Tag: AI Max

Your Google Ads campaign can hit its reported target and still make a decision you would never approve. It can enter a competitor bidding war, learn from queries you already know are irrelevant, favor sales with weak margins, or expand into inventory you did not intend to buy.

You do not need to rebuild every campaign around manual bidding or revive an account full of single-keyword ad groups. You need a control system: clear business objectives, enough consolidated data for automation to learn, explicit boundaries on where it may explore, and a verification loop that catches strategically wrong behavior before it becomes expensive.

Key takeaways

• Automate execution inside boundaries you define. Google Ads can optimize an objective, but it cannot infer every commercial constraint behind that objective.
• Consolidate campaigns when fragmentation deprives Smart Bidding of conversion data. Split them only when the parts genuinely require different economics, budgets, policies, or market strategies.
• As a working benchmark rather than a universal platform rule, look for at least 30 monthly conversions per campaign, with 60 or more providing a stronger foundation for consistent automated bidding.
• Configure negative keywords, brand controls, network settings, and reporting before enabling wider expansion. Do not pay an algorithm to relearn exclusions your business already knows.
• Inspect search terms, match sources, networks, brand exposure, conversion quality, and profitability from the start. A strong top-line ROAS does not prove that the underlying traffic is acceptable.
• After a material change, respect conversion lag. Google has advised waiting one to two conversion cycles before drawing conclusions, unless a hard budget or policy boundary is already being breached.

Define what automation is allowed to decide

Advertiser control no longer means making every auction decision yourself. It means retaining ownership of the decisions that shape those auctions.

Google Ads can observe signals, predict the likelihood of a conversion, adjust bids, and expand targeting. It cannot automatically know that a particular competitor must be avoided, that returned orders erase the apparent profit from a product category, or that your sales team cannot handle another wave of low-value leads. Those are business facts, not auction facts.

The distinction matters because the platform’s definition of success is not automatically the advertiser’s definition. Google benefits when advertisers spend money. You benefit when additional spend produces acceptable incremental business outcomes. Those interests can overlap without being identical.

Write an automation contract for each campaign

Before changing a bidding strategy or enabling AI-driven expansion, write down the following decisions in plain language:

1. Business objective: Name the result the campaign is supposed to produce. Do not substitute ad position, traffic volume, or spend for a business result.
2. Economic target: Record the CPA, ROAS, margin, or other threshold the business actually uses. If different products have different economics, state how those differences will be represented.
3. Permitted expansion: Specify whether the system may explore broad queries, competitor searches, Search Partners, new geographic areas, or additional channels.
4. Prohibited behavior: List the queries, brands, locations, offers, audiences, and traffic sources that are unacceptable even if their reported conversion performance appears strong.
5. Conversion definition: Identify which recorded actions represent real value. Separate primary outcomes from actions that are useful for observation but should not steer bidding.
6. Evidence required: Name the reports you will inspect to verify search terms, match sources, networks, conversion quality, and economic performance.
7. Intervention rule: Define the conditions that require a pause, exclusion, target adjustment, or deeper review. Use thresholds approved by your business rather than inventing them after spend accelerates.

This contract prevents a common mistake: evaluating automation only by the metric it was instructed to optimize. If a campaign reaches target ROAS by entering strategically unwanted auctions, the bidding system may have completed its assignment perfectly. The assignment was incomplete.

Treat every platform recommendation as a hypothesis about execution. Ask which part of your contract it supports, which new permissions it requires, and where its effect will be visible. If you cannot answer those questions, investigate before applying it.

Consolidate learning without flattening business differences

The old response to uncertainty was often more structure: single-keyword ad groups, duplicated match types, traffic-sculpting negatives, and numerous narrowly defined campaigns. Much of that tactical granularity has become unnecessary under automated bidding and matching.

Excessive structure now creates a different risk. Every additional campaign divides the available conversion history. Automated bidding then has fewer observations from which to estimate performance, while each segment receives a smaller share of the account’s traffic and budget.

A useful working benchmark – not a guarantee and not a reason to ignore your own variance – is at least 30 conversions per campaign each month, ideally 60 or more, for Smart Bidding to operate consistently. Before creating a split, estimate how much recent conversion volume each resulting campaign would retain. If one side would fall well below that range, the business reason for separating it needs to outweigh the loss of learning density.

Use a business test for every proposed split

Create a separate campaign when at least one of these conditions is true:

• The segment needs a genuinely different CPA, ROAS, or profit target.
• Its budget must be protected or capped independently for a clear commercial reason.
• Its geography, availability, compliance requirements, or operating capacity differs from the rest of the account.
• Its brand, competitor, query, network, or channel policy must be different.
• The business intends to make a distinct investment decision about that segment and cannot obtain the necessary control through reporting, labels, or exclusions.

Do not create a campaign merely because a reporting dimension exists. Reporting taxonomy and bidding structure are different tools. You can often preserve a consolidated learning pool while using labels and reports to analyze meaningful groups.

Margin is a good example. An account divided into many narrow margin buckets, each carrying its own ROAS target, can look financially rigorous while fragmenting the data the bidding system needs. The resulting campaigns may be too small to achieve the targets that justified the structure.

Instead, use custom labels for information such as margin, sell-through rate, and return rate. Labels do not magically convert profit into a bidding signal, but they let you organize products, inspect performance, and make campaign decisions with context Google does not inherently possess. If you later separate a segment, you can do so because the data reveals a material economic difference, not because a spreadsheet had another row available.

Put guardrails in place before expansion starts

Automation should discover what you do not know. It should not spend your budget rediscovering what you already know.

This is especially important when broad matching or AI-driven expansion can reach searches outside your initial keyword set. Broad match defaults have long created a situation in which inexperienced advertisers can pay to teach the system lessons their businesses could have supplied in advance. If a query category is known to be irrelevant, exclude it before launch rather than waiting for wasted clicks to prove the point.

Configure the controls that correspond to the risk

1. Query risk: Add negative keywords for known irrelevant intent. Review whether exclusions need to apply at the campaign or account level based on how broadly the rule should operate.
2. Brand risk: Decide how your own brand, excluded brands, and competitor brands should be handled. AI Max provides brand inclusions and exclusions, but the advertiser still has to define the policy.
3. Network risk: Decide whether Search Partner Network traffic is permitted. Set the available network control deliberately, then evaluate actual network performance rather than relying on a general assumption about where AI Max will expand.
4. Economic risk: Make margin, returns, sell-through, and other meaningful product differences visible through your feed organization, labels, conversion values, reporting, or campaign design.
5. Measurement risk: Confirm that the conversions guiding bidding represent outcomes the business values. A campaign cannot optimize toward profit if the recorded objective rewards a weak proxy for it.
6. Visibility risk: Make sure the team knows where to inspect search terms, AI Max match type, match source, network delivery, and brand exposure before more traffic arrives.

Competitor traffic shows why these controls cannot be reduced to a performance metric. In one documented rollout, AI Max expanded traffic by targeting a much larger competitor. The reported performance looked good, but the advertiser had intentionally avoided those searches to prevent a bidding war. The system found an opportunity inside the data while violating a strategy that had never been encoded.

That is not an argument against AI Max. It is an argument for declaring competitor policy before enabling it and checking search terms from the first review. Strong aggregate results should increase your curiosity about where the gains came from, not end the investigation.

Be equally careful with conclusions drawn from a small number of campaigns. Early AI Max observations appeared to show a preference for Search Partner traffic, but additional data did not support that as a general rule. Use account-level findings to form a testable question. Do not turn them into a platform-wide belief until the evidence warrants it.

Verify patiently, then keep strategy human

Good oversight separates two jobs that are often confused. Verification asks whether the system is doing what you authorized. Evaluation asks whether the result is good enough to continue. Verification starts immediately; evaluation may need to wait for conversions to mature.

Inspect behavior in the right order

Use this sequence when reviewing an automated campaign:

1. Delivery: Check where spend occurred, including networks, locations, channels, and any other enabled expansion surface.
2. Matching: Inspect search terms, match type, and match source. Identify which traffic came from your explicit targeting and which came from automation.
3. Strategic fit: Look for prohibited brands, competitor auctions, irrelevant intent, or traffic that conflicts with your operating policy.
4. Conversion quality: Determine whether the reported conversions represent qualified leads, completed sales, or another outcome the business can actually use.
5. Economics: Review CPA or ROAS alongside the margin, returns, sell-through, capacity, and customer value information relevant to the decision.

This order keeps a blended efficiency metric from concealing an unacceptable mechanism. If the campaign reaches its ROAS target through traffic your business has explicitly rejected, you have enough evidence to tighten the boundary even before the long-term average settles.

Allow for conversion lag without tolerating a breach

After a platform update or material campaign change, immediate performance claims are unreliable when conversions take time to arrive. Google has advised advertisers in this context to wait one to two conversion cycles before assessing the effect.

That waiting period is not permission to ignore the account. Continue checking spend, query relevance, network delivery, and other hard boundaries. If automation exceeds an approved budget limit or enters prohibited traffic, intervene. If the guardrails hold but the efficiency metric fluctuates, let the relevant conversion window mature before declaring success or failure.

Avoid defensive target changes made only because other advertisers appear worried. Advertisers have raised target ROAS even when their campaigns were not budget-limited, a reaction that can reduce participation without solving an identified problem. A stricter target may be appropriate, but changing it can materially reduce volume. Require an account-specific reason and preserve a baseline against which the effect can be judged.

Keep a decision log, not just a change history

For every material adjustment, record:

• The date and exact setting changed.
• The business problem the change is meant to solve.
• The expected effect on traffic, conversions, CPA, ROAS, or profit.
• The relevant conversion lag or evaluation window.
• The reports that will confirm where the effect came from.
• The hard boundary that would justify intervening early.
• The final decision after enough data has accumulated.

When possible, avoid stacking several material changes into the same evaluation window. If you alter the target, budget, network access, negatives, and campaign structure together, even a clear performance movement may not tell you which decision caused it.

Apply the same skepticism to controls whose labels sound clearer than their mechanics. An emerging Performance Max control for channel importance may give Google greater tolerance around a CPA or ROAS target when a channel receives more importance. Do not assume that increasing importance simply buys more of a channel at unchanged economics. Document the intended outcome, monitor actual allocation and efficiency, and reverse the change if the observed tradeoff is unacceptable.

Finally, do not confuse prominence with profit. Paying whatever it takes to hold the top ad position was an expensive mistake in earlier paid search, and position-driven bidding can sacrifice economics for prestige. Automation does not change that principle. Your objective should describe the business outcome you want, not the visible status you hope to occupy.

Start with one automated campaign this week. Write its automation contract, remove any split that lacks a business reason, encode known exclusions, capture a baseline, and schedule the evaluation for the end of its relevant conversion window. You will have given the system room to find demand without giving it authority to redefine what your business considers a good customer, an acceptable auction, or a profitable result.

References

• Search Engine Land – Matt Van Wagner shares what’s changed, what disappeared and what PPC still gets wrong
• Search Engine Land – Google Ads AI needs guardrails, not blind trust ft Mike Ryan