AI-Powered Advertising Enters Execution: Brands Need a Common Measure of Growth

01 Facts & context

Platform arrangements and measurement discussions are two different types of information

In an official announcement on 8 month 12 day, Google arranged to gradually migrate eligible campaign-level broad match and standalone auto-created assets to AI Max starting from 9 month 1 day. Dynamic search ads are not part of this round, and their automatic migration timing has been adjusted to 2027 year 2 month. On 9 month 2 day, another official Google content focused on how to combine real-time attribution, incremental testing, and marketing mix modeling to measure longer purchase journeys. These publications respectively explain platform arrangements and measurement approaches, and one cannot infer that a particular account has completed migration or directly deduce that advertising performance has improved.

02 GSP analysis

The modeling perspective: calibration is not a permanent budget rule

Google's Meridian documentation explains that experimental results can inform priors, while experiments and marketing mix models may differ in the outcome being estimated, geography and time window. Connecting an incrementality experiment to a mix model does not automatically make their outputs comparable. This is methodological guidance from the tool provider, not evidence that a particular advertising account already has a reliable return estimate.

Before using an experiment in a broader budget discussion, GSP proposes attaching a statement of applicability: which products and audiences were tested, at what spending level, over what purchase cycle, and for which channels? An experiment on new customers in selected cities cannot simply become an annual, nationwide rule for existing customers. The additional assumptions should be recorded and the range of plausible conclusions widened rather than retaining one apparently precise number. Teams can then examine how sensitive the budget recommendation is to those assumptions. If a small change in the prior reverses the channel ranking, the available data do not yet support a stable allocation decision. That uncertainty is useful managerial information. It identifies the next data or testing need, rather than requiring analysts to make every model return the same answer. A calibration record also allows future reviewers to distinguish an updated business condition from a genuine change in advertising effectiveness.

03 GSP analysis

The field-research perspective: randomization may still leave an imprecise answer

Lewis and Rao's 2015 study examined 25 large advertising field experiments and found that return estimates could remain highly imprecise even under randomization. The experiments concerned US retail and brokerage settings of that period; they do not predict AI Max's effects in 2026. They offer a perspective distinct from platform documentation: a design that reduces particular biases does not guarantee that the available sample and observation period can answer the business question.

GSP's interpretation is to define, before testing, the difference that would actually change the decision. Expected sample size, outcome variability and the observation period can then be assessed against that need. A small pilot may establish that data are returned correctly, leads can be followed and sales staff can apply a common scoring rule. It may still be unable to establish advertising returns precisely. If the resulting interval includes both outcomes that justify expansion and outcomes that do not, the conclusion is that the alternatives remain unresolved. It is neither a reason to publicize the positive point estimate as success nor proof that advertising has no effect. The next step may be longer observation, more suitable grouping or clearer definitions of a noisy business outcome. This gives operational piloting, effectiveness estimation and scale decisions distinct jobs instead of asking a limited test to carry the entire growth decision.

04 GSP analysis

Before automation, clarify what constitutes valuable customer progress

Before automated advertising expands matching scope, brands must answer a more fundamental question: what counts as valuable customer progress? If only form submissions are treated as success, the system will continuously optimize for such behaviors, but these may include misclicks, price comparisons, or non-target audiences. GSP recommends breaking down customer progress into verifiable stages: a qualified inquiry must meet clear criteria, such as complete contact information and a business need that matches the brand's capabilities; after matching the need, it enters sales communication, and the final feedback is the partnership outcome. Each stage should define business event fields and set exclusion conditions, such as duplicate submissions or clearly irrelevant requests, to prevent invalid data from being treated as growth signals.

This definition is not a one-time document but a prerequisite for automation operation. Suppose a B2B service provider sets its advertising objective as form submissions but does not distinguish inquiry quality. As a result, the sales team receives a large number of low-intent leads, yet the system increases spend because submission volume rises. If the provider had predefined that a qualified inquiry must include a corporate email and budget range, and exclude competitors or student surveys, the conversion signals fed back would be closer to real opportunities. GSP believes that without such conditions, automation will only amplify existing measurement biases rather than improve growth quality.

When implementing, the team should review existing conversion events, annotate their business meaning and exclusion logic one by one. For example, the form page can add hidden fields or validation rules to distinguish manual submissions from auto-fills; the CRM should record lead source, first contact time, and subsequent status, and regularly compare advertising data with sales feedback. The key is that all definitions must be confirmed jointly by business and sales, not set solely by advertising personnel. If sales considers a certain type of inquiry non-valuable, it should be removed from the optimization objective or placed in a separate observation group.

Refutable signals include: form submission volume rises but the qualified inquiry rate declines, or after sales follow-up, the partnership conversion rate does not improve correspondingly. In such cases, the team should check whether the event definition is too broad or whether exclusion conditions are effective. GSP recommends reviewing stage conversion rates weekly and recording definition changes to trace performance fluctuations. If the proportion of qualified inquiries improves after adjustments, it indicates the definition is working; if there is still no change, the team should examine landing page information or target audience matching. In summary, clarify each step of customer progress first so that automation can optimize in the right direction.

05 GSP analysis

Data feedback needs to express both value and uncertainty

Data feedback is the key bridge between automated advertising and real business, but feedback itself must express both value and uncertainty. If only form submissions are used as conversion signals, the system will continuously optimize to obtain more submissions, but it may ignore the quality differences behind them. For example, the same form may come from a clear need, a price-comparison observer, or a misclick. If such data is fed back without distinction, the model will mistakenly treat invalid behaviors as positive signals, forming a reinforcement loop.

To establish an effective feedback mechanism, the team should record source, business stage, update time, missing information, and delays, and clarify the responsible person. Specifically, define stage fields from form submission to qualified inquiry, sales follow-up, and partnership closure, and set exclusion conditions (e.g., duplicate submissions, invalid contact information). Suppose a B2B service provider treats 'downloading a whitepaper' as a high-value conversion, but in reality, most downloaders are not target customers. If quality validation is not set, advertising will skew toward low-quality traffic.

Feedback should be based on clear conditions, not blindly transmitting all data. For example, only feed back when the user completes verification or meets preset rules, and annotate data confidence. At the same time, avoid claiming unauthorized personal data and respect privacy boundaries. In practice, first map the existing data flow, identify which fields reflect business stages and which have delays or gaps, then design feedback rules.

Refutable signals include: after feedback, conversion rates rise but sales lead quality declines, or the model frequently triggers duplicate submissions. In such cases, check whether the feedback logic is too loose. It is recommended to establish a monitoring mechanism to regularly compare feedback data with sales feedback. If discrepancies are found, adjust field definitions or exclusion conditions in a timely manner. In this way, data feedback can both drive automation and maintain business authenticity.

06 GSP analysis

Attribution answers path questions and cannot independently prove incremental results

Attribution models answer the question of how to allocate credit among recorded touchpoints in the conversion path. They describe how credit is divided among touchpoints, not whether those touchpoints truly generated incremental business results. In the context of AI-driven advertising automation, attribution data is often used as an optimization signal, but if path allocation is equated with incremental contribution, the role of certain ads may be overestimated, causing budget decisions to deviate from reality.

Take a company providing long-cycle consulting services as an example. Suppose its customers typically take several weeks of comparison before deciding to cooperate. Attribution reports may show that the last-click ad contributed to the conversion, but customers may also have signed up due to organic search, industry reputation, or sales follow-up. If the team adjusts advertising solely based on attribution results, it may ignore these alternative explanations and mistakenly attribute organic demand or brand accumulation to ad effects.

Therefore, attribution should be seen as a starting point for understanding the path, not as the endpoint for proving incremental results. To distinguish between the two, the team needs to combine incremental testing, such as pausing or increasing ad spend in specific markets and observing changes in business outcomes. At the same time, recording customer feedback during sales conversations, such as whether they mention that ads influenced their decision, can provide supplementary evidence to help verify the reliability of attribution conclusions.

A practical approach is to establish an independent business outcome tracking table outside the attribution platform, recording the source of each lead, sales follow-up status, and final partnership situation. When attribution data and business outcomes diverge, prioritize checking whether conversion definitions are consistent, whether the observation period covers the decision-making process, and whether there are untracked touchpoints. Refutable signals include: attribution conversion rates rise but qualified inquiry volume does not increase correspondingly, or incremental tests show that business volume does not drop significantly during ad pauses. These signals suggest that attribution may overestimate the true contribution of ads, requiring adjustment of the measurement definition.

07 GSP analysis

Incremental validation starts with a comparable question

Incremental validation first requires writing the problem as a comparable business judgment: under conditions as close as possible, whether a clear change in investment brings additional suitable customer progress. The team needs to explain what they plan to change, which stage they expect to impact, how results will be recorded, and what factors may interfere with the comparison. If simultaneously adjusting creatives, pages, sales scripts, and customer screening criteria, even if conversion changes, it is difficult to explain which action caused the effect. It is recommended to narrow the problem scope first, then choose a comparison method that can be implemented, rather than inferring conclusions from a preset growth magnitude.

When conditions permit, consider a controlled arrangement with clear allocation rules, and check whether different groups are comparable in business scope, observation opportunities, and capacity to handle. If only before-and-after changes can be compared, simultaneously record changes in season, promotions, supply, and sales processes as observations to be explained, and do not attribute the entire difference to automation. Regardless of the method, the observation period should match the real decision-making process; earlier received forms and later completed sales confirmations need to be connected according to a consistent time boundary.

Taking a hypothetical enterprise service company as an example, the team can first define what kind of inquiry is worth further sales communication, then record how these inquiries enter subsequent exchanges. Before comparison, confirm whether the judgment criteria are clear, whether different colleagues can use them consistently, and how duplicate submissions are handled. The actual pilot revolves around a change in investment that has been checked for applicable scope, retaining settings, implementation time, and exceptions, and not arbitrarily changing the definition of qualified inquiries during the pilot. In this way, the team tests the same business problem, not multiple criteria changing simultaneously.

The change log can include business problem, comparison scope, expected impact stage, responsible person, observation window, result definition, and anomalies to be explained. Anomalies should not be directly deleted after seeing them: website unavailability, supply changes, or sales response delays may reveal real constraints that prevent investment from converting into value. In review, separately examine reach, valid inquiries, and subsequent progress, and retain parts where the cause cannot be determined. If records are insufficient to distinguish changes in customer intent from changes in sales handling, the next step should be to supplement comparison conditions; only when results can be reasonably explained and the business can handle them, discuss whether to expand the scope.

08 GSP analysis

The portfolio perspective must bring allocation discussions back to business reality

Cross-channel budget allocation cannot rely solely on a single model output but must return to business reality. In terms of time granularity, weekly or monthly aggregation may mask short-term fluctuations; seasonal supply affects traffic costs and conversion rates; sales capacity determines the upper limit of lead handling; data boundaries limit which channels can be reliably measured. If these conditions are ignored, model conclusions may be detached from reality.

When conclusions from different measurement methods conflict, first check their respective assumptions. Attribution models assume touchpoints can be recorded and allocated; incremental experiments assume control groups can isolate external factors; mix models assume historical relationships persist. If conflicts arise from definition differences, unify conversion definitions and observation periods; if they arise from data gaps, supplement fields rather than forcibly averaging. Suppose a B2B service provider, the mix model shows a channel has high ROI, but incremental tests reveal that most of its conversions are organic. In this case, prioritize the experiment because it is closer to true causality.

A practical operation is to establish a business constraint checklist: clarify the sales cycle length, seasonal peaks, team processing capacity limits, and record the data coverage scope of each channel. When the model suggests increasing budget for a channel, first verify whether the leads from that channel enter qualified inquiries and closures, rather than just looking at form volume. Refutable signals include: model conclusions seriously contradict sales feedback, channel costs fluctuate abnormally without business event explanations, or data feedback delays exceed the decision cycle.

Ultimately, budget allocation should be a continuous calibration process, not a one-time calculation. The team should regularly compare model predictions with actual results and update assumptions. If the model cannot explain business changes, narrow its application scope or pause its use. Investment rationale must be based on explainable business value, not model complexity.

09 GSP analysis

Brand and sales controls must become conditions for automation operation

When AI-driven advertising enters the execution layer, brand and sales controls must become prerequisites for automation operation; otherwise, the system may optimize in the wrong direction. The specific business issue is that automation systems continuously adjust based on feedback data. If target definitions, creative maintenance, lead validation, and change records lack clear responsibility, advertising will amplify initial setting biases. For example, if the sales team does not provide timely feedback on lead quality, the system may treat a large number of invalid forms as positive signals, continuously increasing similar traffic, leading to budget waste.

In terms of causal mechanisms, automation optimization relies on a data loop: objectives confirmed by business determine the optimization direction; creatives and pages maintained by the brand affect conversion quality; leads validated by sales provide quality signals; change records managed by operations ensure traceability. If any link is missing, the system will operate under incorrect assumptions. The condition for establishment is that each role's responsibilities are clear, and data feedback includes business stages and exclusion conditions, such as distinguishing qualified inquiries from general browsing, to avoid duplicate submissions reinforcing erroneous patterns.

Consider a daily business example: a B2B service provider runs advertising for lead generation, and sales finds that many leads come from non-target industries. If a sales validation mechanism is not established, the system will continue to find similar audiences. A practical operation is: sales marks in the CRM whether leads match needs and feeds back status daily; operations set exclusion rules to pause invalid sources; the brand regularly reviews landing pages and copy to ensure alignment with target customers.

Refutable signals include: lead volume rises but qualified inquiry rate declines, or sales feedback contradicts advertising data. In such cases, pause automation expansion, check data definitions and responsibility execution, rather than simply adjusting bids. By clarifying control points, the team can judge the source of changes, keeping automation within an explainable boundary.

10 Suggested actions

Use a change log to launch an explainable pilot

The first step in launching an explainable pilot is to verify whether the account is eligible for migration and to fully record current settings. The team should create a change log table, listing each change item: change content, time, responsible person, expected mechanism, and comparison method. This table is not an administrative burden but a prerequisite for judging performance later: without records, it is impossible to distinguish whether performance changes come from AI Max, creative adjustments, or external environmental fluctuations.

In specific operations, a common problem is that multiple variables change simultaneously, making results difficult to attribute. For example, a team migrates to AI Max while also changing the landing page and adjusting sales scripts. Ultimately, lead volume rises, but it is impossible to confirm which step truly worked. Suppose we only modify the ad matching method, keeping the landing page and sales process unchanged. Then, observed conversion changes are more likely to come from the advertising side. Therefore, the pilot should limit to a single variable and clearly define an observation period that covers the full decision cycle to avoid misjudgment due to too short a period.

A practical operation is: first select a business line or geographic region as the test scope, set a core objective, such as qualified inquiry volume, rather than form submission volume. Record the number of leads entering sales communication each day, quality scores, and final closures. At the same time, set refutable signals: if the qualified inquiry rate does not improve within two weeks, or sales feedback indicates a decline in lead matching, pause expansion, check settings, or roll back. These signals should be defined in advance to avoid subjective post-hoc interpretation.

Finally, the change log should include an exception handling column to record execution deviations, such as creative review delays or budget fluctuations. After the pilot results are available, the team should compare the test group with the control group and check whether the assumptions in the log hold. If conclusions are clear, then discuss expanding automation to more channels. The entire process emphasizes explainability rather than pursuing short-term data growth, providing a reliable basis for subsequent scaling decisions.

11 Suggested actions

Resolve divergences between conclusions before scaling

When automated advertising expands from a single account to more channels, divergences between different measurement conclusions will first expose management issues. Attribution shows a certain path brings conversions, but incremental tests indicate limited actual increment; mix models give budget recommendations, but the sales team reports insufficient lead quality. These differences are not tool errors but rather each answers different questions. GSP believes that the prerequisite for scaling is not to find a unified number but to establish a review mechanism that can explain divergences, allowing the team to make reversible decisions amid uncertainty.

Suppose a B2B service provider runs both search and display ads. Attribution reports show display ads have significant assist value, but pausing that channel does not significantly reduce overall lead volume. This may be because organic search and brand awareness can cover some conversions, while the attribution model assigns credit to the last touchpoint. If the team does not conduct incremental comparisons, it will misjudge channel value. In this case, check data definitions, observation periods, and conversion definitions, and clarify the hypothesis: the effect of display ads may be overestimated. Only after clarifying the source of conclusion differences can one decide whether to expand budget.

A practical approach is to create a change log table, recording each adjustment's content, time, responsible person, expected mechanism, and comparison method. For example, change an ad group from broad match to AI Max while keeping a similar audience as a control, with an observation period covering the full decision cycle. If sales feedback indicates qualified inquiries have not increased but form volume has risen, prioritize checking whether feedback events count invalid submissions as conversions. The team needs to set clear signals: when quality feedback and incremental judgments do not support expansion, pause or roll back rather than continue increasing investment.

Scaling should also consider resource capacity conditions, including sales follow-up capability, creative supply, and data feedback stability. If customer service cannot respond to new leads in a timely manner, or the CRM does not record subsequent stages, any volume growth from automation will be difficult to convert into explainable value. It is recommended to review quarterly, based on quality feedback, incremental results, and capacity, to decide whether to maintain, modify, expand, or pause. Investment rationale should be based on explainable business value, not the launch of platform features.