Quick answer: what are AI ads and how do you run them well?
AI ads are campaigns where the ad platform's AI decides who sees the ad, where it runs, how budget moves and often which creative version shows. You run them well by sending accurate conversion data, giving the AI enough budget and creative to learn, changing things less often, and judging results on profit, not platform ROAS.
What "AI ads" means in practice
The phrase covers three different things, and mixing them up is where most confusion starts. For the basics of how these ads are built, see our explainer on what artificial intelligence ads are. This guide is about running them.
Layer 1: AI that buys the media
The platform decides who sees each ad, on which placement, at what bid. On Meta this is the delivery system behind Advantage+ settings. On Google it is Smart Bidding and campaign types like Performance Max, which Google describes as a goal-based campaign that reaches all Google Ads inventory from one campaign.
Layer 2: AI that changes the creative
The platform adapts or generates versions of your ad: new crops, background changes, text variations and, on Google, generated headlines and images inside an asset group. You supply the base material and approve what runs.
Layer 3: AI that manages the account
Third-party AI ad managers and in-platform rules read results every day and pause, scale or flag campaigns. They sit on top of the first two layers and act on what those layers produce.
Why the difference matters
Each layer fails in a different way. Media-buying AI fails when it gets bad data. Creative AI fails when the base material is weak or off-brand. Management AI fails when its rules are not tied to your real margins. Knowing which layer is struggling tells you what to fix.
How the learning loop works behind every campaign
Every AI ad system on Meta and Google runs the same basic loop. Once you see it, most "the AI is broken" moments make sense.
- You set a goal and a budget. For example, purchases at the best cost, or a target return.
- The platform shows ads widely at first. It tests audiences, placements and creative versions to find early signals.
- Conversions come back as data. Pixel, server events and offline uploads tell the system who bought.
- The system shifts delivery toward what converts. Budget, bids and creative weight move toward the best performers.
- It keeps exploring a little. A share of delivery keeps testing so results do not go stale.
The learning phase
Meta says an ad set needs about 50 results in the week after its last significant edit to leave the learning phase. Until then, results swing more and cost more. Big budget changes, new creative and new targeting all count as significant edits and restart the clock.
What this means for you
Batch your changes. Make one meaningful edit, wait a week, then judge. Brands that edit daily keep their campaigns in permanent learning and then blame the AI for unstable results.
If your budget cannot produce about 50 purchases a week in one ad set, optimise for an earlier event such as add to cart, or combine ad sets so data is not split too thin.
Automated products on Meta and Google compared
Both platforms push you toward their automated products. The table shows what the AI decides in each and what stays in your hands. For depth on each platform, read our guides to Advantage+ creative on Meta and Performance Max on Google.
| Product | Platform | What the AI decides | What you still control |
|---|---|---|---|
| Advantage+ sales campaigns | Meta | Audience, placements, budget split across ads | Goal, budget, creative, exclusions of existing customers |
| Advantage+ creative | Meta | Crops, enhancements, text and media variations per viewer | Base creative and which enhancements are switched on |
| Performance Max | Channel mix, bids, audiences, asset combinations | Goal, budget, assets, audience signals, brand exclusions | |
| Smart Bidding on Search | Bid for each auction | Keywords, ad copy, target CPA or ROAS, budget |
Which to start with
For most online stores, the starting pair is one Meta sales campaign with Advantage+ audience and one Google Search campaign on your brand and best product terms. Add Performance Max once conversion tracking is clean and you have product feed data in Merchant Center.
Signals you can still give the AI
Automation does not mean silence. On Google you can add audience signals to Performance Max, such as your customer list or people who searched your brand, to show the AI where to start. On Meta you can exclude existing customers from prospecting campaigns, set a minimum age where your product requires it and choose which Advantage+ creative enhancements are allowed. These are hints and guardrails, not hard targeting, but they shorten the time the AI spends exploring the wrong people.
When manual still wins
Manual control still helps for small tests of a new market, for strict brand-safety needs, and for accounts with too little conversion data for the AI to learn from.
The four inputs the AI needs from you
You no longer control who sees each ad. You do control what the AI learns from. These four inputs decide most of your results.
1. Clean conversion data
Send purchases from both the browser and your server. Meta recommends a redundant setup with deduplication so each purchase counts once. On Google, enhanced conversions send hashed first-party data with each conversion so more of them can be matched. Our guide on what AI needs from your pixel, CAPI and store covers the setup.
2. Enough budget per campaign
Spread a small budget across many campaigns and none of them learns. Fewer, larger campaigns usually beat many small ones.
3. Creative variety
The AI can only choose between the ads you give it. Give it different angles, not five copies of one idea in different colours: a problem-solution video, a customer review, a product close-up and an offer-led static.
4. A goal that matches the business
If you optimise for purchases but half your orders are cash on delivery that never arrive, the AI will chase the wrong buyers. Send the conversion that best matches money in the bank.
A quick input audit
Place a test order and check it shows once, with the right value, in both platforms. Count how many genuinely different creatives are live. Count how many campaigns share your budget. Those three checks find most problems.
Creative that works with automated delivery
When the platform chooses the audience, your creative does the targeting. A video about joint pain finds people with joint pain; a video about gifting finds gift buyers. That makes creative the biggest lever you still hold.
Think in angles, not variations
An angle is a different reason to buy: a problem the product solves, a result, a comparison, social proof, a price or offer, or a founder story. Five colour versions of one image are one angle. Five angles give the AI five different groups of people to find.
Mix formats
Give each campaign a short vertical video, a static image and, where it fits, a carousel. Placements differ, and the AI can only use the formats you supply. Vertical video matters most for Reels and Stories inventory.
Refresh on a schedule
Watch frequency and click-through rate week by week. When frequency climbs and click-through rate falls for two weeks in a row, the same people are seeing the ad too often. Replace the weakest creatives with new angles rather than editing the winners.
Checking AI-generated versions
Generated crops, backgrounds and text can change claims, prices or product details. Preview every enhancement before it runs, switch off the ones that change your product's look, and never let generated text make health, income or before-and-after claims you cannot prove.
What a good creative brief contains
One angle, one audience problem, the proof you will show, the offer and the first three seconds written out. A brief like this takes ten minutes and saves days of guessing, whether a designer, a creator or an AI tool makes the ad.
Launch new angles inside the existing campaign rather than a new one, so they compete for the same budget and the AI does not restart learning from zero.
A 30-day launch plan
This plan suits a store spending roughly ₹1 lakh to ₹10 lakh a month on ads. Scale the numbers to your budget, but keep the rhythm of one meaningful change per week.
Week 1: fix the data
Check the pixel, server events and Google conversion tags. Remove duplicate purchase events. Decide which event each campaign optimises for.
Week 2: launch a simple structure
One Meta sales campaign with broad targeting and four to six genuinely different creatives. One Google Search campaign for brand and top product terms. Leave both alone for seven days.
Week 3: refresh creative, not targeting
Pause the weakest one or two creatives by cost per purchase and add two new angles. Do not touch audiences or budgets in the same week.
Week 4: move budget on evidence
If one campaign is clearly above break-even ROAS, raise its budget by 15 to 20 percent. If both are below, fix the offer or landing page before spending more.
What to write down each week
Spend, purchases, cost per purchase, ROAS against break-even, and the one change you made. After four weeks you will know which change actually moved results.
Keep a change log
One line per change, with the date and the reason. When results shift, you can see whether the shift followed your edit or happened on its own, for example during a sale or a festival week. A plain spreadsheet is enough; what matters is that every person touching the account writes in it.
Judging results on profit, not platform ROAS
Platforms report the revenue they can attribute to their own ads. That number is useful for comparing campaigns, but it is not your profit. Two adjustments matter most: your margin and the share of orders that actually complete.
Worked example: a fashion store
The order economics
Average order ₹1,800. Product cost ₹600, shipping ₹120, payment fees and GST ₹270, returns and failed cash on delivery ₹180. That leaves ₹630 per order before ad cost, a 35% margin.
Break-even and real ROAS
Meta reports a ROAS of 3.2 for the month. But 15% of the orders it counted were cancelled or returned, so the real ROAS is 3.2 × 0.85 = 2.72.
Break-even ROAS = 1 ÷ margin before ad cost
1 ÷ 0.35 = 2.86. The real ROAS of 2.72 is below 2.86, so this campaign loses a little money on every order, even though the dashboard shows 3.2.
The fix is not more budget. It is a better offer, fewer returns, prepaid incentives or a higher average order, then let the AI find buyers again.
Three numbers to check every month
New customer share: how many purchases came from people who had never bought before. Order completion: the share of attributed orders that were paid and not returned. Profit after ad spend: margin per order times completed orders, minus ad spend. If those three move the right way, the AI is working for you, whatever the dashboard ratio says.
For the full cost picture, including tools and people, see what AI advertising really costs.
What most guides get wrong
They treat AI as a switch
Turning on Advantage+ or Performance Max is the start, not the result. The AI amplifies whatever data and creative you give it, good or bad.
They skip the data layer
Guides jump straight to creative tips. If the pixel misses or double-counts purchases, every downstream decision the AI makes is built on wrong numbers.
They measure the wrong thing
Platform ROAS rises easily when the AI shifts spend to people who would have bought anyway. Check new customer share, real order completion and profit after ad spend, not only the dashboard ratio.
They ignore how often you edit
Daily edits keep campaigns in learning. The most common cause of "unstable AI results" is the person, not the algorithm.
Fix data, structure and creative first. A paid AI ad manager helps most once those basics are right; compare options in our list of AI advertising platforms.
Read next in this series
- AI Ads in India: Costs, Tools and Case Studies for 2026
- Can AI Run Your Ads Fully? What It Handles Alone, Where It Fails and What to Keep Human
- AI Ads and Privacy: What Changed After iOS Tracking Rules and the Cookie Debate
- AI Ads Targeting: Broad vs Interest Targeting in 2026
- AI Ads Budget: How Much Should You Spend, and How Fast Should You Scale?
- AI Ads for E-commerce: A Practical Playbook From Product Feed to Profit
- AI Ads for Growing Brands: Getting Started at ₹10 Lakh+ a Month
- AI Ads vs Traditional Ads: Results Compared, With the Fine Print
- How AI Ads Work on Meta and Google: The Auction, the Signals and the Learning
Key takeaways
- On Meta and Google, AI now makes most delivery decisions. Your job moves from pulling levers to feeding good inputs.
- The four inputs that matter most: conversion data, budget, creative variety and a clear goal.
- Frequent edits reset learning. Plan changes weekly, not hourly.
- Platform-reported ROAS is not profit. Measure against break-even ROAS from your real margin.
- A 30-day plan with one change per week beats a month of daily tinkering.
Ai ads: frequently asked questions
What are AI ads?
Are AI ads better than manual campaigns?
How much budget do AI ads need?
How long should I wait before judging an AI ad campaign?
Why does my AI campaign show good ROAS but no profit?
Should a small brand use AI ads or manual campaigns?
Do AI ads replace the need for creative?
Sources
- Google Ads Help: About Performance Max campaigns
- Meta Business Help: About the learning phase
- Meta Business Help: Conversions API best practices
- Google Ads Help: About enhanced conversions
Let ScaleWin read your ad data every day
ScaleWin checks your Meta and Google Ads daily, judges every ad on profit and tells you exactly what to change.
Start free, no card



