--- name: manage-google-recommendations description: >- A monthly, strategy-first pass over the Google Ads Recommendations queue: pulls everything outstanding via gads_list_recommendations, audits the last 30 days of change history for auto-applied changes that slipped in, evaluates each recommendation category against the account's actual strategy (repairs get benefit of the doubt; spend-increasing and automation-expanding suggestions get skepticism), and executes only the accepted ones through matching VigilDog write tools such as gads_add_negative_keywords or gads_create_sitelink. Use it monthly, before meetings with a Google rep, or when optimization score becomes a stakeholder topic. Do not use it for daily breakage detection (run-a-daily-account-health-check) or for diagnosing a performance change an auto-applied edit may have caused (investigate-performance-anomalies). --- # Review Google's Recommendations Without Ceding Control ## Purpose The Recommendations page mixes genuinely useful repair prompts with suggestions whose primary beneficiary is auction revenue, and the optimization score attached to it creates pressure to accept things that conflict with your strategy. This skill runs a disciplined monthly pass: verify nothing is being auto-applied behind your back, judge every suggestion against your own plan, act on the few that earn it, and keep the score presentable without letting it steer the account. ## When to run - Monthly, as a standing section of the monthly performance review. - Before a call with a Google account representative (arrive knowing your queue and your answers). - When a stakeholder raises the optimization score. - When an unfamiliar recommendation type appears and needs a documented position. ## When NOT to run - Daily scanning for breakage, run-a-daily-account-health-check already skims repair items. - Performance moved and an auto-applied change is a suspect, investigate-performance-anomalies owns that diagnosis (this skill's change-history audit feeds it). - You independently decided to build sitelinks, add negatives, etc. just use the relevant build skill; no recommendation needed as justification. ## Prerequisites - Read access to the account via VigilDog. - A current statement of strategy: bidding approach, match-type policy, budget ceiling, creative standards. Recommendations are judged against this, so it must exist before the pass. - Last month's recommendation log, if one exists, for score and pattern comparison. ## Procedure Consult `gads_policy_guardrail` before the first Google Ads write of the session. All writes are previewed first (validate_only=true default) and applied only after explicit user approval. 1. HUMAN STEP (outside VigilDog): verify auto-apply settings in the Google Ads UI (Recommendations page → auto-apply controls). Expected posture: every category off, with at most a deliberately chosen subset of repair-type items enabled. VigilDog cannot read or set these toggles; record the user's confirmation. 2. Audit for silent changes regardless of step 1's answer. Pull `gads_get_change_history` for the last 30 days filtered to auto-applied recommendation sources. For each hit, judge against strategy: - Consistent with what you would have done → note it and move on. - Conflicts with strategy → queue a revert through the owning skill's write tool (preview → user approval → apply), and have the user re-check the auto-apply toggles. 3. Fetch the queue with `gads_list_recommendations`. Group by category and impression/spend context so high-leverage items surface first. 4. Evaluate repairs first, they have the best hit rate. Broken conversion setups, disapproved entities, and genuinely conflicting negative keywords are usually worth fixing. Dismiss a repair when the "problem" is deliberate design (intentional exclusions, deliberately empty structures). 5. Evaluate bidding and budget suggestions with the default posture of skepticism. Accept only when the suggestion matches a change already justified by your own data (e.g., you separately concluded the target should loosen). Reject anything whose only argument is "spend more" or "loosen targets" absent business context. 6. Evaluate keyword and targeting suggestions. Accept keyword ideas that independently pass your own research standards, and negative-keyword-conflict fixes that verifiably block wanted traffic. Reject broad-match migration pushes, audience expansion, and search-partner enablement unless they are already part of your plan. 7. Evaluate ad and asset suggestions. Accept concrete asset gaps you meant to fill anyway (missing sitelinks, callouts). Reject auto-generated copy, AI-modified creative, and dynamic asset enablement that bypasses your creative standards. 8. Treat new-campaign-type suggestions (new automated campaign launches) as strategy proposals, not queue items: the default is reject; adopting a new campaign type happens through its own planning skill with its own economics, never as a one-click acceptance. 9. Execute accepted items through VigilDog where a matching tool exists, e.g., `gads_add_negative_keywords` for conflict fixes, `gads_create_sitelink` for missing sitelinks, other build skills for larger accepted items. Preview first, apply only after user approval. HUMAN STEP (outside VigilDog): pressing Apply or Dismiss on the recommendation cards themselves is UI-only, VigilDog has no apply/dismiss endpoint, so the user clears the cards: Dismiss for everything rejected, Apply-equivalent already done via VigilDog for accepted ones (dismiss the card afterwards to clear it). 10. Manage the score honestly. Note the score before and after. Because dismissing raises the optimization score exactly as much as applying, the score-pressure answer is always "dismiss the misaligned items", never "accept them to look good". 11. Log the pass: counts applied / dismissed / deferred with one-line rationales, auto-applied findings, score movement, and any new recommendation type encountered with the position taken. Next month's pass starts from this log. ## Decision rules - Default postures by category: repairs, lean accept; bidding/budget, lean reject; keywords/ targeting, reject expansion, accept verified conflicts; ads/assets, accept gaps, reject generated copy; new automated campaign types, reject in this context. - The only universal acceptance test: would you make this change if Google had not suggested it? If no, dismiss. - Dismiss and apply move the optimization score identically, so a target score never justifies accepting anything. - Budget-increase suggestions additionally require confirmed business capacity (inventory, fulfillment, cash) before acceptance; "headroom in the auction" is not capacity. - Any auto-applied change found while toggles were believed off is a standing agenda item until the source is identified. - Defer (leave uncleared) only items that genuinely need data you lack; a queue that rolls over month after month means decisions are being avoided. ## Common failure modes - Score-driven acceptance: applying misaligned suggestions because a rep or dashboard framed the score as a KPI. Dismissal achieves the same score with zero strategic damage. - Assuming auto-apply is off because it was off once. The change-history audit in step 2 is the control that actually catches drift. - Blanket cynicism: dismissing everything unread and missing real repair items, broken conversion configurations surface here before they surface anywhere else. - Accepting keyword expansions that quietly rewrite match-type policy one suggestion at a time. - No written rationale, so next month re-litigates the same cards from scratch. ## Related skills - run-a-monthly-performance-review, the parent cadence this pass belongs to. - run-a-daily-account-health-check, daily detection of auto-applied changes between passes. - investigate-performance-anomalies, when an auto-applied change correlates with a KPI move. - analyze-search-term-reports / set-up-ad-extensions, the independent standards accepted keyword and asset suggestions must pass.