AI for Google Ads management: what it can do – and who decides

Last updated: 27 August 2026

AI in Google Ads management evaluates auctions, checks large volumes of data and handles recurring patterns at a rhythm that a team cannot sustain by hand. How much time that saves compared with how you work today depends on the account. AI cannot set business strategy, and it does not reliably know which conversion is commercially valuable. Reviews at fixed intervals examine a sample; maitiq workflows run on a rhythm of their own and set out the measure, the reason and the supporting data for every proposal. This guide shows the three levels, seven controls and a read-only pilot template.

Which three levels must be kept separate?

The three levels differ not in the degree of automation but in the question of who is accountable for something and who is allowed to change it.

Level 1 – Google AI inside the advertising platform. It receives objectives, budgets, bid targets, signals and assets, and decides the bid and the delivery from them. It is part of the advertising product and is developed further by Google, even without any action on your part.

Level 2 – external analysis and operations workflows. They read account data, run recurring checks, connect signals across account levels and produce proposals with reasons and evidence. They change the account only within an explicitly granted authorisation.

Level 3 – human governance. It defines the business objective, the valid conversion, brand and legal limits, budget accountability, approvals and stop rules. It carries the responsibility, while risks arise at all three levels.

Two things above all flow between the levels: evidence towards governance, and specifications and authorisations towards execution. What matters is that it stays visible which level reads, which decides and which carries out authorised changes.

Level 2 is worth it only if it takes on an operational task that the platform does not solve: bringing signals together across accounts and campaign types, running checks on a schedule, linking recommendations with their source and reason, taking business rules and protected terms into account, making human decisions visible and keeping authorised changes traceable. The comparison software or managed service clarifies whether you need a licence or a managed service for this.

Which AI is already inside Google Ads?

Google AI is already active before any external provider comes into play. Smart Bidding sets bids per auction and covers the Target CPA, Target ROAS, Maximize conversions and Maximize conversion value strategies; Google describes auction signals such as device, location, time of day, search query text, browser or seasonality. Performance Max is an objective-based campaign type across YouTube, Display, Search, Discover, Gmail and Maps; you supply conversion goals, budget, assets, audience signals and optionally a product feed, and steer it through search themes, excluded keywords, brand exclusions and final URL expansion.

For ads in AI-powered search, Google names different levers: bidding strategy and CPA or ROAS targets, budget, choice of conversion action, conversion value rules, custom goals, data exclusions, seasonality adjustments, keyword exclusion lists and the pinning of ad assets. Do not confuse the levels: search themes steer Performance Max, whereas brand exclusions also apply to Search campaigns and to AI Max. For each lever, check which campaign type it relates to.

In addition, with AI Max for Search campaigns Google has introduced a separate feature set made up of improved search-term matching, text customisation and final URL expansion, and automatically switches suitable campaigns over to it. This changeover has two different dates, and anyone who conflates them plans for the wrong deadline:

• From September 2026, the automatic changeover affects campaigns with automatically created assets (ACA) and the campaign-level broad match setting.
• From February 2027, the phase-out of Dynamic Search Ads (DSA) and their automatic changeover to AI Max begins.

The source is Google’s updated announcement “DSA upgrade to AI Max”; Google postponed the DSA date from September 2026 to February 2027 in June 2026. Before planning, check both dates directly in the account and in the announcement: Google moves such windows. A platform-side changeover is a governance event — it changes matching and assets without anyone in your team having made a decision. For each date, record who notices it, who assesses it and which controls are re-established afterwards. How to keep auto-apply, Smart Bidding and external tools in one register, check them and stop them is described in the guide controlling automation with guardrails.

None of these pages promises a result. Automation does not remove the responsibility for measurement and goal selection. A conversion goal groups conversion actions; several primary actions can be valid at the same time, and secondary actions in a custom goal can be used for bidding. What matters is each action's identity, meaning, counting method and time window – and that multi-stage journeys are not counted twice. If an action marked as primary is commercially worthless, a system optimises very efficiently towards the wrong goal.

Which decisions must AI not take on its own?

These questions need a named person responsible: which business objectives take priority? Which conversion is genuinely qualified? How are margin, availability and customer benefit weighted? Which brand or legal limits apply? Which budget change is commercially sustainable? Is an unusual data movement real or a measurement error? Which recommendation fits the current sales or product strategy? Who accepts the remaining risk?

AI can structure evidence on these questions. It cannot take on corporate responsibility for them.

Which seven controls does AI-supported management need?

Missing evidence is not success. A system must distinguish between not observed, observed zero and observed positive; data gaps are boundaries.

How maitiq implements the three levels

How maitiq works: level 2 runs as workflows on a schedule – budget pacing, Target CPA, review of search terms and keywords to exclude with conflict checking, ad and keyword preparation, and reporting. Not every workflow runs daily; the schedule follows the task and the account. Every proposal names the proposed action, the reason and the supporting data; it carries the old and the new value where applicable. Level 3 stays with you, accompanied in the managed model by a Client Success Manager. People set objectives, rules and authority; within an explicitly configured standing authority, eligible proposals are approved and implemented automatically – an explicit, recorded Exclude decision for the named competitor keywords also authorises the corresponding exclusion, without another approval. All other proposals wait for a decision, and every implementation runs through the guarded, logged apply path. By its own account, maitiq offers its management fee at half of what a classic Google Ads agency charges for a comparable scope of services; the media budget is unaffected by this.

Which data crosses which boundary?

“Read-only” describes the fact that no campaign change is carried out. It does not describe which data leaves your account, your country or your legal jurisdiction. These are two different questions, and reputable providers answer both separately.

Ask every provider – including us – in writing:

1. Which fields are transmitted for an AI inference: only aggregated metrics, or also account identifiers, campaign and ad group IDs, keyword texts, website URL, brand terms or free text from your specifications?
2. Does anonymisation or pseudonymisation take place before transmission, and how is it documented and tested?
3. Which inference provider is used, in which country does it process data, and on what contractual basis?
4. Where are the application, database and backups hosted, and who operates the infrastructure?
5. How are OAuth credentials stored and rotated, and how are they invalidated on revocation?
6. Which retention periods apply per data type, and how is deletion carried out?
7. Does the free audit itself already trigger an external inference?

Demand answers that match a configuration you can check. maitiq answers these seven questions in conversation and in the contract documents – account-specific, with a date and a named person responsible. Hold every provider to the same precision: an answer that pulls “read-only” and “data boundary” into one sentence answers only one of the two questions.

How should vendor figures be assessed?

Platform providers publish aggregated improvement figures for their own AI features. Such figures typically come from the provider's internal analysis, relate to a selected segment and are published provider averages, from which no guarantee follows. Check the source, the period, the segment, the comparison basis and whether the figure applies to your campaign structure at all. Never adopt a vendor figure as a planning input or as a budget justification. The same applies to statements that service providers make about themselves.

How should an AI pilot be set up?

A good pilot starts read-only and answers one concrete operating question, for example: “Does the process deliver relevant search query or budget signals earlier, with traceable evidence, without making changes?”

Assess evidence quality, coverage, working time saved and decision value separately from short-term advertising performance. Define the success criteria before the start; a pilot must not “win” through metrics chosen after the fact.

In the maitiq audit, this read-only start is the normal case: the assessment runs on an agreed connected account; native reports are supplied separately and reviewed in a separately arranged pass, and there is no self-service upload. Every check is shown as complete, limited or not carried out, with a reason. A scoped initial assessment shows which paid media tasks maitiq demonstrably supports in Google Ads today and which adjacent channels need a separate implementation review.

What does AI really improve – and what does it not?

AI can check more parts of the account more often, apply patterns systematically and present signals earlier. That improves the basis for decisions. Whether advertising performance improves additionally depends on auctions, tracking, the offer, prices, creative, landing pages, the market and human decisions.

Therefore assess separately what improves: analysis coverage and the decision basis on the one hand, the advertising result on the other. A particular advertising result is not guaranteed – neither by a platform feature nor by a service provider.

For measurement you need a baseline, a documented conversion goal with its actions and a change log before the start. Record each action's identity, meaning, counting method and time window, and do not count multi-stage journeys twice. Do not compare arbitrary periods and do not automatically attribute movements to AI.

Does AI management suit a large Swiss account?

Four preconditions matter for suitability: complexity of the account structure, data quality, internal roles and risk limits. A team with a strong in-house SEA function decides differently from a team without operational capacity. In both cases people remain responsible for the business objectives and for the authority they grant; maitiq supplies the recurring analysis and acts only within that authority.

How do you classify AI features across platforms?

Across platforms the same initial question applies: which feature decides what, on which supporting data and with what degree of control? The classification can be transferred to native AI features of other advertising platforms. For Google Ads, the separation of operation, approval and follow-up control remains; for other platforms the point is classification, not operating procedures.

“AI-powered” does not mean the same thing everywhere in paid media. A platform may use it to describe audience recommendations, creative drafts, delivery, budget suggestions, ad ranking or assistance in the interface. Anyone who compares product labels alone therefore places apparently identical features side by side, even though they see different signals, make different decisions and trigger different actions. A workable comparison begins with six boundaries: objective, signal, decision, action, authority and comparability of the evidence.

An inventory instead of a single AI category

First record every native feature with its official name, the platform, the account or market and the date of the check. “Automatic campaign” or “AI Creative” is too imprecise as an entry. The name must lead to a visible product surface or current official documentation. Features that are not available in an account are not inferred to be active from a global help article.

The feature is then classified. Does it support setup, recommend an audience, generate a draft, prioritise delivery, suggest a budget, make an auction decision or explain a report? Several tasks can sit inside one feature suite. That does not make the feature suite a single decision. Each task gets its own row so that its boundary stays visible.

Six fields make features comparable

The objective describes what the feature works towards according to its configuration. A platform objective is not yet a shared business objective. Two similarly named objectives can use different events, time windows or counting rules.

The signal names the information available to the feature: for example interactions inside the platform, an event supplied by the advertiser, an asset, a URL or an audience specification. “Uses data” is not enough. Origin, definition and platform boundary must be recognisable.

The decision is the concrete choice the system makes. It can suggest content, determine a likely audience, rank ads or choose a distribution. The action is what visibly follows from it: a draft in the interface, a served ad, a placement or a recommendation. Decision and action are recorded separately, because a suggestion is not a delivery and a delivery is not a creative draft.

The authority shows which specifications an advertiser can set, check or exclude, and what the platform determines itself within those specifications. It is not a blanket statement about “human control” but a list of visible boundaries for precisely this feature.

Comparability of the evidence finally answers whether a result is comparable outside the platform. An identically worded metric can contain a different definition, a different window or a different modelled component. Without aligned definitions it remains platform-local evidence.

Four platforms show four different maps

Google groups its current AI Essentials into four areas: AI Data Strength, AI Content Strength, AI Performance Strength and agentic capabilities. This grouping is useful for not hiding different roles inside Google Ads under a single “Google AI” heading. But it is Google’s own product map, not a neutral standard for paid media.

Meta, in turn, describes AI roles in advertising assistance, creative generation and platform-side ad ranking in its current corporate and product communications. Ranking, creative origin and visible labelling are three different boundaries. A label proves neither accuracy nor rights clearance nor effect.

LinkedIn Campaign Manager lists auto-targeting, “Draft with AI”, suggested budgets and Maximum Delivery as separate features. That shows directly why an AI inventory must not merely collect product names: audience recommendation, draft, planning signal and auction logic answer different questions. A product name alone is therefore not a stable comparison key.

TikTok describes the updated Smart+ experience as a spectrum between manual, partial and full automation. The documentation separates, among other things, audiences and placements, budget, creative, campaign structure and reporting. Here too, “Smart+ active” is not a sufficient description. What matters is which sub-feature is used in which form and which setting is visible in the specific account.

These examples are not a product ranking. They show that four providers explain their native AI along different levels and feature suites. The common ground lies not in the name but in the six questions that must be answered anew for every feature.

Product names and availability are time-dependent

Platforms move features, unify interfaces and retire campaign types; TikTok’s updated Smart+ experience is a current example. A table created only once therefore ages silently. Every entry needs a check date, a documentation link and a status such as “visible in the account”, “documented only”, “limited availability” or “retired”.

When a name changes, it is not automatically assumed that objective, signal, decision, action and control options have stayed identical. The old and the new form are initially kept as two versions. Only a documented field-by-field check permits them to be merged. That keeps a product update visible as an evidence boundary.

Do not combine platform reports into a misleading comparison

A shared overview may place platform-local observations side by side but must not calculate with them unchecked. “Conversion”, “lead”, “view”, “engagement” or “AI-generated” can have different triggers and scopes. A platform-side forecast, recommendation or quality indicator is also, first of all, a statement within its own system.

For an honest comparison, every row records whether definition, time window, attribution, currency, market and data status have been aligned. If one condition is missing, the result is not “worse” or “better” but “not portable”. Causal effect and cross-channel attribution belong in a separate measurement question.

Flag four typical wrong comparisons early

The first wrong comparison equates a feature suite with a feature. A campaign type can influence audience, creative and distribution at the same time; a similarly named assistant elsewhere may only create a draft. Both carry an AI label but do not share a decision unit.

The second wrong comparison equates a recommendation with a machine-made choice. A displayed budget signal is an output for review. A platform model that ranks ads during delivery, by contrast, continuously makes a native choice. The visible control points and the evidence required are different.

The third wrong comparison treats manual, partially automated and automated as stable tiers across providers. Without a field list, these words do not say who determines audience, placement, budget, creative or delivery. The comparison overview therefore never adopts a tier label on its own but records the authority per feature.

The fourth wrong comparison turns a transparency indicator into a quality verdict. An AI label can provide context about origin. It confirms neither facts nor rights, brand fit, audience suitability or business impact. Conversely, a missing visible label does not prove that no AI was involved. The platform rule and the specific asset context must be checked separately.

These four errors receive no ranking score. They create an unresolved question and a person responsible for it. That keeps the overview fit for decisions without building an apparently precise platform rating out of unequal product mechanics.

Name native platform AI and external analysis separately

A native feature works within the product and data boundaries of its platform. An external analysis can read several exports or reports but does not automatically possess a shared, fully comparable truth as a result. It may flag differences and pass questions to the responsible people. It must not replace missing definitions with a single uniform AI score.

Authority is platform-specific too. A visible switch, a recommendation and a platform-side delivery decision are not the same action. Detailed operating, stop or change procedures for Google Ads are set out in the Google Ads guides; procedures for other platforms are neither offered nor invented here.

From the comparison overview to the responsible party

If a row shows a Google Ads-specific comprehension problem, the next step leads to the guide controlling automation with guardrails; a tracking diagnosis belongs to the Google Ads conversion tracking audit.

Frequently asked questions

Sources and how to read them

Platform documentation and provider publications explain features and limits; this page relies on those platform and provider sources, not on supervisory authorities or independent price research. Read them accordingly, not as a general market price or proof of success. The linked destinations are in English. Information about how maitiq works is available at maitiq.com.

Have maitiq review your specific case.