Evidence-based AI ad optimization
Connect your Meta and Google Ads accounts. Mervin compares each proposed action against real outcomes from similar campaigns, then shows the sample size, median effect, confidence interval, and its own historical hit rate before you approve anything.
Similar actions were followed by a positive ROAS change in 67% of observations. Historical association is not a guarantee.
The recommendation, fully unpacked
See the action, cohort definition, observed outcomes, uncertainty, limitations, and the exact write operation before making a decision.
RECOMMENDATION · META ADS
TW Commerce / Meta Ads · Q4 Prospecting / Broad Value
ROAS has remained above the account target for 7 days while impression share is constrained by budget.
EVIDENCE FROM SIMILAR ACTIONS
In similar historical cases, this type of budget increase was followed by a positive ROAS change in 67% of observations. The result is statistically meaningful, but it does not guarantee the same outcome for this campaign.
Historical association does not establish causation. Auction pressure, creative fatigue, or tracking changes may produce a different result.
Nothing changes until you approve.
The black-box problem
“Increase the budget by 20%.”
“In 186 comparable cases, similar budget changes were followed by a median 8.4% ROAS increase, with a 67% positive outcome rate.”
Not “the AI thinks.” The data shows.
Evidence Engine
Every condition remains inspectable. Adjust the platform or industry to see when the historical signal is meaningful, and when Mervin should say there is not enough evidence.
Active cohort: Meta Ads, E-commerce, Conversions, Taiwan, Q4, budget increase of 10–20%, evaluated over 14 days.
COHORT RESULT · ILLUSTRATIVE DEMO DATA
For this cohort, budget increases were historically followed by a median 8.4% ROAS change over 14 days. 67% of observed outcomes were positive. Historical evidence is not a guaranteed forecast.
How Mervin learns
Mervin turns actions and subsequent observations into a decision record that can be reviewed, challenged, and scored.
Meta and Google Ads accounts
Account and performance history
Historical advertising actions
Performance after each action
Relevant comparable cases
An evidence-backed action
The result and its own hit rate
The statistical engine is designed to learn from dozens of ad accounts, hundreds of thousands of daily performance records, and more than ten thousand historical actions, with daily updates. These are scale ranges, not exact performance claims.
A day with Mervin
Mervin reads continuously, but the human stays at the decision point.
Yesterday's Spend, ROAS, CPA, conversions, and unusual changes are summarized across accounts.
Mervin flags a budget, creative, keyword, audience, or performance issue that needs judgment.
Comparable actions are matched and evaluated for sample size, median effect, positive rate, confidence, and significance.
The exact proposal is sent through the official API, logged, and monitored only after a person approves it.
Mervin OS
A work-focused command center built around the question that matters: why is this action worth considering?
Spend is budget-constrained while 7-day ROAS remains above the account target. The matching cohort excludes learning-phase campaigns.
Transparent accountability
A recommendation is not counted as correct or incorrect until its observation window closes. Pending outcomes never inflate the score.
OVERALL HIT RATE
71.2%Correct outcomes among recommendations whose 14-day evaluation window has closed.
What counts as a hit? The recommendation's stated outcome must beat its pre-defined baseline after the full evaluation period. Recommendations still inside that period are pending and excluded from the hit rate.
Approval-first execution
Inspect the exact operation, adjust it if needed, then approve. Try the demo below to move a proposal from draft to execution monitoring.
Core capabilities
Discuss an account like you would with a senior optimizer: inspect numbers, test reasoning, and adjust the direction.
Every proposal includes sample size, median effect, confidence interval, and statistical significance.
Review how past recommendations performed after their observation windows closed.
Every write action needs human approval. Newly created ads start paused.
Analyze and manage both advertising platforms from one consistent decision workspace.
Who it is for
Review more accounts without giving up professional judgment. See where attention is needed and why.
Explore MervinTurn senior optimization experience into a reviewable decision record that teams can discuss, approve, and hand over.
Explore MervinMove beyond 'the AI recommends.' Inspect the evidence and decide whether the action is worth taking.
Explore MervinA clearer decision system
Mervin combines human judgment with a visible statistical record and a controlled execution path.
Transparent by default
No “proprietary AI” shortcut. Just clear definitions, limits, and operating rules.
Ask about your accountsMervin does not stop at a recommendation. It shows the comparable cohort, sample size, historical outcome distribution, confidence interval, limitations, and Mervin's track record for that recommendation type before any action is approved.
Similarity is defined with visible filters such as platform, industry, objective, market, season, action type, change range, tracking quality, and campaign state. The cohort definition appears on every recommendation so a reviewer can challenge it.
The engine learns from advertising actions and subsequent performance observations available to the product. Results are presented as aggregated cohorts. This page uses illustrative demo data, not customer performance claims.
No. Significance means the observed pattern passed a defined statistical threshold in the historical cohort. It does not prove causation or guarantee that a future campaign will behave the same way.
A recommendation is evaluated only after its observation window closes. Its outcome is compared with the stated baseline and success definition. Pending recommendations are excluded from the rate, and the sample size is always shown.
Mervin labels the proposal 'Insufficient evidence,' explains which cohort is too small, and avoids presenting statistical significance. The user can still inspect the account context and decide manually.
Recommendation evidence is designed to show aggregated cohort results, not another advertiser's campaign names, creatives, audiences, or raw account records. The exact production safeguards should be confirmed during security review.
No. Approval mode is always on. Budget changes, pauses, audience edits, and newly created ads all remain proposals until a person approves the exact operation.
A paused default prevents a new ad from spending immediately after creation. A reviewer can inspect the creative, targeting, budget, and tracking setup before activation.
The interface is designed for budget, status, audience, keyword, bid, and new-ad proposals. The exact supported write operations depend on the connected platform and the current product release.
The engine is designed to refresh daily as new performance observations complete their evaluation windows. Each evidence panel shows its own data freshness timestamp.
Yes. Mervin OS is designed around an account selector, shared approval queue, action history, and evidence records so agency teams can review decisions across accounts without losing context.
Every recommendation comes with evidence.
Connect an account, review your first evidence-backed recommendation, and decide whether it is worth acting on.
