AI-Driven Advertising: What to Hand to Machine Learning

Google, Meta, LinkedIn and TikTok have built machine learning into the way their ads are bought and made. They use it to set a bid for each auction, to decide who sees an ad, to assemble and test creative, and to decide which ads get credit for a sale. Some of it is the default: Google Ads, for example, uses data-driven attribution for most conversion actions unless you change it.

So the practical question for a marketing leader is less whether to use AI in advertising than which decisions to hand over, on what evidence, and what your team has to keep. This guide explains where machine learning already runs in paid media, what it needs from you to work, where it goes wrong, how to measure what it is really doing, and how to test a handover before you move budget.

DisclosureJake Jorgovan is a partner in Fieldtrip, a group of marketing agencies, which is mentioned in this article.

What AI-driven advertising means in practice

In paid media, "AI" usually means two kinds of system. The first is predictive: models that estimate how likely a person, at a given moment, is to take the action you want, and then act on that estimate by raising or lowering a bid, choosing an audience or picking a placement. The second is generative: tools that write headlines, produce image and video variations, or build a whole campaign draft from your website.

Both work from what you give them: a goal, a conversion to count, creative assets, budgets and limits. The platform decides how to reach the goal. What the goal is, and whether reaching it is good for the business, stays your decision.

Where machine learning already runs your ads

The table maps the main jobs to what the platforms' own documentation says their systems do, and what the advertiser still sets.

Bidding

Google describes Smart Bidding (opens in a new tab) as strategies "that use Google AI to optimize for conversions or conversion value in every auction". At auction time it weighs signals such as device, location, time of day, the actual search query, whether the person is on one of your remarketing lists and, on Display, their activity on your site. It needs conversion tracking switched on. Google makes advertisers responsible for using it within its personalized advertising policies, and asks them to review the signals it uses to decide whether Smart Bidding suits their business.

Programmatic display and video work the same way. Google's Display & Video 360 (opens in a new tab), a platform for buying ads across many publishers, uses automated bid strategies that optimize a line item's bid from two predictions: how likely an impression is to lead to a click, conversion or view, and whether it is available at a good price compared with the market. Its signals include inventory, audience lists and creative effectiveness. New line items default to an automated bid strategy. You can set an average price ceiling per thousand impressions, but Google warns that it can be exceeded at times.

Audience

On Meta, Advantage+ audience (opens in a new tab) lets the delivery system expand the pool of people it can reach beyond your targeting. Meta's developer documentation lists the constraints it does not expand, including location, minimum age, language and custom audience exclusions. Before you switch it on, check that every limit you need is one of them, and do not assume an audience suggestion is a hard limit.

Creative

A responsive search ad (opens in a new tab) takes up to 15 headlines and 4 descriptions, and Google tests combinations over time to learn which work best for which queries. The assets can appear in any order, so each one must make sense alone; Google says text that must appear in every ad, such as a legal disclaimer, belongs in headline 1 or description 1, pinned to that position. Meta's standard enhancements (opens in a new tab) create variations of a single image or video ad, such as a different aspect ratio or a template, and show each person the variation it expects them to respond to.

Generative tools go further. TikTok's Symphony Creative Studio (opens in a new tab) generates videos from simple inputs and images from text prompts, and adds narrator or product avatars. LinkedIn's getting-started guide for Accelerate campaigns (opens in a new tab) says it reads a page you supply to describe your product and audience, then uses generative AI to write ad copy and select images with the highest predicted click-through rate.

Whole campaigns

Performance Max (opens in a new tab) is Google's goal-based campaign type that reaches all of its inventory from one campaign. Google's own summary divides the work clearly. You define the conversion goals and values, supply text, images and video, share audience signals, and set budgets and brand-safety settings. Google bids, generates images and video from your assets, and moves the budget toward the placements that perform. With final URL expansion on, Google may send people to a different page on your site and write matching headlines, and the help page states that you remain responsible for the accuracy of that page content and of every generated asset.

What the system needs from you

A conversion that reflects value

The bidding systems optimize toward the action you tell them to count. For a business that sells online, that can be the sale itself. When a campaign generates leads, the action the platform sees is a form fill or a demo request, and any sale happens later, in your CRM. Optimize to form fills and the system is built to find more people who fill in forms, whether or not they ever buy.

Each platform has a route for sending the later outcome back:

• Google offline conversion imports (opens in a new tab): you store the click ID with the lead and, when that person signs a contract, send it back to Google Ads with the conversion. Google recommends its enhanced conversions for leads, which match hashed details such as email addresses.
• Meta Conversions API (opens in a new tab): a direct connection from your server, website platform, app or CRM, which Meta says lets ads optimize for later actions, including customer scores.
• LinkedIn Accelerate: the guide says LinkedIn Conversions API events can inform the campaign's targeting and measurement. It separately suggests audience inputs that reflect the real goal, such as a list of past leads who booked demos.

Each of these sends customer data to the platform, so it needs the permission your privacy notice and the law require. Google's page says to get consent where the law or its policies require it.

Volume and time

The systems learn from results, and they need enough of them. Meta says an ad set usually leaves its learning phase (opens in a new tab) after about 50 results in the week after its last significant edit, and that during learning, performance is less stable and cost per result is usually higher. Significant edits restart the learning, and splitting a budget across many similar ad sets leaves each with less to learn from. Google recommends judging Smart Bidding over a period with at least 30 conversions, such as a month or longer, or 50 for Target ROAS. Those are evaluation guidance, not a minimum for switching it on, and Smart Bidding can draw on data from your other campaigns. For data-driven attribution, Google recommends at least 200 conversions and 2,000 ad interactions in 30 days.

A handful of qualified leads a month makes any of this hard to judge. Consider consolidating similar campaigns, or optimizing to an earlier qualified stage that happens more often, while you keep checking what those leads become.

For B2B companies looking to generate more qualified leads, working with the best cold email agency for B2B leads (opens in a new tab) can complement paid advertising by reaching relevant decision-makers directly.

Image

Zoom

The approach described in this article, shown as a visual guide.

What to hand over and what to keep

The familiar mistakes are running AI on messy or disconnected data, personalizing beyond what the buyer finds useful, letting model output replace the judgment of people who know the accounts, ignoring consent and bias, and adding tools faster than anyone can measure them. The table gives each decision an owner.

Two of these deserve more than a table row.

Ownership of generated creative. The U.S. Copyright Office concluded in January 2025 (opens in a new tab) that generative AI output is protected by copyright only where a human author has determined sufficient expressive elements, which "the mere provision of prompts" does not do. Whether you may use the output commercially is a question of each tool's terms; TikTok's page, for example, says outputs belong to the user unless otherwise stated. If owning a campaign's creative matters, plan for human authorship, arrangement or modification, and take legal advice.

Data and consent. Platform targeting and conversion matching run on personal data. Build on first-party data collected with the permission your intended use needs, and use contextual placements where you lack that permission or the subject is sensitive. The behavioral targeting guide covers the browser, platform and legal limits in more detail.

How to measure what it is really doing

Platform reports show conversions the platform attributes to its own ads. Google's data-driven attribution (opens in a new tab), for example, compares the paths of people who convert with those who do not and gives more credit to the ad interactions that appear to matter, across Google's own Search, Shopping, YouTube, Display and Demand Gen ads. That is useful for steering within Google. It does not tell you whether a sale would have happened without the ad, or what your other channels did.

The gap can be large. In a 2018 working paper (opens in a new tab), Brett Gordon and Florian Zettelmeyer of the Kellogg School, with two Facebook researchers, compared 15 randomized advertising experiments at Facebook with the estimates that common observational methods produced from the same data. The observational methods mostly overstated the effect, sometimes understated it, and in half the studies were off by a factor of three for purchases. In one campaign, simply comparing people who saw the ad with people who did not implied a 316% lift; the experiment measured 73%.

To estimate the conversions the advertising caused, use a lift study with a group withheld from the ads:

• Google Conversion Lift (opens in a new tab) splits people, or geographic areas, into a group that sees your ads and one that does not, and reports the conversions the ads caused. Google says it is not available to every account; its page directs advertisers to their Google account representative.
• Meta Conversion Lift (opens in a new tab) works the same way and lists, among the questions it answers, whether your attribution model agrees with a controlled experiment. Meta's guide, read in October 2026, starts from a campaign launched in the past year with at least $5,000 in spend and 500 conversions, counted under a 1-day click, 7-day click or 1-day view attribution setting. Both figures scale up for campaigns longer than 90 days. Self-serve tests also need a qualifying data source, such as the Conversions API.

A lift study compares ads with no ads. To compare a new setting with your current one, use a split-campaign experiment instead. Google custom experiments (opens in a new tab) split a campaign's traffic and budget between the existing campaign and a copy with your changes, for Search, Display, Video and Hotel campaigns. Google recommends a 50% split and warns that changing either version mid-test makes the result harder to read.

B2B deal volume and sales cycles also limit what a test can establish. On a Leaders of B2B episode in December 2024, I described how many B2B companies close 20 deals a year, each worth half a million dollars, and how a few interactions that no one can track can move those decisions. As I put it then, "there's intuition plus data as a guiding light, but it can't be the full piece there." In that episode (opens in a new tab) I was talking about brand and messaging.

An ad algorithm has the same blind spot, because it can only optimize what it can count. Where deals are few and large, judge it on pipeline and revenue in your own records, and keep brand and message decisions with people.

How to test a handover before scaling

1. 1

   Choose one campaign and one decision

   Pick a campaign with steady conversion volume and hand over one decision, such as bidding or audience expansion, while the rest stays as it is. Several changes at once leave you unable to say which one worked.

2. 2

   Fix the conversion signal first

   Make sure the conversion the campaign optimizes for reflects value. For lead generation, send qualified leads or closed deals back to the platform before the test starts, so the system learns from the outcome you want.

3. 3

   Compare it with your current setup

   Run the handover as a split test against the settings you have now. On Google, use a custom experiment with a 50% split where the campaign type supports one. A lift study answers a different question, whether the ads cause conversions at all; run one on the setup you keep if your account qualifies. Agree on the success measure, the budget and the end date before launch.

4. 4

   Leave it alone long enough

   Avoid significant edits during learning. Set the test's length from the conversion volume and sales delay you expect: the platforms' learning and evaluation figures do not, on their own, show that a test has enough evidence to name a winner.

5. 5

   Judge it on your own records, then scale or roll back

   Compare cost per qualified lead, pipeline or revenue for each arm of the test in your CRM, alongside the platform's report, and allow time for leads to mature. If the difference is too uncertain to support a decision, record the test as inconclusive. Scale a supported improvement, keep the controls that protected the test, and move to the next decision.

Image

Zoom

The approach described in this article, shown as a visual guide.

To compare whole campaign systems instead, LinkedIn's Accelerate guide describes an A/B test against a Classic campaign. For lead generation, it requires single-image ads, English copy, at least 14 days and a lifetime budget of $3,000 per campaign ($6,000 in total). That compares the systems together; it does not isolate which automated decision caused the difference.

Questions to ask your team or agency

Common questions

How is AI used in advertising?

Mainly in four places: setting bids in each auction, choosing who sees an ad, assembling and generating creative, and dividing credit for conversions among ads. The platforms build it into their bidding strategies and campaign types, and some of it, such as data-driven attribution in Google Ads, is the default.

How do Google and Meta use AI for ads?

Google's Smart Bidding sets bids from predicted conversions, responsive search ads test combinations of your text, and Performance Max runs bidding, creative and placement across Google's inventory from one campaign. Meta's Advantage+ settings expand audiences within the limits you set and create creative variations, and its delivery system learns from results during the learning phase.

Can I use AI-generated ads commercially?

That depends on each tool's terms, which you should read before using its output in an ad. Owning the output is a separate question: the U.S. Copyright Office's position is that prompts alone do not make AI output copyrightable, so human creative work on the final ad matters if exclusivity does.