About 2.3%. That is how little reported weekly manual optimization time fell while third-party AI bidding adoption climbed eight percentage points.
The arithmetic is straightforward: reported manual work fell from 10 hours 12 minutes to 9 hours 58 minutes a week. Adoption moved. The workload barely did.
I would not read that as proof that AI bidding failed. The figures do not establish causation, isolate Google Ads, or measure MCP agents. I read them as a reason to reject a much easier sales pitch: that an automation label is evidence of labor saved. A tool earns that claim through completed, verified work and the review time it leaves behind.
The arithmetic: higher adoption, barely lower reported hours
Start with what was actually measured. The PPC Land writeup of DoubleVerify’s survey figures reports the following changes:
| Measure | 2025 | 2026 | Change, calculated from the reported figures |
|---|---|---|---|
| Manual optimization per week | 10h 12m | 9h 58m | −14 minutes, about −2.3% |
| Third-party AI bidding adoption | 42% | 50% | +8 percentage points, about +19% relative |
The time calculation uses the earlier weekly total as its baseline: 14 minutes divided by 612 minutes. The adoption calculation uses its own starting share. Those are different measures, not competing bars on the same scale.
That distinction matters. The figures show a small reported reduction in manual hours alongside a larger relative increase in adoption. They do not show what happened to the workload of an individual marketer who adopted AI bidding.
The mismatch is a question worth investigating, not a causal finding.
What the survey cannot tell us about Google Ads MCP
The figures come from 2,020 marketing decision-makers across 25 countries, surveyed in February 2026 by Sapio Research. PPC Land’s analysis appeared in November. I treat that as a February snapshot, not a live reading of accounts at publication.
The sponsor also matters. DoubleVerify owns DV Scibids AI bidding, which its own writeup positions as a solution to the manual work it reports. That does not make the findings false. It makes the distinction between the findings and the sales argument worth keeping visible.
The same analysis flags a 2.2-percentage-point margin of error at 95% confidence, no breakdown by title, company size, or budget, and missing respondent bases on most charts. Those limits leave several questions unanswered:
- Did AI bidding cause the reported reduction in manual work?
- Is the difference in reported hours statistically meaningful?
- Did Google Ads operators experience the same change?
- Did any particular agent reduce the work required to manage an account?
The supplied figures answer none of those questions. The sample covers marketing decision-makers broadly, and the available cuts do not isolate search.
Nor is this an MCP performance study. Google’s official MCP server launched on October 7, 2025, as an open-source, read-only tool for reporting and diagnostics, with no execution at launch.
Use the survey to sharpen the questions asked of an agent, not to score its performance.
Automated bids do not empty the work queue
Smart Bidding handles a particular part of account management: adjusting bids at auction time for the query, device, and audience in front of it. That is not the same job as moving budgets between campaigns, excluding irrelevant queries, adding or pausing keywords, and fixing creative.
That remainder still needs monitoring and intervention. It helps explain why automating bids is not equivalent to automating account management. It does not, by itself, explain the survey’s reported hours. I keep those claims separate.
The broader adoption figures also deserve a closer read:
| Reported measure | Figure | What it does—and does not—show |
|---|---|---|
| Planned AI use in media planning, buying, and optimization | 97% | Intent, not completed implementation |
| AI use to activate campaigns | 35% in 2025, versus 31% in 2024 | Reported use in a specific workflow |
| AI use to optimize bidding | 29% in 2025, versus 24% in 2024 | Another workflow measure, not interchangeable with third-party bidding adoption |
The report also claims 98% of AI-bidding users saw performance gains. But the question asked, “How did AI bidding solutions improve campaign performance,” which assumes improvement. Reported benefits included time savings at 44%, ROAS at 37%, and CPA at 34%.
That wording limits what the headline can establish. A question that assumes improvement is not a neutral test of whether improvement occurred. Ask what work disappeared, not just whether respondents selected a benefit.
Ask ChatGPT for an accountable work list
If bidding is only part of the timesheet, I want an AI tool to account for the rest.
With groas, the starting point is to connect the account and add groas to ChatGPT. My request is deliberately scoped: identify what needs attention this week, show the account evidence behind each item, and classify it as a human decision, an execution step, or no action.
That is a more useful test of Google Ads AI tools than asking for general optimization advice. A plausible recommendation is easy to write. An account-specific task with checkable evidence is harder to bluff.
The work list should separate:
- Search terms that spent without converting.
- Budgets and bids that hit limits.
- Keywords and account structure that need additions or pauses.
- Creative or landing pages that may be dragging conversion rate.
A flag is not a diagnosis. If the available data does not establish why a landing page underperforms, I want the answer marked unresolved rather than dressed up as certainty.
Execution capability matters next. The cited overview of Google’s official server describes version 0.0.6 as having seven tools: three reads and four guarded writes for assets and labels. It lists no tools for bids, budgets, keywords, or campaign pauses. On that description, the server cannot clear the broader work list.
With groas, every change is logged with its reason and can be undone in one step. That gives me a record to inspect, not permission to stop inspecting it. No evidence, no decision. No verified execution, no completed task.
Follow a search-term task through to completion
For an illustrative account spending $20,000 a month, I would start with a narrow search-term review. The spend is context, not a measured result or a threshold for using an agent.
The prompt is boring on purpose:
Show search terms from the last month with spend, conversions, and cost per conversion. Flag terms that spent without a qualified action. Explain each candidate using the business context, then propose exact negatives. Do not apply anything.
The model reads the account through groas. The output is a proposal. Nothing changes yet. That pause is an approval checkpoint, not a control group and not proof that the proposal is correct.
I would then work through the task in order:
- Check intent. Is the term irrelevant, a close variant worth keeping, or a tracking gap where the conversion never fired?
- Approve a precise scope. Name the negatives and the campaign. Leave no room for adjacent “helpful” changes.
- Execute only that scope. Apply the approved list and nothing else.
- Read the account back. Confirm that the campaign negatives match the approval.
- Check the record. Reconcile the result with the reasoned change log and undo path.
That is the useful distinction from the official server’s documented limits: a reporting bridge can surface the work, but it cannot perform the bid, campaign-pause, or keyword changes described here.
An approved change that is verified in the account beats a recommendation left for someone else to implement. The first-time autopilot runbook carries the same setup discipline: decide which decisions stay human, then hold that boundary.
Make the tool survive the labor equation
Here is the math I would use to evaluate AI agents for Google Ads:
Net human time saved = baseline human task time − prompting, review, correction, and exception-handling time.
The baseline is how long the task takes by hand, with a timer running and no agent involved. From that, subtract every minute of human work the agent adds back: writing the prompt, checking proposed negatives against actual intent, fixing errors, and reverting changes that should never have passed approval.
Waiting for a read belongs in the record too, with a clear distinction between time spent waiting and time that can be used elsewhere. Otherwise, a tidy calculation can hide an untidy working day.
Skip the review and repair costs and the result measures the demo, not the week.
I would keep a plain task log during a short trial:
| Record | Why it belongs in the test |
|---|---|
| Task and baseline manual minutes | Establishes the work being replaced |
| Prompting, review, correction, and exception minutes | Captures the human work that remains |
| Proposed, approved, and verified actions | Separates suggestions from completed work |
| Abandoned or reverted actions, with reasons | Keeps failed work visible |
A rejected proposal still costs review time. A negative that needs undoing costs review time plus repair time. Neither belongs in a vendor’s completed-work total.
If a vendor claims weekly savings, I want to know which baseline it timed and which review minutes it counted. The survey cannot supply that evidence on the vendor’s behalf: it reports weekly hours across marketing decision-makers, not timed Google Ads tasks before and after a specific agent.
Count the work removed after supervision, not the work the demo appears to do.
More edits do not mean less work
Before calling a trial successful, I would ask ChatGPT to reconcile approved actions with the groas change log:
List what I approved, list what actually changed, and flag anything that changed without approval.
An agent can do plenty without doing what was asked. Edit volume is neither labor saved nor performance gained. Adding keywords is not automatically useful if the review creates more work than the implementation removes.
I would keep budget limits and approval boundaries in writing, leave autopilot off while measuring review time, and enable it only for a narrow scope once verified completion is consistent. The distinction between tasks that can run unattended and those that still break is covered in what genuinely runs unattended. That boundary belongs in the operating rules, not in the sales deck.
CPA or qualified pipeline needs a separate record. A tool can reduce review time and still lose money if the negatives are wrong. A clean log proves traceability; it does not, by itself, prove better advertising.
Verify execution, measure labor, and track business performance separately. Let none of them stand in for the others.
The buying verdict: require both ledgers
I would run a Google Ads AI trial on two ledgers and accept no substitute for either:
- Human work remaining: baseline task minutes; prompting, review, correction, and exception time; abandoned or reverted actions.
- Business outcomes: spend, qualified conversions, CPA or ROAS over the same period, and whether tracking stays intact.
The outcome ledger needs the same restraint as the survey analysis. Results during a trial are not automatically results caused by the tool. But without that record, there is no basis for deciding whether less manual work comes at an acceptable business cost.
Neither a survey average nor a lower subscription price proves that an account improves. Subscription cost belongs in the decision, but so does the human time required to make the tool useful.
This test will not fit everyone. For a small search campaign checked monthly, timing prompts and reconciling logs may cost more effort than the tool removes. I would stay manual rather than build an administrative department around a tiny account.
For accounts with recurring work across search terms, budgets, and creative, I would require both ledgers before renewing. The survey’s 2.3% decline in reported manual hours is not a verdict on MCP. It is a reason to demand better evidence than adoption figures and automation labels.
The next tool earns its place when verified work gets done, review minutes fall, and CPA or qualified pipeline holds. That is the buying decision the numbers support.

