Both platforms now use AI at every step. They predict who is likely to take the action you want, bid in each auction on that prediction, choose where the ad appears, and assemble or even generate the creative. You supply the goal, budget, conversion data and raw assets; the machine decides who, where, when and which version.
The catch that most owners miss: the AI optimises relentlessly for whatever signal you give it. If you tell Meta a "conversion" is someone opening your booking page, it will find people who open booking pages, not people who book. If Google counts a click on your phone number as a win, it will buy clicks from people who tap numbers by accident. Understanding how the machinery works matters mostly because it shows you where your inputs control the outcome.
The four decisions the platforms now make for you
Ten years ago, running ads meant making four decisions by hand. Now the platforms make them, using models trained on enormous numbers of past auctions.
- Who sees the ad. Instead of you picking interests and age bands, the system predicts each person's likelihood of acting and treats your targeting as a hint.
- How much to bid. Each auction gets its own bid, based on the predicted value of showing the ad to that person at that moment.
- Where it appears. Feeds, stories, search results, video, email tabs, maps: the system shifts budget to wherever results are cheapest.
- Which version. Headlines, images, video and text are mixed, matched, cropped and sometimes generated, then shown in whichever combination performs.
Your job has moved from operating those levers to feeding the system good inputs and checking its work.
Inside Meta: how one ad reaches one person
When someone opens Instagram or Facebook, Meta has a fraction of a second to choose which ads to show from everything that could run. It does this in stages.
- Retrieval. Meta's engineering team described this stage in December 2024, when it introduced a system called Andromeda: it narrows tens of millions of candidate ads to a few thousand that are relevant to this person. Andromeda runs on specialised hardware and was built to cope with the explosion in the number of ad variations advertisers now upload.
- Prediction and ranking. For each remaining ad, models estimate how likely this person is to take the action the advertiser optimises for: a purchase, a lead form, a message.
- Auction. Meta ranks ads by what it calls total value: roughly your bid multiplied by the estimated action rate, plus a measure of ad quality. A small business with a relevant, well-received ad can beat a bigger bidder whose ad people hide or ignore.
- Delivery and learning. The result (did they act?) feeds back into the models.
On top of that sits Advantage+, Meta's name for its automation features. Meta describes two kinds: end-to-end campaigns that apply AI across audience, placements and budget together (Advantage+ sales, app and leads campaigns), and single-step options you can switch on individually:
- Advantage+ audience. Your age, interest and lookalike choices become suggestions. You keep firm controls such as location and a minimum age, but the system can go beyond your suggestions when it predicts better results.
- Advantage+ placements. Ads run across Facebook, Instagram, Messenger and Meta's partner network, with spend moving to what works.
- Advantage+ creative. The system can adjust brightness, crop images, add music, generate text variations or backgrounds, and choose which variation each person sees. Individual enhancements can usually be switched off.
- Budget sharing. Spend moves between ad sets towards the ones performing best.
Creative mixing is easiest to understand with an example. An illustrative optician uploads five headlines, four images and two short videos to one campaign. The system can assemble dozens of combinations and learns which work for which people. In the first week, the best-performing combination paired the headline "Children's eye tests, no wait" with a photo of a child choosing frames, which made sense. The second-best paired "Designer frames from $89" with the same child photo, which didn't match the offer at all. Nothing was wrong technically; the system simply found that the photo got attention. The fix was to split the assets into two ad sets, one for children's tests and one for frames, so the machine could only mix pieces that belong together.
Meta publishes its own headline results for these campaigns, such as a 20% better cost per acquisition for Advantage+ sales campaigns. Treat vendor averages as a reason to test, not a forecast. A local business with a small audience and few conversions often sees very different numbers. Whether it suits you is covered in should a small business use Advantage+.
Inside Google: bids, matching and self-directing campaigns
Google's AI works in three layers that stack on each other.
Smart Bidding sets the price of every auction
Smart Bidding strategies (Maximise conversions, Maximise conversion value, Target CPA, which aims for a cost per acquisition, and Target ROAS, which aims for a return on ad spend) set a separate bid for each search. They use signals such as device, location, time of day, language and whether the person has visited your site before. A search for "emergency glasses repair" at 5pm on a phone might get a much higher bid than the same words typed on a laptop at 2am, because the model predicts different outcomes.
Matching decides which searches you appear on
Keywords are no longer an exact list. Broad match lets Google show your ad on searches it judges related in meaning. AI Max for Search campaigns goes further with three features: search term matching (finding queries from your keywords, ads and landing pages), text customisation (generating headline and description variations from your assets and site), and final URL expansion (sending people to whichever page on your site it predicts will work best). Google began automatically upgrading campaigns that use automatically created assets and campaign-level broad match to AI Max in September 2026; for Dynamic Search Ads, Google moved the automatic upgrade to February 2027. AI Max includes controls such as brand inclusions and exclusions and URL exclusions, and the search terms report now labels AI Max matches so you can see what it bought.
Campaign types that choose their own channels
Performance Max runs one campaign across all of Google's inventory (Search, YouTube, Display, Discover, Gmail and Maps) from a single budget, mixing your images, text, video and product feed into ads for each place. Demand Gen focuses on visual placements such as YouTube and Discover. Both hand most channel decisions to the system. The mistakes small businesses make with Performance Max are serious enough to have their own list: see nine costly Performance Max mistakes.
What the machine learns from, and why tracking comes first
Every one of these systems learns from conversions: the actions you tell the platform count as success. That makes conversion setup the most important decision you make, more than any audience or creative choice.
An illustrative optician shows why. It set up Google Ads with "click on phone number" and "visit to booking page" as its conversions. Within three weeks the campaign reported 140 conversions at under $3 each. Actual booked eye tests from ads that month: nine. The system had done exactly what it was told. It found people who tap phone numbers and open pages, many of them on mobile browsing idly, and bid for more of them. When the optician changed the primary conversion to "completed online booking" and imported phone bookings from its diary, reported conversions dropped to around 20 a month, each costing more, but booked tests rose.
The signals worth giving the platforms, roughly in order of value:
- The real outcome: a purchase, a completed booking, a qualified enquiry form.
- A value for each outcome where values differ: a $400 varifocal order is worth more than a $30 lens-cleaning kit.
- Offline results fed back in: which leads became customers, uploaded from your booking system or CRM.
- Customer lists, uploaded under the platform's terms and your privacy notice, so the system can recognise existing customers.
If you only fix one thing before spending, make it this; the setup steps are in set up conversion tracking before you let AI spend your ad budget.
The learning phase in plain numbers
Automated campaigns start out uncertain and improve as conversions arrive. Meta says an ad set generally leaves its learning phase after about 50 optimisation events in the week following its last significant edit. Google recommends running a new Performance Max campaign for at least six weeks and avoiding frequent changes to budget, bid strategy or status in that period, because changes can reset learning.
Now the sum a small business has to do. An illustrative nail salon spends $20 a day on Meta, optimising for completed bookings, at about $25 per booking. That's $140 a week, or around five or six bookings a week. At that volume, the ad set will never reach 50 events a week. It will run in "learning limited" indefinitely, and results will swing week to week.
Options, none perfect:
- Consolidate. One campaign and one ad set rather than four, so all conversions feed one learner.
- Optimise for a higher-volume action that still predicts bookings, such as starting the booking form, while tracking completed bookings separately.
- Accept noise and judge monthly, not daily, against your own booking numbers.
- Raise the budget for a defined test period if the margins allow it.
How much to spend before judging results is worked through in how much to spend on AI-run ads before judging them.
A garden centre's spring campaigns on both platforms
To see the machinery work together, follow an illustrative garden centre through its spring push. It has a website shop for click-and-collect plant orders and a budget of $1,200 a month for eight weeks.
Setup. Conversions: completed online orders with order value, plus imported in-store redemptions of an ad-only voucher code. Google gets $700 a month: a Search campaign on Maximise conversion value with broad match on eight core keywords and brand exclusions set, plus a Performance Max campaign using the shop's product feed. Meta gets $500 a month: one Advantage+ sales campaign using the product catalogue and twelve images from the centre's own photos, with generated backgrounds switched off.
Weeks 1 to 2. Results jump around. Performance Max spends heavily on Display and reports low-value orders. Meta finds a surprising audience of new homeowners buying starter herb kits. The manager resists changing anything except adding negative keywords for "jobs" and "wholesale".
Weeks 3 to 6. The Search campaign settles at about $12 cost per order, with an average order of $54. Meta settles at about $15 per order, average $38. Performance Max, after the channel performance report shows most of its Display spend producing nothing, has its asset groups rebuilt around three themes (bedding, fruit trees, garden furniture) and improves.
Weeks 7 to 8, the reality check. Platform reports claim 118 orders in total. The shop's own system shows 96 ad-attributed orders, because both platforms claimed credit for some of the same customers. The owner judges each channel on the shop's numbers, not the platforms', and moves $150 a month from Performance Max to Search for the summer.
The lesson isn't that one platform won. It's that the AI worked better once it had real conversions and values to learn from, and that the owner's own sales data, not the platforms' dashboards, made the final decision.
What you still control on each platform
| Control | Meta | |
|---|---|---|
| Budget and schedule | Yes, at campaign or ad set level | Yes, per campaign |
| What counts as a conversion | Yes: the optimisation event | Yes: primary conversion actions and goals |
| Where you advertise geographically | Yes, a firm control even with Advantage+ audience | Yes, including presence versus interest settings |
| Who is excluded | Some exclusions, such as existing customers in certain campaign types | Negative keywords, brand exclusions, customer lists, placement exclusions |
| Creative inputs | Your images, video and text; enhancements can be switched off | Your assets; automatically created assets and video enhancements can be limited |
| Landing pages | Yes | Yes, with URL exclusions when URL expansion is on |
| Where ads appear | Placement choices and brand-safety controls | Limited in Performance Max; more control in Search |
Settings and menu names change often on both platforms, so check the current help pages when you set a campaign up.
Where the automation serves the platform as much as you
The platforms' AI and your interests overlap a great deal, but not completely. Four misconceptions cost small businesses money.
- "The recommendations tab knows best." Google's recommendations and optimisation score measure how much of Google's suggested setup you've adopted, not whether your campaign is profitable. Recommendations to raise budgets or broaden matching can be sensible or not. Check whether auto-apply is switched on in your account, and switch off anything you haven't consciously chosen.
- "More conversions reported means more customers." Automated campaigns are good at finding people who were going to buy anyway, especially people searching your business name. Brand exclusions and new-customer goals exist for this reason.
- "Generated creative is free creative." Automatic text and image variations can drift off-brand or claim things you don't offer. Review what's being shown, not only what you uploaded.
- "Broad is always better." Broad targeting works when there are enough conversions to learn from. With five conversions a week, the system is guessing, and it guesses with your money.
A useful habit is a 20-minute monthly review: check search terms, placements and the creative combinations being served; compare platform-reported conversions with your own sales; and look for any setting that changed without you. If you'd like to run the campaigns yourself, managing Google Ads without an agency covers the routine, and keeping your brand safe when AI places ads covers the exclusions to set.
Further reads
- How Much Does AI Marketing Cost a Small Business Each Month? — Where ad spend sits in a small business's monthly AI marketing budget.
- 12 AI Marketing Mistakes Small Businesses Make (and the Fixes) — The wider marketing mistakes that also sink ad budgets.
- How to Write Facebook and Instagram Ad Copy With AI — Write the creative inputs the machine mixes and matches.
- How to Test Ad Variations With AI on a Small Budget — Test creative properly when budgets are tight.
- What Are AI Marketing Agents and Should a Small Business Use One? — Where AI agents that manage campaigns fit, and where they don't.
- How to Measure Whether AI Is Improving Your Marketing Results — Check platform-reported results against real sales.
- How Marketing Agencies Use AI to Automate Client Reporting — A four-layer setup for automated client reports: data pipes, the right reporting tool, a commentary prompt that can't invent wins, and a five-minute check.
- How to Track Competitors' Prices and Offers With AI — Set up page monitors, a newsletter inbox and an AI-built price log so you see rivals' price changes and offers weekly, without drowning in alerts.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Meta Engineering post on Andromeda (December 2024); Meta for Business Advantage+ pages; Google Ads Help on Performance Max, AI Max for Search campaigns and the Dynamic Search Ads upgrade; Google Ads optimisation tips for Performance Max.