AI Advertising: Paid Media Guide 2026

The short answer

Google and Meta now make most in-auction decisions for you: bids, placements, much of the targeting and, increasingly, how creative assets are combined. That does not make the advertiser less important. It moves the work upstream. The platform optimises towards whatever you tell it counts as success, using whatever creative you give it, within whatever budget and exclusions you set. If those inputs are weak, automation spends efficiently on the wrong thing.

The framework above is the control loop: hypothesis, creative, tracking, optimisation. This guide explains what the platforms automate, the signal work that decides whether that automation helps you, how to use AI for creative within UK ad rules, and how to check profitability honestly.

What the platform decides, and what you still control

Division of labour in automated ad platforms

Area	What the platform automates	What you control	What goes wrong if you do not
Bidding	Bid per auction towards your goal (Smart Bidding on Google; similar systems on Meta)	The goal, the conversion event and, if used, its value	Efficient buying of low-value conversions
Targeting	Finding people likely to convert, often beyond the audiences you suggest	Audience signals, exclusions, geography, customer lists	Spend on existing customers or unsuitable regions
Creative assembly	Mixing headlines, images and videos into ads per placement (for example Performance Max asset groups)	The assets, their claims and how they are grouped by theme	Combinations that are off-message or make claims you cannot support
Budget allocation	Distributing spend across placements and times within a campaign	Campaign budgets, structure and when to change them	Budget drifting to cheap, low-quality inventory
Measurement	Modelling conversions it cannot observe directly	Tag setup, consent, offline outcomes sent back, and your own profit calculation	A platform report that looks healthy while margin falls

The right-hand column is the job now. Most of it concerns the signal.

The conversion signal is the strategy

Google Ads Help says Smart Bidding learns from your conversion data and that inaccurate tracking causes fluctuations in volume, cost per acquisition and ROAS. Meta's Conversions API exists for the same reason: to send events from your server or CRM so its systems can optimise and measure. Whatever event you mark as the goal is what the algorithm will find more of.

For ecommerce, the natural goal is a purchase with its value. For lead generation, a form fill is a weak goal, because the cheapest form fills often come from the least interested people. Three steps move the signal closer to a real outcome:

1. • Send qualified outcomes back. Google's enhanced conversions for leads uses hashed details the person submitted (such as an email address) to match leads later qualified in your CRM back to the ads that produced them. Meta's Conversions API can carry CRM events in a similar way. The platform can then optimise towards leads your sales team accepted, not just any submission.
2. • Give conversions a value. Google's conversion value rules let you adjust the value of conversions by conditions such as location or audience, so bidding can favour the leads worth more. A simple starting value is close rate × average gross profit: if 10% of qualified leads become customers worth £5,000 in gross profit, each qualified lead is worth about £500.
3. • Get consent right. Google requires advertisers using its measurement and personalisation features to collect consent from users in the UK and EEA and pass consent signals through consent mode. Without it, both measurement and remarketing degrade, and you may be in breach of Google's policy as well as UK rules.

Then leave the system alone long enough to learn. Google's guidance is to allow at least two weeks without changes for an initial learning period, and that recalibration after a change can take up to around 50 conversion events or three conversion cycles. Constant tweaking resets that learning; a weekly rhythm with clear decision rules works better than daily intervention.

Creative: AI for volume, people for claims

Automated campaigns reward a varied supply of assets, and AI is well suited to producing it: headline variants, copy angles for different audiences, resized images, script drafts. Performance Max, for example, assembles ads from asset groups, and Google recommends separating asset groups by theme, product category or audience so combinations stay coherent.

Two constraints matter more than the tooling.

• • Test ideas, not word swaps. Twenty AI-written headlines that all say "save time" test one idea. Structure variants around distinct hypotheses: the problem, the outcome, the proof, the objection. The decision table above describes when a winning idea earns more budget.
• • Claims must be true regardless of how they were made. The ASA and CAP state that ads created with AI are subject to the same rules as any other ad, and that disclosing AI use will not fix an ad whose message is misleading. An AI-generated image that exaggerates a product's effect is a misleading ad, with or without a label.

Check profitability yourself

Platform-reported ROAS is a starting point, not a verdict. It counts revenue the platform attributes to itself, which can include sales that would have happened anyway, and it ignores your margin. Use the calculator with your real gross margin.

Calculator

ROAS and break-even calculator

Check whether ad revenue is profitable once your gross margin is taken into account.

Ad spend (£)Revenue attributed to the ads (£)Gross margin (%)

The starting figures are illustrative round numbers, not AfM prices or client results. Replace them with yours.

ROAS

3.00x

Attributed revenue divided by ad spend

Break-even ROAS

2.50x

1 divided by gross margin: below this, the ads lose money before any other cost

Contribution after ad spend

£1,000

Revenue × margin, minus ad spend

Assumptions

• Attributed revenue depends on the attribution model; platform-reported revenue often overlaps between channels.
• Contribution excludes fees, returns, fulfilment you have not counted in margin, and overheads.

Worked example

Illustrative example: an online retailer spends £10,000 in a month; the platform attributes £32,000 of revenue; gross margin is 40%.

• • ROAS = £32,000 ÷ £10,000 = 3.2.
• • Break-even ROAS = 1 ÷ 40% = 2.5.
• • Contribution after ad spend = (£32,000 × 40%) − £10,000 = £12,800 − £10,000 = £2,800.

On those figures the campaign is profitable. Now suppose a holdout test or a period with ads paused suggests that about a quarter of the attributed revenue came from existing customers who would have bought anyway. Incremental revenue is then £24,000.

• • Incremental ROAS = £24,000 ÷ £10,000 = 2.4, below the 2.5 break-even.
• • Contribution = (£24,000 × 40%) − £10,000 = £9,600 − £10,000 = −£400.

Enter £10,000, £32,000 and 40% in the calculator above to reproduce the first result (3.20x, 2.50x, £2,800), then change revenue to £24,000 to see the second (2.40x and −£400).

The same campaign, judged on incremental revenue, loses money. The 25% figure is an assumption for illustration; your own figure has to come from a test, such as a geographic holdout or excluding existing customers and watching what changes. Until you have one, treat the platform's number as an upper bound.

For lead generation, replace revenue with the value of qualified leads (from the value rule above) and the same logic applies: judge spend on the outcome, not the form fill.

A weekly review that keeps control

A short, fixed agenda prevents both neglect and over-tinkering.

1. • Signal health. Did conversion counts in the platform match your CRM or shop within a sensible tolerance? A gap usually means broken tracking, which invalidates everything else.
2. • Quality. What share of last week's leads or orders were qualified, or from new customers?
3. • Profit. Contribution after ad spend, using real margin.
4. • Creative. Which ideas (not variants) are winning, and which are fatiguing?
5. • Decisions. Scale, hold or cut, per the decision table above, with one new test queued and written down.

Budget changes are best made in considered steps at this meeting rather than reactively during the week, so the platform's learning is not repeatedly reset.

A published AfM example

AfM's published record for an anonymised global technology platform describes about 11 months of paid media (Q4 2024 to Q3 2025) that Jakub ran before founding AfM: LinkedIn only at first, then Meta as the main channel from Q3 2025, with creative rotation using AI-assisted copy in German and English, segmentation by audience and country, and quarterly strategy reviews. Cost per lead fell from $270+ in the engagement's own first quarter, on LinkedIn only, to $8.86 blended across Meta and LinkedIn in the last, a 97% reduction, across 1,700+ leads and $110K+ of managed spend. The record is candid about the cause: most of the fall came from the move to Meta, since LinkedIn alone still cost $201 a lead in the final quarter, and it holds no lead-quality figures. The case record has the detail. It is one engagement in one market; it shows the method, and why a signal and channel review matters, not a figure any other account should expect.

AfM runs paid media for clients as part of its managed service, with the client approving budgets and creative. If you want a view of your own account's setup, the free written breakdown works from your public ads and site and any figures you choose to share; it cannot see inside your ad account. For budget-level decisions between channels, Facebook Ads versus Google Ads for B2B goes further.

FAQ

Should I use Performance Max or Meta Advantage+ campaigns?

They can work well when the conversion signal is strong and you supply good, well-grouped assets. They are riskier when your conversion event is a low-quality form fill or when you need tight control over where brand or existing-customer spend goes. Fix the signal first, then test.

How much budget do automated bidding strategies need?

Enough to generate a steady flow of conversions. Google describes calibration after a change as taking up to around 50 conversion events or three conversion cycles. If a campaign produces only a handful of conversions a month, consider optimising towards an earlier, still meaningful event, or consolidating campaigns.

Do I need to label AI-generated ads?

There is no blanket UK labelling requirement in the ASA and CAP guidance cited here. The question is whether the audience would be misled; if so, a label will not cure a misleading message. Ads made with AI are judged by the same rules as any other ad.

Why does my platform report more conversions than my CRM?

Common causes are modelled conversions, different attribution windows, duplicate tags, and conversions counted for people your CRM treats as existing customers. Reconcile a sample week line by line before trusting either number.