Google Grey Ops

Google Grey Ops

Reviewed 2026-09-02. Baseline: Sonnet 5 / Claude Code subagent / 2026-09-02; 2026-09-03 blind cascade task [A] — payment+email cascade named only when prompted. Launch/operate via API on your own MCC credentials (googleops, 10) + survive aggressive verticals without losing accounts, domains, payment identities.

googleops/gmcops are NOT yet live-verified (2026-09-03): offline tests only, no run against a real customer. A human watches the first live plan → apply; record the result in 10.

Start here

Agent writes a JSON spec; googleops validates, creates PAUSED, reads back, activates. Never hand-assemble a MutateOperation or call a mutate outside the CLI. New shape → extend gads_spec/gads_build/gads_verify and test_googleops.py.

Boundary: "buy well" → google-ads · "don't get killed / source accounts / scale" → here · "feed and Merchant Center" → google-feed-ops · "count and sync" → tracker-ops · "portfolio and TL decisions" → senior-buyer-ops.

Authority: this skill governs grey-vertical execution; google-ads clean-marketing guardrails govern compliant accounts. Route by lane; do not merge the two normative stances.

The Google model is not the Facebook model

Discipline transfers from meta-grey-ops; the risk map does not. They differ:

Full 7-row comparison (destination review, cascade scope, burn radius, moderation timing, appeal posture) → references/04-failure-forensics.md.

On Facebook you protect the session. On Google you protect the billing identity and the destination. Porting the Facebook playbook wholesale over-invests session hygiene, under-invests payment consistency and landing-page integrity.

Non-negotiables

1. Classify the enforcement track first. Egregious (no warning, permanent, propagates) · non-egregious (mandatory 7-day warning) · Limited Ad Serving (throttle, not suspension). Most costly error: treating recoverable as terminal or vice versa. See 04.
2. Know which replacement you are doing. New Google identity after your suspension = Circumventing-systems violation, separately charged, propagates. Reseller swapping the seat (their MCC, their invoice) is a different product — still cascade-risk if GTM, domain, phone, or payment profile overlap. See 05.
3. Never submit false information during verification. Failing honestly pauses the account; lying suspends it permanently — fabrication is never lower-risk. Org path skips the selfie; grey self-reg does not pass BOV — 06.
4. Never chargeback. Named account-level suspension trigger; burns the payment method's future usability.
5. Sequence payment events deliberately. No card add, method change, or threshold charge from an inconsistent context.
6. Rotate on signals, not a timer; change one variable at a time — else the next burn is unattributable.
7. Know whose MCC you are in. Cascade takes every account sharing it. Reseller who can't name the MCC owner = unknown compliance history inherited.

Route references

Scripts

googleops is the agent write boundary (from this folder: uv tool install ., or prefix uv run --project .). Workspaces inside the skill store are rejected; state lives in the project's .googleops/.

Env: GADS_DEVELOPER_TOKEN + (GADS_CLIENT_ID/GADS_CLIENT_SECRET/GADS_REFRESH_TOKEN or GADS_JSON_KEY_FILE), GADS_PROXY or GADS_ALLOW_NO_PROXY=1; GMC_JSON_KEY_FILE or GMC_REFRESH_TOKEN for gmcops. Never pass a credential on argv.

Policy taxonomy, certifications, and the full appeal process live in google-ads/09 — read it before any regulated vertical.

New-job bootstrap

1. Collect access — MCC/account IDs, billing arrangement + owner, Merchant Center, tracker campaign URL, domains, proxy. Into gitignored project notes, verbatim.
2. Establish the enforcement baseline — policy status, active strikes, verification state, Limited Ad Serving status (needed to diagnose later changes).
3. Confirm certification requirements per target geo (google-ads/09) before building. Gambling and financial services changed multiple times in 2026.
4. Wire tracking before spend. Tracking template with {lpurl}/{gclid}, tracker campaign mapping; check whether the developer token can still onboard offline conversion import at all (2026-06-15 cutoff, google-ads/06).
5. Naming before first launch. Campaign name must encode whatever the tracker splits on; mapping is not automatic (tracker-ops).
6. Launch through googleops (00): doctor → plan (validate_only) → apply PAUSED → verify → activate.
7. Review layer. Final URL = the white, same registrable domain, cloak off until serving. No PMax cloak. No App-campaign cloak (08). Filter stack in 05. Confirm tz/currency and OFAC vs serve-geo (07) before first spend.
8. Agree kill rules in writing with the TL: spend-without-conversion cap, CPA cap, account verdict threshold, who decides on a ban wave.

Diagnose a failure in order

which track (A/B/C) -> account-level or item-level -> payment/verification or policy/content
-> domain-specific or account-wide -> then act

Do not skip to remediation. A domain rotation on an account-level signal problem wastes a domain and teaches nothing.

Guardrails

• Redirects are not the violation. Content that differs conditioned on who is requesting it is — that's Circumventing systems (no warning, permanent, cascade). Filter stacks/same-domain whites (05) don't make the track non-egregious.
• Compromised site = disapproval tier. Malicious software = egregious. Not the same death sentence.
• "Temporarily paused" ≠ "suspended". Verification pauses recover by completing the task.
• Multiple accounts per business are not banned; the violation is the evasion pattern, not the count.
• Vendor pricing, spend ceilings, appeal-success claims in these files are vendor-reported, no published methodology — treat every number as a prior, not fact.
• Where a compliant path reaches the same goal, say so — 01 maps each grey practice to its legitimate alternative; several are cheaper long-term (trust accrues to you, unrevocable by a third party).

Output

Lead with the track classification and whether the state is recoverable. Give the exact next action, the review timeline to expect, what to observe before the next change, and what single variable is being changed so the result stays attributable.