a bad use of it.
Rule-based signals decide the bid number.
CPA delta, CTR, conversion rate, cost data — all computed in compute_bid_signals() from keyword_bid_snapshots and campaign_performance. Deterministic; same campaign state always produces the same recommendation. No ML model deciding the bid.
Source · keyword_bid_snapshots · campaign_performanceThe Lyra Agent writes the explanation per row.
Once a recommendation is shortlisted, the Lyra Agent is asked to explain the heuristic output in human language — what changed in the signals, why this delta, what the projected impact looks like. The Agent never invents signals; it only narrates.
Source · Lyra Agent · constrained to signal dataSmart Bidding campaigns get analysis-only.
Only Manual CPC and Enhanced CPC campaigns get bid recommendations. Smart Bidding (tCPA, tROAS, Maximise Conversions) gets analysis-only output — Google's algorithm controls those bids, so a keyword-level mutation would be ignored or overridden.
Source · KEYWORD_BID_ELIGIBLE_STRATEGIES filterBids on a leash , reasoning on the record.
Each row: keyword, current bid, suggested bid, ±%, plus a Lyra-Agent reasoning paragraph grounded in the heuristic signals. Editable before execute, logged after.
Analyze returns signal data only — no bids changed. Preview returns the full draft list. Execute applies via Google Ads BatchJob. You pick the mode every run; no scheduled cron triggers it.
Every recommendation respects your per-keyword ceiling and floor. Cooldown days block back-to-back adjustments to the same keyword, even if the signals would justify one. Both are hard rules; recommendations that would breach either are dropped before you see the list.
Smart Bidding campaigns get analysis-only output — Google's algorithm owns those bids. No surprise mutations on Smart-Bid campaigns; the tool surfaces signal data so you can decide whether to switch strategies.
The honest answers about the engine.
Does the tool apply bid changes automatically?
No. Three modes — analyze (signals only), preview (full draft), execute (applies). You pick the mode at run time. Default is preview. Execute writes to Google Ads via the BatchJob API; no scheduled cron triggers it. The tool only runs when you ask it to, from the AI Advisor chat or directly via API.
Where does the AI actually sit?
The Lyra Agent writes the per-recommendation reasoning. The bid math itself is deterministic — heuristic signals over your campaign data, no ML model deciding the number. The Agent is asked to explain a recommendation, not produce one. Asking an LLM to do arithmetic is a bad use of it. (See AI Advisor for what powers the Agent under the hood.)
What stops a runaway adjustment?
Two things: bid caps (per-keyword ceiling and floor you set) and cooldowns (a minimum number of days between adjustments to the same keyword). Both are hard rules — if a recommendation would breach either, it's dropped before the list is shown to you.
Why doesn't this work on Smart Bidding campaigns?
Because Google's algorithm controls bids on Smart Bidding strategies (tCPA, tROAS, Maximise Conversions). We can't propose a keyword-bid change that the system will respect. For those campaigns, the tool runs in analyze-only mode and surfaces the same signal data so you can decide whether to switch strategies.
What's logged after a run?
Every applied change writes to bid_change_history (per-keyword before/after/reason) and an entry to account_diary_entries summarising the top increases and decreases. Reverting an apply means running the tool again with the inverse delta — there's no one-click rollback on this one.
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