AI placed a competition point from generating creativity to signal governance and control

01 Facts & context

October update to link construction plans, control plans, quantitative effects.

2026-10-27, Microsoft Advertising, October product update, with the theme of building, directing and measuring advertising activities through AI. The first change disclosed by the source was the creation of an AI-driven advertising campaign, whereby advertisers registered to provide websites, marketing plans, brand guides and pictures of products, and the system generated an edited business snapshot covering product services and advertising expressions, allowing for the publication of the first Performance Max advertising campaign after a review by the advertiser and the possibility of continuous revision of the budget, bid strategy, text and material.

The second change focuses on the controllability of maximizing impact: age and gender orientation are officially available and equipment orientation will be introduced to all advertisers in the coming weeks; Targeters exclude access to open pilots, brands are excluded from fine-tuning, and there is a choice to retain brand searches for commercials that merely serve as text advertisements; The search term report and the bid auction insight report are available and the search word landing page report will be rolled out gradually from mid-July 11. The third is the introduction of information guidelines into the open pilot, which applies to maximizing impact and the introduction of AI Max ' s search advertisements, with a maximum of 40 rules to regulate claims, themes, concepts, tone or otherwise avoided expressions. The fourth is the official use of the conversion API (CAPI) through server-to-server connections, real-time access to the web site, applications, CRM and the porter sales point system for conversion data, and parallel use with the common event tracking label.

02 GSP analysis

The real strategic choice is to authorize boundaries, not to use AI.

These updates appear to have improved efficiency, essentially putting a governance issue before advertisers: how much judgement is passed on to the system and what is left in the hands of others. The AI construction plan, which reduced “from zero” to “reviewed snapshots”, saves start-up time, but the snapshot is a simplified expression of business, and the deviation is continuously magnified by subsequent automatic bids and automatic generation once the advertising owner has been hastily confirmed. The first level of strategic choice is therefore whether the enterprise is willing to allocate real professional manpower to “review”, rather than treating it as an over-click confirmation. This choice determines whether the AI construction plan is an accelerator or a deviation amplifier.

The second layer is chosen to control particle size. The openness of age, gender, equipment orientation and audience exclusion means that maximization is no longer an entirely unexplained system, and advertisers can choose which dimensions to limit. The more limited they are, the smaller the space for systematic exploration, and the longer the short-term learning period is likely to be; The fewer restrictions, the more efficient they may be, the greater the risk of deviation from brands and compliance requirements. (b) As an hypothetical example, a business-oriented software service provider who is fully free of gender and age may spend its budget on a clearly unqualified population; A brand for young consumers, on the other hand, could miss a trans-equipment route if the equipment is overstretched. The qualification should be a limitation when the exclusion of a certain dimension has a clear operational justification (compliance, unit price, channel capacity); When limitations arise only from mistrust of the system and lack of data support, space for exploration should be retained.

The third level of selection is the priority for signal input. The conversion of API is officially available, turning “signal quality” into an investable asset. The willingness to invest in engineering resources to access the CRM and door-to-door systems depends on whether the enterprise can say what transformations really represent value. If enterprises continue to submit subsurface forms as their sole conversion objective, a stronger automatic bid can only be optimized to the wrong direction. The strategic trade-off is that it is preferable to reduce the number of conversion targets and to ensure that each goal has an understandable link to income or retention. This could be done through a conversion-calibre audit that lists and aligns the respective “success” definitions of websites, CRMs and shops, and decides which sources to access; At the same time, a reviewer responsible for the results of the operations was appointed, rather than a snapshot confirmed by the executive level.

03 GSP analysis

Values shift from generation efficiency to signal quality and binding quality

Understanding these updated value mechanisms requires a distinction between the two types of AI capabilities. The first category is capacity: writing, structure, material. The marginal value of such capabilities is declining rapidly, as competitors can generate and the homogenization of the text increases and the rate of generation no longer constitutes an advantage. The second is the ability to judge: to know what is a success, to know what is not possible and to know which groups of people should not be reached. Such capacity depends on the enterprise ' s own input and is difficult to replicate, and is therefore a source of differentiation. The information guidelines and the transformation of API correspond to the second type of capacity, which requires advertisers to display tacit knowledge.

The mechanism for signal quality is worth developing. The automatic bid is essentially a learning system that allocates the budget on the basis of the conversion signal, and the more complete the signal, the closer it is to the real value, the closer it is to the best distribution. The transformation of API activities that may be lost through server-to-server connections to supplementary browsers and parallel to labels improves signal coverage and continuity. But increasing coverage also reveals an old problem: if the definitions of transformation from different sources are not consistent, such as using registration as a conversion on a website, CRM as a conversion, the merged signals are diluted and the model's “success” is blurred. The realization of value is therefore premised on a unified transformation of calibres, followed by access. The test is to extract the same user before access, comparing the label to the difference between the number of returns and the conversion time of the API channels, which is too different to indicate that the calibre is not aligned.

Conversely, the mechanism for binding quality is: information guidelines reduce ex post work by pre-declaring boundaries. The rationale behind the suggestion that the guidelines should be used primarily for ideas, themes, concepts, tone and language that should be avoided is that broad rules are more stable than thin ones. Example of a hypothetical case: a high-end retailer with “not to suggest low prices for products”, a rule that could cover a large number of case scenes; And if there are dozens of specific phrase bans, they are difficult to maintain and easy to fail when models generate new expressions. The institutional conclusion is that the constraints should be at the level of principles and that the review should remain at the level of exceptions. It is recommended that enterprises allocate 40 ceilings as budgets, give priority to compliance red lines and brand core ideas, leave the remaining amounts to seasonal themes, and regularly calculate the rules that trigger the highest frequency, thereby adjusting the wording instead of adding new entries on a continuous basis.

04 GSP analysis

The bottlenecks in the user's journey are moving from construction to review and attribution.

By placing updates in the actual operating journey of advertisers, a shift in bottlenecks can be observed. The first half of the journey involved registration, provision of materials, the generation of snapshots and the release of plans, which were significantly reduced by the AI construction plan. During the journey, adjustments are being placed: a search term report will judge the intention to match, a competition will look at the location of the judgement, a land page report will be used to judge the correspondence of the intention to the page, and a direction will be used to exclude narrowness or liberalization. This period has not been compressed, but has been increased by the increase in the number of reports available. The latter part of the journey is attributable to the disk: the conversion of API provides a richer signal, but also requires that someone be able to explain the signal changes.

One of the organizational consequences of moving the bottlenecks forward is that most of the time spent on building the team in the past is now fast-tracked and time is flowing towards review and interpretation. If the team's formation remains unchanged, there will be the paradox of "the faster the plan is built, the more hasty the review". As a hypothetical example, a retail enterprise uses AI to build lots of multiple impact maximization activities before the season, and a snapshot review of each activity takes only a few minutes, resulting in a deviation between brand and commodity language budget allocations until search terms are reported online. This example is inspired not by the rejection of the AI construction plan, but by the need for clear checklists and those responsible for the review process.

There was also an easy-to-neglect break point during the journey: the search for the ground page report revealed an attempt to map the page, which linked the delivery problem to the site issue. If the landing page does not match the search intent, then the exact orientation cannot be converted. This means that the optimal boundaries of the advertising team extend to site experiences and require a fixed rhythm of collaboration with the product or site team, rather than the landing page as an established condition.

05 GSP analysis

The division of labour is to be redrawd by decision type rather than by platform

The impact of these updates on the organizational division of labour is not due to an additional tool, but to a change in the type of decision-making. Investment decisions can be divided into three categories: structural decision-making (targets, transformation caliber, budgetary framework), parameter decision-making (direction, bid, exclusion, guidelines), exceptional decision-making (exceptional signals, brand risk, competitive variation). AI is good at day-to-day implementation of parameter decision-making, and humans should retain the power to define structural decision-making and the power to adjudicate exceptional decision-making. If the division of labour continues to be divided according to "who controls which platform", structural problems will become unaccountable and inefficient for the consideration and approval of parameters.

In the specific case of the person responsible, it can be said that the head of the growth or market is responsible for the transformational caliber and objectives, since only he can judge which behaviour represents the value of the business; (a) The person responsible for the placement is responsible for the configuration of the parameters and maintenance of the guidelines, since only he is aware of the system ' s behaviour; The data or engineering manager is responsible for the transformation of API access and signal consistency; Brands or legal services are responsible for the presentation of compliance in information guidelines. The input output should also be clear: the operator enters a value definition and brand boundary, the technical party exports a reliable signal pipeline and the sender exports an accountable adjustment record. There is a need for a shared list of guidelines and a conversion dictionary among the three parties as a common language for collaboration.

The trade-off lies in concentration and fragmentation. If the guidelines were to be maintained by the respective personnel, they would be of varying calibres; If all are concentrated at headquarters, there will be a loss of responsiveness to local markets. The more feasible approach is to stratification: principles such as advocacy, compliance and tone are managed centrally, and enforcement levels such as local marketing and language practices allow limited delegation of authority and maintain audits. The condition for the determination is to concentrate when a norm involves legal risks or CCCs; Decentralization when its impact is limited to single activities and can be quickly rolled back.

06 GSP analysis

The sequence of inputs and the rebuttal of optimistic interpretations

When resources are limited, the sequence of inputs is in itself a strategy. The more logical order would be to harmonize the conversion caliber, then connect the conversion API, then establish information guidelines and, finally, expand the scope of use of the AI construction plan. The reason is that the latter relies on the former: there is no clear definition of success and automatic bids optimize the noise; In the absence of stable guidelines, volume-generated documents add to the review burden; Without reliable signals, the more reports are more difficult to judge. In turn, the pursuit of up-to-date speed would concentrate the technical debt on a positive season.

There are several optimistic interpretations of this sequence that need to be tested on a case-by-case basis. The first explanation is "tools with their own best practices, access is visible." Such statements are derived from the product distribution context, the effect of which depends on the quality of the enterprise ' s own signal and binding quality, the publication of the statement cannot be used as proof of effect, nor should it be assumed that the manufacturer ' s publicity has been independently validated. The second explanation is that "the increase in the number of reports equals the increase in transparency". Search words and competition insights do increase visibility, but visibility becomes value only if someone regularly reads and shapes the adjustment action, otherwise it is simply a data pile. The third explanation is "the closer the orientation, the more precise". In cases where the sample is limited, too fine orientation weakens learning outcomes, and the value of age and gender orientation depends on whether there are clear grounds for exclusion from the business.

Another alternative explanation is noteworthy: it was suggested that the AI construction plan would lower the professional threshold for placement and could therefore reduce the number of professionals. This inference confuses the operational threshold with the threshold of judgement. It is true that operations become simpler, but a review of snapshots, maintenance of guidelines and interpretation of signals require a higher degree of judgement. Example of a hypothetical case: An SME that has an administrative staff member who will be able to go online in the short term, but no one will be able to locate a site, attribution or orientation problem when a conversion signal is abnormal, the loss exceeds the human cost saved.

07 Suggested actions

Order of implementation: list of enforceable actions within four weeks

The first week of the conversion dictionary, led by the Head of Growth, lists all current transformational behaviour item by item, indicates the causal distance between each and the income or retention, combines repeats the caliber, removes the shallow layers of behaviour that have no operational meaning, and eventually forms a page of narratives. The output is a conversion dictionary and a list of duty bearers, and the test is that each transformation target is able to answer “why it represents value” on the spot. This step, which hardly requires technical input, determines the direction in which subsequent signal access and bids are optimized because of the wrong calibre and the greater the incomplete deviation in access. It is proposed to begin with three categories of transformation objectives: direct transactional, high-intensity trails, deep interaction and the establishment of a single numerical rule for each category.

Signal access assessment during the second week, whereby API access is converted by the data owner against the control, i.e., directly integrated, through the label manager or through a technical partner, to take stock of the coverage gaps of existing labels and prioritize access to sources that complement lost activities restricted by the browser, such as the CRM and the Door System. The output is the access programme and the certification plan, which is to indicate whether the change in signal levels after the line is in line with expectations, rather than just an increase in totals. In cases where enterprises do not have access to CRM or porter systems, it is recommended that the scope be limited to websites and applications to avoid one-time openings leading to confusion of calibre, and that manpower be difficult to verify multiple sources simultaneously.

In the third week, information guidelines will be set up, which will be drafted jointly by brands and placements, covering no more than twenty principle-level rules covering ideas, tone, avoidable expressions and language requirements, to be piloted in a limited number of activities and to observe whether the production of documents requires frequent back-to-work. The output is the standard list and review record. A controlled experiment was performed in the fourth week: a group of activities was selected to enable age gender orientation and audience exclusion, while another group was kept open, comparing the length of the period of study with the quality of transformation, while regular reading of search phrases and competition insight reports began, resulting in weekly adjustment sessions. Under the scenario, a retail enterprise could simultaneously record weekly cost curves between the experiment group and the control group in the fourth week as a basis for a retreat in the next phase.

The whole list is sorted along the following lines: caliber first, signal first, back-up guidelines, and finally control experiments, as the latter rely on the clean input of the former. In the event that output falls short of the target in a given week, for example, there are still objectives in the conversion dictionary that do not indicate value, it is recommended that the follow-up steps be suspended and completed instead of being pursued in parallel. Such collusive arrangements sacrifice a bit of speed, but allow each adjustment to be traced back to specific changes, avoiding multiple factors that make it impossible to judge the source of the effects.

08 Suggested actions

Indicator and experimental design: Quality rather than Quantity

The indicator design, in order to avoid being misled by the generation efficiency, suggests three layers: signal layers to see how much conversion coverage is consistent with the transformation calibre, i.e. how many should be recorded and whether the conversion definitions from different sources are consistent; The control layer looks at the rate of return to work following the triggering of the guidelines and the changes in the audience structure resulting from the exclusion rules; The result layer looks at unit conversion costs and conversion values, but must be compared to the control group. It is not possible to distinguish between tools and seasons, especially during the second half of the year, when seasonality is the main alternative to interpretation, and it is therefore recommended that control groups be used as a pre-condition for all conclusions.

The experimental design suggested a comparison over time rather than a comparison over time. Assumptions are given that maximization activities are divided into two groups by area or by product line, with one group enabling new orientation and exclusion, one group maintaining the original set-up and running a learning period sufficiently long to be compared and avoiding drawing conclusions during the learning period. The information guidelines could be assessed using the return rate and review time as proxy indicators because of the high direct measurement cost of brand consistency. The API conversion assessment should focus on changes in signal increment and bid behaviour, rather than simply on increases in conversions, since the increase in signal numbers may in itself be merely an expansion of calibre.

Stop conditions need to be set. If the quality of transformation of a group of experiments has not improved after the learning period and the cost has increased, then the set-up should be reversed instead of continued additions; If the guidelines result in a large number of rejections of the generated case, they indicate that the rules are too detailed to accommodate the principle layer; If the number of signals after access increases but the unit transformation value decreases, the conversion dictionary verification caliber should be returned. The value of the indicator is to support retreat rather than to justify the input, and it is therefore recommended that the regression threshold and the decision maker be identified before the experiment begins, avoiding ex post facto justification for the input.

09 Suggested actions

Applicable conditions and when to adjust judgement

These updates are of different value to different types of advertisers. Firms with clear transformation links, first-party data and high brand expression requirements have the greatest benefit because their signals and guidelines can form positive feedback and the more complete and clear access is, the better the basis for automatic optimization. Enterprises with long conversion cycles and complex decision-making chains, such as large durables or enterprise-level services, need to be more cautious: server-end signals improve coverage, but the attribution of long-cycle conversion needs to be manually judged and it is not appropriate to set too narrow an automatic bid target. Small and medium-sized enterprises with a weak data base can begin with a uniform calibre and a small number of guidelines, thereby delaying complex access works.

There are several types of signals that need to be fine-tuned. If the signal volume increases after the conversion of the access API but the unit conversion value decreases, indicating that the calibre is mixed into low-value behaviour, the conversion dictionary correction should be returned; If the orientation and post-exclusion learning period is introduced and the cost is significantly prolonged and does not improve, it indicates that the constraints are excessive and that some dimensions should be liberalized; If the maintenance costs of the information guidelines continue to be higher than the savings from back-to-work, they should be rewritten rather than updated to indicate that the rules are designed to deviate from the principle layer; If a snapshot of AI ' s planned output repeats similar errors in a number of activities, indicating insufficient or vague input of branded material, then the input should be refined rather than increased manual review. Assuming that an enterprise finds the same type of error in two consecutive reviews, the issue should be included in the branding material template.

Finally, it is important to retain the judgement that the need to increase the controllability of instruments does not amount to a reduction in the need for control. The more controls are available, the more the need for clarity as to who will decide what to control, otherwise the controls themselves will become a new management burden. Businesses should review the list of conversion dictionaries and guidelines on a regular basis, e.g. on a quarterly basis, remove the rules that no longer apply and add to the emerging operational constraints. This review mechanism is better than any single configuration to determine the long-term effects and is a real entry point for transforming product renewal into business capability.