Microsoft Ads: more control over AI, but what results?

When an advertising platform promises to launch a campaign in a few minutes, I think about the work that comes before that button. What offer are we bringing to the market? What words can we use? What must happen for a request to be considered useful? These are less spectacular questions than automatically generating an ad, but they determine what the system will try to achieve with our money.

The Microsoft Advertising update at the end of October interests me for this: alongside the speed of creation, tools appear to direct messages, read searches better and connect the results to company systems. The question is understanding how to use them without mistaking a new setting for an control problem that has already been solved. A written limit can be ambiguous. A correctly transmitted event may represent the wrong result. A more detailed report may lead to action being taken too early.

My position is to start from a small and verifiable perimeter: commercial promise, landing page and result to be measured must convey the same work.

The short answer: Microsoft Advertising changes can make Performance Max more observable and better targeted, but I wouldn't increase the budget solely because these features have been released. I would first check the availability in the individual account, the correctness of the texts and the correspondence between advertising conversions and business results. This article is a documentation-based analysis with a proposed verification method, not the report of a test on real campaigns.

What has been announced, and what is yet to come

Performance Max search terms, auction reports, age and gender targeting are available in the October 28, 2026 Microsoft Summary. Device targeting is rolling out in the following weeks; the report that links terms and pages will begin rolling out in mid-November. Audience exclusions and messaging guidelines are in open pilot; for brand exclusions the supplier describes a pilot refinement. The guidelines, up to 40, concern Performance Max and Search with AI Max. Conversions API is declared generally available and works alongside UET. Assisted creation uses the site and company materials for an editable profile before launch.

Documents were verified on November 3, 2026. An Italian account has not been verified: the summary does not offer a complete matrix of countries, languages ​​and accounts. Open pilot means an access phase to be verified, not universal availability. The report expected from mid-November remains future as of the date of this analysis.

This distinction has an immediate practical consequence. In the team plan I would separate what we can review today from what requires a support request and what we can only put on the calendar. Building a procedure around an option that is still absent generates work to be redone and makes it difficult to understand whether a problem depends on the strategy or on access to the product.

The before and after: more tools to ask better questions

Performance Max is a campaign that entrusts some of the choices on distribution, combination of materials and optimization towards the set objectives to an automatic system. It does not mean that every decision disappears from the account: it means that human work must clearly define the field within which the system searches for opportunities. Automation alone does not know the minimum margin that makes a sale sustainable or the conditions that make a contact suitable for the service.

The historical starting point matters. The May 7, 2026 Microsoft Document on Performance Max Transparency already described publisher-site reporting with clicks, spend, and conversions, landing page information, and the gradual rollout of search-term reporting. Auction information was then presented as subsequent. I therefore do not attribute the birth of all this visibility to October: the new summary updates a development that began earlier.

For me the interesting change is organisational. Previously, a discrepancy could end up in a general discussion about the opacity of the algorithm. With more observable elements we can formulate more precise hypotheses: are we intercepting a need that is not very relevant? Does the ad promise something that the page doesn't clarify? Are we reading behavior that never leads to a negotiation as successful? Each question suggests a different check.

However, I would not consider a report to be the complete record of what the model thought. A table shows what the platform exposes according to its reporting rules. I would use it to identify inconsistencies to investigate further, avoiding turning every weak line into an automatic exclusion. The distinction between observing and intervening is essential when the number of events is small or the purchase path is long.

A hypothetical case: selling a service without attracting the wrong need

Let's imagine a company that designs office furniture and works with companies in a specific geographical area. It offers a site visit after an initial evaluation of the project. It does not sell individual chairs online and cannot promise immediate deliveries. This is an invented teaching example; it it it it does not represent a client or campaign for which I am reporting results.

The main page contains photographs of environments, a description of the method and a form. If the team defines success as simple form submission, even a request for a home chair could appear as a positive conversion. The contact exists, but the salesperson must take time to explain that this is not the service offered. The distance between metrics and business activity arises before the algorithm: it is in the definition of what we asked to achieve.

I would start from a shared sheet between marketing and sales. The form describes the person to be reached, the type of project, the area served, the information necessary for an evaluation and the reasons for exclusion. There is no need to collect more personal data in principle. We need to distinguish what is necessary to understand the need from what adds friction to the form without helping any decision.

From this brief I would derive three consistent elements. The message advertises an office design consultancy. The page explains which projects it is aimed at and which steps follow the request. The measurement distinguishes the form received from the contact rated as relevant. If one of these links is missing, a new advertising option can speed up a confusing process.

Messaging guidelines: making brand rules explicit

The English name indicates instructions with which to orient the generated text. Their usefulness, in my method, begins with the quality of the formulation. Writing “be professional” leaves many interpretations open. For the imaginary company I would instead write: describe the service as the design of work spaces; avoid promises on delivery times; do not present the inspection as automatically available for any request. They are examples of proposed editorial criteria, not instructions whose effectiveness I have verified in the product.

Each criterion requires a reason. If I ban a reference to immediate delivery, I need to know what the company can actually promise. If I ask to avoid the language of the low price, I must have an understandable alternative proposal: analysis of the spaces, quality of the materials, coordination of the project. Otherwise I get an advert that is formally conservative but lacks a reason to be chosen.

I would keep the rules short, distinct, and ordered by importance. First the statements that could deceive, then the constraints of the offer, finally the tone. A guide that contains the entire company manual can hide the crucial instruction. The problem is not filling the available space: it is allowing those who approve a text to immediately recognize which rule they are checking and why.

A written limit remains an instruction to be verified. I would not treat this as a technical guarantee that no problematic formulations will appear. If a promise would be unacceptable even once, the process must include monitoring of visible materials and a responsible person capable of stopping or correcting the spread. The presence of the field in the product does not transfer knowledge of the real conditions of the service to the platform.

Here I see a direct link to the content factory with AI and human control: Producing more variants makes sense when the team knows how to review them and keep them consistent. For the Ads radar the principle becomes even more concrete, because an approved text can start to spend budget and generate expectations that the service must then support.

Creativity is evaluated together with the page

In the furniture example, a headline can be accurate in isolation and become misleading next to an image. A very broad environment, a promise of simplicity and a context-free form can suggest a different offer than what is available. For this reason I would review the combinations as units of meaning: text, image, call to action and landing page. The absence of grammatical errors is not enough.

I would put a question next to the material: what would a reasonable person expect to receive after seeing it? If marketing's response differs from that of sales, the work is not done. The conflict must be resolved on the service and the promise, before asking the automatic system to multiply the advertisements. Creative variations become more useful when they explore different reasons for choosing the same offering, without implicitly changing it.

Page control deserves the same attention. A general page may be suitable for presenting the company but not suitable for collecting requests for a specific service. When the report suggests a discrepancy, I would check on the site what information is immediately visible, whether the form names the service and whether the essential conditions are also clear on a mobile phone. Correcting these points may make more sense than trying to find new advertising wording.

However, I wouldn't attribute any improvement to a single setting if in the meantime we rewrote the page, changed the form and replaced the images. It would be an overall intervention on the path, to be described as such. This honesty is also useful within the company: it allows you to understand which work to retain and which hypothesis requires a separate experiment.

Conversions: what we want the system to learn to look for

A conversion is an action that we have decided to count as a result: a purchase, a request, a reservation or another defined event. It is not a universal synonym of acquired customer. In our hypothetical company I would at least distinguish the request sent, the relevant request after evaluation and the assignment actually accepted. They are different moments, with different times and different possible errors.

The CRM, i.e. the system that collects and organizes commercial relationships, can help describe the path following the click. But before linking it to advertising I would ask people to agree on the meaning of the sales stages. “Qualified” can mean that the contact details are valid or that the project has been reviewed and is eligible for service. If two colleagues use the same label with different meanings, the data will only appear orderly.

My priority would be to get a repeatable definition. For example: project for a workspace in the area served, request sufficiently described and availability for a subsequent discussion. It is a hypothetical definition to be adapted to the company, not a Microsoft standard. Its usefulness is to allow you to verify why a contact was classified in a certain way without depending on the memory of whoever managed it.

The delay must also be admitted. A request today could be evaluated later and become an assignment again later. If we compare the expenditure just incurred with the results already concluded, we risk judging a group of contacts still in progress as incomplete. I would look at comparable sets of requests by entry date and maturation, indicating which states remain open. Waiting does not mean giving up control: it means respecting the time of the process we are measuring.

CAPI, explained without confusing it with a guarantee on results

An API is an interface with which two software systems exchange information according to defined rules. The Conversions API Microsoft Learn documentation, accessed November 3, describes sending events from the server and recommends use in conjunction with the Universal Event Tracking, or UET, tag that operates in the browser. For the same event sent by the two channels, it indicates to keep the tag identifier, event name and stable event identifier consistent. The documentation also distinguishes accepted request and actually visible conversion, with controls on objective, timing and attribution identifiers.

In simple words, let's think of a request recorded both by the site and by the management system. They are two descriptions of the same fact, not two different people. Reconciling them means preventing the second message from causing an additional result to appear. The analogy serves to understand the problem; does not itself describe all the technical details of the integration and does not replace the supplier's specification.

I would ask for controlled proof of the correspondence between the company event, the sending and the result read in the advertising system. Implementers must be able to reconstruct the path of a test event without exposing personal data in work reports. An accepted message confirms a technical step; it does not automatically demonstrate that the data is complete, correctly attributed or useful for an economic decision. They are distinct controls that deserve distinct outcomes.

The server connection is not a shortcut to bypass people's choices about their data. In the project I would include those who manage consent, access and information flows, so as to establish which events are appropriate and under what conditions they can be transmitted. You don't have to wait for an accident to assign these responsibilities. This is an organizational criterion, not a legal evaluation of the specific data processing.

To understand why a connection cannot determine the meaning of data on its own, see the guide to MCP, APIs and files. The means by which information arrives changes; we still need to decide who produces it, what it represents and who can correct it.

Budget and distribution: the question before moving money

The budget is a choice about what opportunities to look for and how much uncertainty we are willing to bear. An ad that's quicker to produce doesn't automatically make a new geographic area, larger audience, or different type of inquiry convenient. In the case of furniture, attracting projects far from the operational area could increase contacts and worsen the sustainability of the work. The quantity of requests must be read together with the concrete possibility of serving them.

Before moving money I would write the economic hypothesis in one sentence: we want to obtain more relevant requests for that service without exceeding the cost that the company can bear and without increasing an unproductive commercial load. The sentence forces you to name the result and the constraint. If the team does not yet know how to quantify the sustainable cost, I would record this uncertainty and limit the exposure, rather than adopting a threshold suggested by someone else's case.

For the cost of acquiring relevant contacts I would use a simple formula: advertising spend of the observed group divided by relevant requests from the same group. I would keep sales-team labor and integration costs separate, or explicitly include them in a second, larger measure. If no relevant requests arrive, the ratio does not become zero: it cannot be calculated as the unit cost of a success. It is a signal to be examined together with the maturation time.

Distribution also requires a choice about the role of brand-related searches. Those who are already looking for the company and those looking for a generic solution may find themselves at different times. I would avoid presenting all conversions attributed to the campaign as evidence of new demand. To evaluate a change you need the context of the other channels and an explicit question about what would probably have happened even without that intervention. Reporting helps discuss; it does not resolve this causal question on its own.

The test I would propose, with clear stopping conditions

I have not run an experiment on these releases. The following is the design I would propose to a team before activating them. I would choose a service, a stable page and an objective that is also understandable to the salesperson. I would record the initial status of the account, the controls available and the access conditions confirmed by support. A missing feature would be left out of the test, with the reason written.

I would separate the technical control from the commercial evaluation. The first must establish that messages respect the rules, pages are relevant and events are not duplicated. The second must establish whether the requests that have had time to progress through the sales process are consistent with the service and whether the cost makes sense. Starting the economic evaluation before having reliable data risks producing a clear verdict on the wrong measurement.

To compare a modification I would choose, when actually available and suitable for the account, a design that keeps the conditions comparable. If there is no possibility of isolating the intervention, I would describe the before/after comparison as observational. Seasonality, promotions, competition and sales rep response times could explain part of the change. I would not call what we have not been able to separate from other causes an "AI effect".

I would define in advance when to suspend: a promise incompatible with the service, an incorrect page, duplicate events or exceeding the agreed spending limit require intervention. I wouldn't wait until the end of the period to fix an error that harms people or data. The presence of normal fluctuations, however, is not enough to justify a daily change of the rules.

Duration would not be a universal number. It would depend on the volume of requests, the cycle with which they are evaluated and the possibility of observing an economically significant outcome. If the sample remains small, the result may be “we don't know yet”. It is a useful conclusion if it prevents you from increasing spending based on a few lucky episodes or turning off a channel after a normal delay in the sales process.

The team load changes, and must be counted

Automation can reduce manual setup work and increase definition and verification work. I would make it visible in the activity plan. The media specialist controls distribution and spending; those who know the product verify the promises; the salesperson returns a coherent opinion on the requests; whoever manages the data verifies the events. In a small team the same person can cover multiple roles, but the tasks must remain distinguishable.

For labor costs I would use actual hours logged per task, not an imagined savings percentage. There are different hours spent once defining the rules and those needed every week to correct texts or reconcile data. If generation is fast but revision continues to increase, the organizational benefit may be less than perceived at launch.

I also see a consequence for those who are learning the trade. If the system directly produces a campaign, a junior person risks seeing the output without reconstructing the choices. I would use the revision to make the reasoning explicit: why this audience, this promise, this event? It's a concrete way to maintain competence in the team while part of the operation becomes more automatic. It is not a measured effect of the release, but a training choice that I consider prudent.

The reading Google Ads and AI Max addresses another advertising system. The link is used to compare the type of work required of the team, without transferring names, settings or availability from Google to Microsoft: products similar in objective are not interchangeable in the procedures.

What could improve, and what would change my mind

I would see an advantage if the team spent less time correcting promises, reconciled events more easily and could explain which inquiries its advertising budget was paying for. To claim an economic benefit, I would also want commercial outcomes that have had time to materialize and costs consistent with the service. More complete measurement can increase reported conversions without increasing accepted projects: that is an informational improvement, which must be distinguished from business growth.

I would consider unfavorable a situation in which the quantity of contacts grows but the sales representative reports increasingly less suitable needs, or in which the texts continue to exceed the limits of the offer. I wouldn't respond with a generic "the algorithm needs more time". I would double-check the definition of the event, message and page. Time can help observe a dynamic; it doesn't single-handedly correct a flawed definition of success.

I would support expanding use when the process remains understandable even to someone who did not configure the campaign: traceable data, clear rules, assigned responsibilities and inquiries assessed consistently. I would postpone expansion if the integration needed constant manual corrections or the account lacked features essential to the test plan. A new feature can remain interesting even when the reasoned decision is to wait.

My choice today is to use this update as an opportunity to make the pact between marketing, product and sales explicit. First let's define what we can promise and what is worth as a result; then we check if the new tools help us respect it. The operational criterion is simple to formulate and challenging to apply: I would increase delegation to the system only when the team can better explain and control what it is delegating.

FAQ

Are the Microsoft Advertising updates already available in Italy?

Not all have the same status. The announcement distinguishes general availability, open pilot and subsequent rollouts; does not confirm every country, language and account combination. Before planning a test, you need to check the features in your account and with support.

Do the messaging guidelines guarantee texts that are always correct?

I would treat them as instructions to guide generation, to be verified on actual messages. They do not replace human control over commercial promises, terms of service and consistency with the page.

Do more conversions after CAPI mean more customers?

Not necessarily. The detection of existing actions could be improved. It is necessary to compare consistent events, avoid duplicates and also read the commercial results achieved in the same group of requests.

Have you tested these releases on real campaigns?

No. It is a documentation analysis verified on November 3, 2026, with a hypothetical case and a proposed verification design. It does not report customer results, an experiment performed, or demonstrated ROAS increases.

Sources and scope of verification

The central sources are the October 28, 2026 Microsoft Advertising summary, the May 7, 2026 reporting precedent, and the Conversions API technical documentation. Documentary verification: November 3, 2026. No performance statistics are used as evidence of effectiveness; no accounts or campaigns were observed. The control of the seven radar platforms is recorded separately in the editorial register, without turning it into seven news items in the text.

Who writes

I am Giorgia Palazzo, CMO, founder and teacher. I connect media, performance, data, AI and automation to marketing decisions. Here I distinguish the availability declared by the supplier from the method I would propose to evaluate them: the responsibilities for promises and results remain part of the team's work.