Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

AI in Marketing › AI Applications in Marketing

AI in Ad Platforms

How AI already runs inside ad platforms: automated bidding, audience expansion, asset generation and combinations, and campaign types that bundle them.
Guideintermediate11 min readUpdated

Automation is already the default

New campaigns on both platforms start with automation on. Google Ads starts new campaigns on an automated bidding strategy unless the advertiser changes it and, in Display and several other types, documents expansion of the chosen targeting as enabled by default; its responsive ad formats assemble the ad per impression from a pool of assets. Meta defaults ad sets to its automated audience option, its placements to automatic selection and its budgets, in many flows, to campaign-level distribution, and its creative options apply enhancements unless they are switched off. Manual control still exists on both, but it is now the thing an advertiser opts into.

This matters for how the work is understood. The adjustments that used to fill an account manager's week, moving bids, adding and removing audiences, choosing placements, are now made by the platform from signals the advertiser never sees. What the advertiser sets instead is the frame: which conversions count and what they are worth, what the campaign may not do, what assets it may use, how it is structured and how much it may spend. Those decisions determine what the automation can achieve, and most of what goes wrong in automated accounts traces back to one of them rather than to the automation itself.

It also matters for how the platforms are read. Both document their automation in enough detail to know what is a suggestion and what is a constraint, what the system needs to work and what it does when it lacks it. Both leave undocumented how the predictions are made, how often expansions happen and what any feature does in a particular account. This guide stays on the documented side of that line and says so where it reaches it.

Bidding

Google documents Smart Bidding as auction-time bidding that uses signals present at the moment of the auction to predict the likelihood and value of a conversion and set a bid accordingly. The advertiser picks what to optimise towards and optionally a target to hold: the documented strategies include maximising conversions, maximising conversion value, and the targeted versions of each, which hold a target cost per action (CPA) or a target return on ad spend (ROAS). Google documents a learning period after a strategy is created or significantly changed, and documents that the strategy depends on the conversion actions included in the campaign's goals and, for value strategies, on the values reported with them. The Smart Bidding strategies guide covers each strategy, the data it needs and how to change targets without restarting learning.

Meta documents a parallel set: a default strategy that aims for the most results the budget allows, a cost-per-result goal, a return-on-ad-spend goal and a bid cap, together with a learning phase during which delivery is less stable and after which the system has learned enough to settle. Meta documents that significant edits to an ad set restart the learning phase, and the learning phase guide explains what counts as significant.

What both systems share is that they optimise exactly what they are told to and nothing else. A campaign told that every lead is worth the same will find the cheapest leads. A campaign whose tracked conversion fires on a page view will find page views. A campaign that is not told about returns, cancellations or lead quality will treat all conversions as equally good because it has no information to the contrary. Bidding automation is therefore mostly a measurement problem in disguise: the conversion tracking setup and the Meta pixel setup decide what the bidding can optimise, and value reporting decides whether it optimises for the right thing.

Audiences

Audience automation takes two documented forms: expansion beyond what the advertiser chose, and audiences the platform builds from a seed the advertiser supplies.

On Google, optimised targeting is documented for Display and several other campaign types as enabled by default and as looking beyond the selected audience segments to find people likely to convert, with the selected segments acting as a starting point; exclusions are documented as respected while it runs, and it can be turned off. In Performance Max, audience signals are documented as suggestions that help the campaign find converting users faster rather than as targeting. Broad match keywords are documented as designed to work with Smart Bidding, which uses auction-time signals to decide which matched queries to bid on. The display targeting guide covers optimised targeting in detail and the Performance Max structure guide covers signals.

On Meta, the automated audience option is documented as starting from the advertiser's inputs and expanding beyond them where the system predicts better results. Meta documents two kinds of input: controls, which cover location, minimum age, language and excluded custom audiences and are honoured as limits; and suggestions, which cover included custom audiences, lookalikes, detailed targeting, gender and an upper age and can be exceeded. The Advantage+ audience guide sets out the distinction and how to test it against defined audiences.

The seed-based form is the Lookalike Audience: Meta documents lookalike audiences built from a source audience the advertiser supplies, at a size the advertiser chooses, and Google documents lookalike-style segments in some campaign types built from the advertiser's own data. The lookalike audiences guide covers sources and sizing. In both forms the advertiser's real influence is on the inputs, the seed quality and the exclusions, and on reading the breakdowns afterwards to learn who was actually reached.

Creative and assets

Creative automation on the platforms is of two kinds: assembly, where the system combines assets the advertiser supplied, and generation, where it creates or alters assets. Assembly is older and better documented. Google's responsive search ads combine supplied headlines and descriptions per query, with pinning as the documented way to hold a position; responsive display ads combine headlines, descriptions, images and logos per placement; Performance Max asset groups do the same across all inventory. Meta's dynamic creative and its catalogue-driven formats assemble ads from supplied assets and products per person and placement.

Generation is newer and the documentation is moving. Google documents an optional Search setting that creates additional headlines and descriptions from the landing page and existing ads, which the advertiser can review and switch off; auto-generated video and other enhancements in Display, which can be disabled; video generated from other assets in Performance Max when none is supplied; and generative tools in the campaign builder that produce image and text assets from a prompt, shown for review before use. Meta documents enhancements under the Advantage+ creative name, including adjustments to images, templates, music, text improvements, image expansion and background generation, with per-enhancement controls to switch each off. The AI for creative guide covers the brand, rights and review questions these raise; the responsive search ads guide and the responsive display ads guide cover the assembly mechanics.

The advertiser's control here is the asset pool and the enhancement settings. What goes in decides what can be assembled; a weak pool produces weak ads however clever the assembly. The enhancement settings decide what the platform may change, and the documented defaults lean towards on. Both platforms document asset-level reporting that shows which assets are used and how they rate, which is the feedback loop for improving the pool; neither documents how combinations are chosen beyond stating that the system predicts which will perform. The tutorial on reviewing AI enhancements in your ad accounts sets out how to list those settings, decide each one against written brand and claim rules, and look again when campaigns change.

Bundled campaign types

The newest campaign types combine bidding, audience, placement and creative automation behind a goal, a budget and a set of assets. The table lists the documented ones and what each leaves to the advertiser.

Campaign typeWhat it bundles (documented)What the advertiser setsGuide
Performance Max (Google)Placements across all Google inventory, bidding, audience expansion, asset assembly, optional generationGoals and values, budget, asset groups, product feed, audience signals, brand and placement exclusions, URL expansion settingsPerformance Max structure
Demand Gen (Google)Placements across feed-based Google surfaces, bidding, lookalike-style segments, asset assemblyGoals, budget, assets, audiences and seeds, exclusionsGoogle Ads hub
App campaigns (Google)Placements across Google inventory, bidding, targeting and assembly from the app listing and supplied assetsGoals, budget, assets, locations and languagesGoogle Ads hub
Advantage+ sales campaigns (Meta; earlier called Advantage+ shopping)Audience, placements, budget distribution, creative enhancements and catalogue assemblyObjective and conversion, budget, creative and catalogue, country and existing-customer settings, exclusions where availableCatalogue ads
Advantage+ app and leads campaigns (Meta)Audience, placements, budget distribution and creative enhancements for the respective objectiveObjective, budget, creative, app or form, countryCampaign objectives
Advantage+ catalogue ads (Meta)Product selection per person from the catalogue, assembly into formats per placementCatalogue and product sets, creative templates, audiences or automated audienceCatalogue ads

Names and groupings change: Meta has renamed and merged parts of this family and Google has added and retired features over time, so the help centre page for the campaign type is the record of what it currently bundles and which settings it exposes. The common pattern is stable, though. The advertiser's influence runs through the goal, the assets, the feed, the exclusions and the structure, and the campaign decides the rest.

What you still control

The levers that remain are few and decisive, and they are the same across platforms.

What counts and what it is worth. The conversion actions in the campaign's goals, their values, and whether offline outcomes such as qualified leads, returns and repeat purchases are reported back. Every other lever is downstream of this one.

What the campaign may not do. Exclusions of audiences, placements, content, brands and URLs; the controls the platform documents as hard limits; the original targeting options where a boundary has to hold. This is where legal, brand-safety and policy constraints are enforced.

What it may use. The asset pool, the product feed, the landing pages, the enhancement settings. The quality of the inputs sets the ceiling.

How it is structured. How many campaigns, split by goal, margin or market, so that each has a target that makes sense and enough data to learn; the account structure guide and the Meta campaign structure hub cover the principles.

How much, and for how long. Budgets and targets, changed in measured steps so that learning is not restarted, and left alone long enough for the result to be read.

Whether to automate at all. Each feature can be declined. Smart Bidding, expansion, enhancements and bundled types are choices, and the right answer for a campaign with rare conversions, a legally bounded audience or a brand that cannot accept altered creative is sometimes the manual option the platform still offers.

Measuring the trade

Handing decisions to a platform is a trade: effort and often results in exchange for visibility. The platform's own reports show what it did by its own attribution, and that is the wrong instrument for judging whether the trade was worth it, for two reasons. Platform attribution counts conversions the platform touched, including ones that would have happened anyway, and automated campaigns are documented to seek out the people most likely to convert, who are often people already on the way. And the reported cost per result or return will look healthy whenever the campaign is optimising towards a conversion that is easy to produce, whether or not it is the one the business wants.

The measurement that answers the question is comparative. Before automating, record the baseline: the business outcome, not the platform's, over a period long enough to average out noise. Then compare with the automation on, ideally with a holdout: a geography, an audience segment or a time period where the campaign does not run. The incrementality testing guide explains how to design such a test and read it. Where a holdout is not possible, the weaker substitute is to watch the business KPI, such as revenue, qualified pipeline or new customers from the CRM, alongside the platform's figures and to treat divergence between them as the signal.

Illustrative example — not a real result

A regional insurer tests automated bidding against its own records

A regional insurer runs search campaigns for home insurance quotes on manual bidding with a cost per quote request of £38 and, from its CRM, a policy-purchase rate of 22% on those requests. It moves to a target CPA strategy at £35 and after 6 weeks the platform reports £31 per quote request, which looks like a clear win. The CRM tells a different story: the purchase rate on requests from the automated period has fallen to 14%, so the cost per policy has risen from about £173 to about £221. The team imports purchases as an offline conversion, switches the campaign to optimise towards policies with a target of £180, and holds one of its four regions on the old setup for 8 weeks as a comparison. At the end of the test the automated regions produce policies at £168 against £176 in the holdout region, and the team adopts the strategy account-wide with the purchase conversion as the goal.

Common mistakes

  • Optimising towards the wrong conversion. The platform finds what it is told to find; a soft conversion produces soft results that look efficient.
  • Reporting no values. Without values, every conversion is equal and the system favours the cheapest.
  • Leaving expansion on without exclusions. Expansion is documented to respect exclusions; without them it has nothing to respect.
  • Changing targets daily. Each significant change restarts learning; the result can never be read.
  • Judging automation on platform reports. The platform's attribution cannot show what it added. Compare business outcomes, with a holdout where possible.
  • Treating suggestions as targeting. Audience signals and suggested audiences are starting points, not boundaries.
  • Ignoring what the platform generated. Generated and enhanced assets run under the brand's name; see them before they do.
  • Automating a campaign that cannot feed the model. Rare conversions and tiny budgets leave the system guessing; manual control is still offered for a reason.

The applications section places platform automation beside the other uses of AI in marketing and gives a way to judge each by value, review cost and risk.

Frequently asked questions

Is it still worth learning manual campaign management?

Yes, because the automation is only as good as what it is given, and what it is given is set manually: conversion tracking, goals and values, exclusions, assets, structure and budgets. The skill has moved from adjusting bids and audiences by hand to deciding what the system optimises towards and checking what it did. Someone who understands the auction, the conversion path and the reports reads automated campaigns far better than someone who only knows which switches to leave on.

Why do automated campaigns report good results while sales do not move?

Usually because the campaign is optimising towards a conversion that is easy to produce rather than one that matters, or because it is taking credit for conversions that would have happened anyway, such as brand searches and returning customers. Platform reports are built from the platform's own attribution and cannot show either problem. A holdout or geographic test that compares outcomes with and without the campaign is the only way to see what it added, and the incrementality testing guide covers how to run one.

Should expansion features be left on by default?

For campaigns with reliable conversion tracking, enough volume and good exclusions, the documented behaviour of expansion features is to find converters the chosen audience would have missed, and leaving them on is reasonable. For campaigns where the audience is a legal, brand or policy boundary, or where conversions are rare, the expansion will wander and the documented way to stop it is to use the original, constraining options. The decision is per campaign, and it should be made rather than defaulted.

Terms used on this page

Where this page is used

Part of these learning paths

Also related

Sources

  1. Google Ads Help — Google Ads Help
  2. Meta Business Help Centre — Meta

Related