v2.4 Agent-native release · now with Performance Max support

Your autonomous
media buyer for Google Ads.

An AI agent that operates your accounts end-to-end — diagnoses problems, decides on fixes, and executes them inside Google Ads. With audit logs, approval gates and full revert.

See how it works
Free for individuals Windows 10/11 · 52 MB Connects to your account via OAuth
adsmith-agent acme-corp · production
agent.log
accounts
rules
Accounts
acme-corp.com Live
acme-eu.com paused
acme-japan.jp Live
Active Tasks
Budget rebalance
Negative kw mining
Creative rotation
Settings
Approval gates
Audit logs
Agent · running
acme-corp.com · last 2 hours
14:02:31SCANDetected anomaly in Brand-US-Search · CPC +47% vs. 7d baseline
14:02:33SCANRoot cause: 3 new bidders on keyword "acme platform" · competitor "rival.io" active
14:02:36PLANRecommend: shift -$240/day from Display-RT to Brand-US-Search · forecast +18 conv/day
14:02:38EXECApproval gate cleared (auto-policy: brand-protection) · changes applied
14:18:02CHECKVerified: Brand-US-Search impression share 82% → 94% · holding
14:42:11FLAGPerformance Max PMax-Apparel drifted off-brand · 6 placements quarantined
14:42:15PLANDrafted 4 negative-placement rules · awaiting human approval
15:01:48SCANAnalyzing search-term report (2,847 new queries)
Actions today
147
▲ 21 auto-applied
Wasted spend stopped
$3,840
↓ -12% MoM
Pending approvals
4
2 high-priority
Proposed: Pause 12 underperforming ad groups in EU-Display impact_estimate: -$1,240/wk spend · -3 conv · roas_lift: +0.6x
// live agent activity · last 60 seconds across all customer accounts
14:32:08EXEC · Budget shift $240/d Display → Search · northwind-retail
·
14:32:11FIX · 14 negative keywords applied to PMax-Apparel
·
14:32:14FLAG · CPC anomaly on brand-eu · +34% vs. 7d baseline
·
14:32:18SCAN · 4,182 search terms processed · apex-dtc
·
14:32:22APPROVED · Daniel R. approved pause of 7 ad groups · apex-dtc
·
14:32:26EXEC · Creative rotation: 3 winners scaled +25% · helios-de
·
14:32:31REVERT · Auto-rollback: CPA breached guardrail · crosspoint
·
14:32:34EXEC · 6 MFA placements quarantined · lumen-asia
·
14:32:08EXEC · Budget shift $240/d Display → Search · northwind-retail
·
14:32:11FIX · 14 negative keywords applied to PMax-Apparel
·
14:32:14FLAG · CPC anomaly on brand-eu · +34% vs. 7d baseline
·
14:32:18SCAN · 4,182 search terms processed · apex-dtc
·
14:32:22APPROVED · Daniel R. approved pause of 7 ad groups · apex-dtc
·
14:32:26EXEC · Creative rotation: 3 winners scaled +25% · helios-de
·
14:32:31REVERT · Auto-rollback: CPA breached guardrail · crosspoint
·
14:32:34EXEC · 6 MFA placements quarantined · lumen-asia
As covered in
TechCrunch
Search Engine Land
The Information
Marketing Week
a16z Future
PPC Hero
"Adsmith is one of the first agent products to credibly own a full operational loop inside a third-party platform, not just generate suggestions." — Search Engine Land, Mar 2026
"The audit-log and revert architecture is what makes this enterprise-deployable. Most agent startups skip this part." — The Information, Feb 2026
"If you run more than $50K/month in Google Ads and don't have someone glued to the dashboard, this is the closest thing to a senior PPC manager you can rent." — PPC Hero, Apr 2026
#1 Product of the Day
Product Hunt · Feb 14, 2026
G2
Leader · Spring 2026
PPC Software · 4.8 / 5 (342 reviews)
Capterra Best Value 2026
Marketing Automation · 4.7 / 5
Google Partner · Premier
Verified API developer · ID 7842139
In production at performance teams running real budgets
Northwind
Apex Retail
Crosspoint
Lumen
Vanta Group
Helios

What the agent actually does.

Not "AI suggestions" — concrete operations on your Google Ads account, with full audit trail. Click any capability for the technical detail.

01
Diagnose underperformance at the query level
Pulls daily search_term_report joined with conversion data. Flags ad groups with CTR < 50% of account median, CPA > 1.5x target, or impression share lost to budget. Produces ranked list of fixes with expected impact in dollars.
02
Mine and apply negative keywords automatically
Scans all serving queries every 6 hours. Applies negatives at the right level (campaign vs ad group vs shared list) based on intent classification and brand-safety rules you define.
03
Re-allocate budget between campaigns hourly
Runs a portfolio-level optimizer. Caps movement at ±15% per hour to avoid learning-phase resets. Respects your floors and ceilings per campaign group.
04
Quarantine drifting Performance Max placements
Monitors PMax channel breakdown via Insights API. Detects off-brand placements (low-quality apps, MFA sites) and applies them as account-level negatives, restoring channel mix toward search.
05
Rotate creatives based on statistical significance
Uses Bayesian A/B with proper multiple-comparison correction. Pauses losers only after 95% confidence and minimum sample size — no premature kills during learning.
06
Alert + auto-revert when KPIs break thresholds
Watches CPA, ROAS, spend pace, and conversion volume. If a tracked metric breaches your guardrail, the agent rolls back its last action and sends a Slack/email with full trace.
07
Generate weekly performance briefs in plain English
Auto-drafts an executive summary every Monday: what moved, why, what changes were made, what to expect next week. Shareable PDF or Notion export.
agent_action.json
verified
action_idact_8f4a31b2
type"budget_reallocate"
from"Display-Retargeting"
to"Brand-US-Search"
delta$240.00 / day
forecast_impact+18 conv/day
confidence0.94
approved_byauto-policy
revert_window2h
audit_logview →

The numbers customers actually see.

Aggregated from 1,847 production accounts running for 90+ days. Median, not best-case.

+24%
Median ROAS lift
After 60 days · vs. account baseline.
p10: +8% · p90: +51%
$0.31
Saved per ad dollar
Wasted spend stopped per $1 spent.
Across all account sizes.
11hrs/wk
Time freed per manager
Median across in-house and agency teams.
Single-account: 6h · Multi-acc: 17h
2.4min
Avg. detection time
From anomaly start to agent flag.
vs. 4–18 hours for human teams.
99.2%
Action accuracy
Approved actions that improved or held KPIs.
Measured at 7-day window.
$2.1B
Ad spend under management
Across 1,847 accounts in 41 countries.
SOC 2 Type II audited.

The algorithms that run your account.

No black box. Each subsystem of the agent is documented, version-pinned, and benchmarked against industry baselines.

bidder/thompson-v3
Bid optimization

Thompson Sampling for portfolio bidding

Bayesian multi-armed bandit that balances exploration vs. exploitation across keywords. Outperforms epsilon-greedy and pure-greedy bidding by 12–18% in our backtests on 2.3M ad-group days.

Latency: 84ms p95 Update freq: 15 min Open-source: bidder-core
forecaster/prophet-x
Revenue forecast

Hierarchical time-series forecasting

Decomposes seasonality, trend, and holiday effects per campaign, then reconciles forecasts across the account hierarchy. MAPE of 7.4% on 30-day revenue forecasts (industry baseline: ~22%).

Horizon: 30 days MAPE: 7.4% Retrain: daily
anomaly/spc-cusum
Anomaly detection

CUSUM + EWMA control charts

Statistical Process Control with adaptive thresholds per campaign. Detects mean-shifts in CPC, CR, and conversion volume faster than fixed-threshold alerts. False-positive rate < 0.4%.

MTTD: 2.4 min FPR: 0.38% Window: rolling 7d
creative/bayes-ab
Creative testing

Bayesian A/B with sequential analysis

Beta-Bernoulli priors with Sequential Probability Ratio Test. Stops tests at 95% confidence with built-in multiple-comparison correction. Saves 38% of test sample size vs. fixed-n testing.

Confidence: 95% Min sample: auto Correction: Bonferroni
classifier/intent-bert
Query intent

Fine-tuned BERT for query intent

Classifies search queries into 14 intent classes (commercial, navigational, informational, brand-protect, etc.) for accurate negative-keyword decisions. F1 = 0.91 on internal benchmark.

F1: 0.91 Classes: 14 Inference: on-device
planner/llm-router
Decision planning

Constrained-decoding LLM planner

Generates action plans in a strict JSON schema. Outputs are validated against a typed policy engine before reaching the executor — no hallucinated parameters, no malformed API calls.

Models: Claude · Llama-3 Schema: JSON validated Fallback: rules engine

How the agent makes a decision.

A four-stage loop. Every step is logged, reversible, and gated by policies you control.

1. Observe

Pulls metrics, search terms, asset performance, and account-change history via Google Ads API. Refreshed every 15 minutes.

2. Reason

Combines a structured rules engine with an LLM planner. Generates ranked hypotheses with expected impact and confidence scores.

3. Approve

Routes the plan through your approval policy: auto-apply, single-approver, or quorum. Slack and email handoff supported.

4. Execute & verify

Writes the change via API. Watches downstream KPIs against guardrails. Auto-reverts if a threshold is broken inside the revert window.

// agent_decision_loop
Google Ads API
campaigns, queries, assets, conv
Agent Core
planner + policy engine
Anomaly detector
Cost optimizer
Bid model
Creative ranker
Brand-safety
Forecaster
Executor
writes via API · logs · reverts

Teams running 8- and 9-figure budgets.

Selected from public case studies. Numbers are 90-day post-deployment.

We had four PPC managers running 18 markets. The agent took over the operational layer in week two. We redirected our team to strategy and saw blended ROAS go from 3.1 to 4.8 in 11 weeks. No additional headcount.
+54%
Blended ROAS
$840K
Monthly spend saved
18
Markets unified
MK
Marina Kovač
VP Performance, Northwind Retail · $42M/yr ad spend
The agent caught a competitor bid war within 2 minutes on a Sunday night. It paused our exposure and saved us ~$11K before Monday morning.
2.4m
Mean time to detect
$11K
Saved in 1 incident
DR
Daniel Reeves
Director of Growth · Apex DTC
Audit logs and approval gates were what got it through our security review. The fact that we can revert any action in two clicks made finance comfortable.
SOC 2
Type II passed
14d
Time to deploy
JT
Jasmine Tan
CFO · Crosspoint

Most "AI tools" still leave you doing the work.

Recommendations require human triage. The agent owns the loop end-to-end.

Operation
Native Google Ads + dashboards
Adsmith Agent
Detect anomaly within minutesCPC spike, budget pacing, conv. drop
Manual review or static alerts
Detects in < 3 minutes
Decide on the right fixacross campaigns, accounts, geographies
Human triage required
Ranked plan with $-impact
Execute the change in-platformwrites to Google Ads, not your spreadsheet
Read-only insights
Writes via official API
Approval gates & quorumspolicy-controlled before any change
Not available
Per-action policies
Auto-revert on KPI breachguardrails on CPA, ROAS, spend pace
Email alert only
Automatic rollback
Full audit logwho/what/when, exportable
Change history only, no rationale
Full reasoning + diff
Multi-account portfolio viewcross-account budget reallocation
One account at a time
Portfolio optimizer

Side by side with named competitors.

Compiled from public documentation, vendor pricing pages, and our own product testing (Mar 2026). Updated quarterly.

Capability Optmyzr Adalysis WordStream Adsmith Agent
Autonomous executionwrites to Google Ads without human click Manual Manual Partial Full
Approval gates & quorumpolicy-controlled per action type
Auto-revert on KPI breachrollback within revert window
Tamper-evident audit logMerkle-chain, SIEM exportable Basic Basic Cryptographic
Self-hosted / VPC deployfor regulated industries
Pricing modelpredictable cost vs. spend % of spend Per-account tier % of spend Flat / account
Starting pricemonthly, for 1 account $249/mo $149/mo $264+/mo $0 Solo / $249 Team
Free tierreal free, not trial 14-day trial 14-day trial Free forever

Built to clear your security review.

SOC 2 Type II audited. SSO, RBAC, audit trail, data residency. The agent runs locally on your machine — your account credentials never touch our servers.

SOC 2 Type II
GDPR
ISO 27001
OAuth 2.0
HMAC-signed logs

Local-first by design.

The agent binary runs on your workstation. It authenticates to Google Ads via OAuth and never sends your refresh tokens or account credentials to our infrastructure. Optional cloud sync is opt-in and encrypted with your own KMS key.

Auth
OAuth 2.0 + PKCE
Encryption
AES-256 at rest
SSO
Okta · Azure AD · Google
Data residency
US · EU · APAC

Role-based access control

Granular roles: Viewer, Analyst, Operator, Admin. Restrict who can approve actions, who can change policies, who can view financials. Per-account or per-campaign scope.

Tamper-evident audit log

Every action is signed, timestamped, and hashed in a Merkle chain. Export to your SIEM (Splunk, Datadog) in real time. Suitable for SOX, internal-audit, and external review.

Approval policies

Define which actions auto-apply, which require single approval, and which need quorum. Different rules for spend thresholds, account criticality, or working hours.

On-prem deployment available

For regulated industries: a fully self-hosted version runs in your VPC. No network egress to our infrastructure. Same agent, same audit, fully isolated.

Plugs into the stack you already run.

The agent talks to the systems your team uses. Add or revoke access in one click.

G
Google Ads API
Connected
GA
Google Analytics 4
Connected
M
Google Merchant Center
Connected
BQ
BigQuery
Connected
Sl
Slack
Connected
Sn
Snowflake
Connected
Sg
Segment
Connected
Hs
HubSpot
Connected
Ok
Okta SSO
Connected
Az
Azure AD
Connected
Dd
Datadog
Connected
Sp
Splunk
Connected

Always-on operations.

Live uptime per subsystem over the last 90 days. Public status page at status.adsmith.ai.

All systems operational
status.adsmith.ai →
Agent Core
100.00%
Google Ads API gateway
99.97%
Planner service
99.99%
Audit log storage
100.00%
SSO & identity
100.00%
Slack & email handoff
99.98%
Recent incidents
RESOLVED Delayed sync from Google Ads API
Apr 28, 2026 · duration 18m
Google Ads API returned elevated latency in EU region. Agent automatically degraded to cached state until API recovered. No customer actions were affected.
MAINTENANCE Planner model upgrade
Apr 14, 2026 · duration 4m
Rolling deploy of planner-v3.2.1. Zero downtime, no service interruption.
RESOLVED Slack handoff delay
Mar 22, 2026 · duration 31m
Slack webhook ingestion delayed for some workspaces. Approvals queued and processed once Slack restored. No actions lost.

What we're shipping next.

Updated every two weeks. You can vote on priorities from inside the app under Settings → Roadmap.

Q1 2026 · Shipped
Done
  • Performance Max full supportv2.4.0 · Mar 12
  • Bayesian creative A/Bv2.3.0 · Feb 04
  • Slack approval threadsv2.2.5 · Jan 18
  • Datadog log exportv2.2.0 · Jan 09
Q2 2026 · Now
Building
  • Demand Gen campaign supportbeta · May 2026
  • Multi-currency portfolio optimizerbeta · May 2026
  • macOS native build (Apple Silicon)private beta open
  • Linear & Jira integration for approvals
Q3 2026 · Next
Planned
  • Microsoft Ads support (parity v1)
  • Custom rule scripting (Lua sandbox)
  • HMAC-signed webhook outputs
  • SOC 2 Type II re-audit
Q4 2026 · Later
Exploring
  • Meta Ads autonomous agentresearch phase
  • Linux desktop build
  • FedRAMP Moderate certification
  • On-device LLM planner

Last 6 releases.

A new build every 1–2 weeks. Auto-update on by default, or pin a version in Enterprise.

v2.4.1
Apr 30, 2026
Patch

Performance Max stability + bidder tuning

  • Fixed: PMax-Insights API timeout when account has > 200 asset groups
  • Improved: Thompson Sampling exploration rate reduced from 0.12 to 0.08 — sharper exploitation on stable accounts
  • Fixed: Audit log timestamp drift on systems with manual clock
  • Security: Updated oauth2-client to 4.8.2 (CVE-2026-1147)
v2.4.0
Mar 12, 2026
Major

Performance Max full support

  • Added: PMax channel breakdown via Insights API · placement quarantine
  • Added: Asset-group-level creative ranking with Bayesian A/B
  • Added: Brand-safety negative-placement library (2,400+ MFA sites pre-loaded)
  • Improved: Budget reallocation now respects PMax learning periods
v2.3.5
Feb 24, 2026
Minor

SSO, RBAC scopes, audit improvements

  • Added: SAML 2.0 SSO with Okta, Azure AD, Google Workspace
  • Added: Role-scoped policies — restrict by account, campaign, or action type
  • Added: Merkle-chain audit log with verifiable export to Splunk
  • Improved: Approval Slack threads now show forecasted impact in $
v2.3.0
Feb 04, 2026
Minor

Bayesian creative A/B engine

  • Added: Sequential Bayesian A/B with auto-stopping at 95% confidence
  • Added: Multi-armed bandit creative scaling — winners get +25% impression share automatically
  • Fixed: Edge case where revert window expired during long-running tests
v2.2.5
Jan 18, 2026
Patch

Slack approvals + EU data residency

  • Added: Slack thread approvals with one-click approve/reject and full context
  • Added: EU data residency option (Frankfurt region)
  • Improved: 38% faster cold-start of the desktop client
v2.2.0
Jan 09, 2026
Minor

Observability + Datadog export

  • Added: Real-time log streaming to Datadog, Splunk, and generic webhooks
  • Added: Prometheus-compatible /metrics endpoint on local agent
  • Improved: Anomaly detector now uses CUSUM + EWMA, MTTD down to 2.4 min from 4.8 min
Full changelog (v0.1.0 → v2.4.1)

Built by operators, not advisors.

Founded in 2023 by three engineers who ran performance marketing at scale and got tired of duct-taping dashboards together.

AK
Adrian Köhler
Co-founder & CEO
Previously Head of Paid Media at Zalando · scaled €180M/yr Google Ads spend across 25 markets. Stanford MS CS.
PN
Priya Nakamura
Co-founder & CTO
Previously Staff Engineer at Google Ads (Auction Infrastructure team). MIT BS Math. Contributor to openrtb-tools.
EC
Elena Costa
Co-founder & Head of ML
Previously Senior Research Scientist at DeepMind (causal inference). PhD Statistics, ETH Zürich. Published in NeurIPS & KDD.
Backed by
Sequoia
a16z
Index Ventures
Y Combinator W23
+ 23 angels

Read the source. Read the docs.

Full SDK documentation, public benchmarks, and an open-source CLI for power users.

adsmith · agent_sdk.py
pythontypescriptgo
# Connect to your Adsmith Agent and stream actions from adsmith import Agent, ApprovalPolicy agent = Agent.connect( account_id="123-456-7890", policy=ApprovalPolicy( auto_apply=["negative_keywords", "creative_rotation"], require_approval=["budget_change"], max_daily_budget_delta=500, revert_window_minutes=120, ), ) # Subscribe to live actions async for action in agent.stream(): print(action.type, action.forecast_impact) if action.requires_approval: await action.approve() # or .reject()
90-DAY
MONEY
BACK

No questions asked. No "win-back" calls.

If the agent doesn't measurably improve your ROAS in 90 days, we refund the full Team subscription. Enterprise contracts include a written performance SLA — credits issued automatically if missed.

2.1%
Refund rate
7.4d
Avg. refund time

Free for individuals. Fair for teams.

No per-impression fees, no take-rate of your ad spend. Pay for managed accounts and operator seats.

Solo
For freelancers and single-account operators.
$0 /forever
No card. No expiry.
  • 1 Google Ads account
  • All core agent capabilities
  • Approval gates & auto-revert
  • Local-only audit log (30 days)
  • Community support
Enterprise
For regulated industries and 9-figure budgets.
Custom
Annual contract. Procurement-friendly.
  • Everything in Team, plus:
  • SSO (Okta, Azure AD, SAML)
  • Self-hosted / VPC deployment
  • SOC 2 / ISO 27001 reports
  • Custom data residency & KMS
  • Named CSM · 1h SLA · 99.95% uptime

Talk to humans. Not bots.

A team of PPC engineers — not a tier-1 outsourced helpdesk — answers every ticket.

Email

Median first response: 47 minutes during business hours.

Slack Connect

Shared channel for Team & Enterprise. Real-time triage.

4h SLA · 1h Enterprise
Community

4,200+ marketers and engineers. Self-serve answers in minutes.

discord.gg/adsmith
Phone (Enterprise)

Direct line to your CSM. Available 24/7 with on-call rotation.

+1 (415) 555-0142

Questions performance leads actually ask.

Does the agent move money without my approval?
Only if you explicitly allow it via policy. Out of the box, every monetary action above your threshold requires human approval through Slack or email. You can also require quorum (e.g. two approvers) for high-risk actions like budget shifts above $1,000/day.
What happens if the agent makes a bad decision?
Every action has a revert window (default 2 hours) during which the agent monitors downstream KPIs. If any guardrail is breached (e.g. CPA jumps above target), the action rolls back automatically and a high-priority alert fires. Outside the revert window, any change can still be reverted manually with one click, with the full pre-change state restored.
Does my data leave my machine?
On Solo and Team plans, the agent runs locally and only sends derived metrics (no raw account data, no creative content) to our planner service for the reasoning step — and only if you enable it. On Enterprise, the entire planner can be deployed in your VPC with zero network egress to us.
How is this different from Google's Smart Bidding?
Smart Bidding adjusts bids inside a campaign. The agent works one layer above: it diagnoses problems across campaigns and accounts, decides on structural changes (negatives, budget shifts, ad group restructures, creative rotation), and can override Smart Bidding when it's clearly miscalibrated. Think of Smart Bidding as the auto-pilot, and the agent as the flight engineer who fixes the plane while it's flying.
Can we limit which campaigns the agent can touch?
Yes. Scope is fully configurable: per account, per campaign, per ad group. You can also restrict it by action type — for example, allow keyword and creative changes but disallow budget changes. All scoping is RBAC-controlled.
What system requirements does the app have?
Windows 10 or 11 (64-bit), 4 GB RAM, 200 MB disk, a modern browser for OAuth. macOS and Linux are in private beta — request access from settings after install. No GPU required.
Is this affiliated with Google?
No. We are an independent product that uses Google's official Ads API as a registered third-party developer. We are not endorsed by, sponsored by, or affiliated with Google LLC. We optimize for your KPIs — not Google's revenue.

Hand off the operational layer.
Keep the strategy.

Download the agent, connect your account in two clicks, and watch the first audit log appear within five minutes.

Talk to sales
Free for individuals
SOC 2 audited
OAuth, never stores tokens
Download started — check your Downloads folder