Meta Andromeda and Google AI Max Explained

The short answer: On Meta and on Google, the system now works out who should see your ad from the ad itself, your landing page and your conversion data. The audience boxes you used to tick matter much less. That leaves you three real levers: how many genuinely different ideas you put in front of the system, how accurately you tell it what a sale is, and how little you fragment your budget. Spend your time there, not on interest lists and bid tweaks.

This guide explains the two changes in plain terms, then gives you a creative matrix, an account structure, a testing budget rule and a tracking checklist.

What actually changed

On Meta: Andromeda. Meta's engineering team describes Andromeda as the retrieval stage of its ads system. It is the first step, which cuts tens of millions of candidate ads down to a few thousand before the ranking models choose what you see. Meta says the new system has 10,000 times the model capacity of the old one and improved retrieval recall by 6%. In plain words, it can now consider far more ads for each person and match them more precisely. Meta built it partly because advertisers, helped by generative tools, are uploading far more creative than before.

On Google: AI Max for Search. Google announced AI Max in May 2025 and says it left beta in April 2026. It adds three things to a Search campaign:

Feature	What it does	The control you keep
Search term matching	Shows your ad on relevant searches beyond your keyword list, using your keywords, ads and landing pages as signals	Negative keywords, brand inclusions and exclusions
Text customization	Writes headlines and descriptions to fit the search, from your site and existing ads	Text guidelines, pinning, removing assets you dislike
Final URL expansion	Sends the click to the page on your site it judges most relevant	URL inclusions and exclusions

Google's figure at launch was 14% more conversions at a similar cost for advertisers who switch it on. That is the seller's own number, so test it in your account before believing it.

AI Max is also becoming harder to avoid. Google has said campaigns using automatically created assets or the campaign-level broad match setting are upgraded to AI Max from September 2026, and Dynamic Search Ads follow from February 2027.

What this means: your ad and your landing page are now the brief you give the machine. A vague ad gets shown to a vague audience.

Build a creative matrix

The common mistake is to make ten versions of one idea: same video, different first line. To a system that reads the creative, those are close to one ad. What it needs is different reasons to buy, shown in different ways.

Plan with a grid. Angles down the side, formats across the top. Here it is filled in for a made-up product: a ₹799 insulated steel bottle.

Angle	Creator-style Reel	Product demo Reel	Static image
Problem	"My water is warm by 11 am on site visits"	Ordinary bottle against ours, left in a parked car	"Still cold at 6 pm." Bottle beside a temperature reading
Proof	A customer shows the bottle after six months of daily use	Ice test filmed over a working day, with a clock in frame	A real review, with the star rating
Offer	"Two for ₹1,399 this week, here is why I bought two"	Unboxing the two-pack	Price, offer and end date, nothing else
Comparison	"I tried three bottles under ₹1,000"	Leak test: three bottles upside down in a bag	A three-row table: ours against plastic and against a costlier brand

Twelve cells. You do not need all twelve at once. Start with six that cover all four angles and at least two formats, then fill the rest as you learn. Change the person on camera and the setting between Reels too: office, gym, school run. Each one tells the system about a different buyer.

Good statics depend on clear product shots, which is where product photography pays for itself. The Reels come out of a planned content creation shoot, not a scramble before each campaign.

Keep the account structure lean

Every ad set has to learn on its own. Meta's long-standing guidance is that an ad set needs roughly 50 of its optimization events in a week to settle. Split a budget across ten ad sets and none of them gets there.

Layer	Meta	Google
Main campaign	One sales campaign, broad audience, one or two ad sets holding your proven creative	Search campaign with AI Max on, grouped by theme, not one ad group per keyword
Testing	One ad set for new concepts, on a fixed share of budget	An experiment comparing AI Max on and off before a full switch
Brand	Not usually needed	A separate brand Search campaign, so brand clicks do not flatter the rest
Shopping and catalogue	A catalogue ad set inside the main campaign	Performance Max or Shopping, read with the channel report
Split further only when	Countries, product lines with very different margins, or separate goals	The same, plus separate budgets you must protect

Guardrails still exist, so use them: location settings, brand exclusions, negative keywords, URL exclusions for pages that should never be a landing page (careers, blog, returns policy).

A testing budget rule, with numbers

Our rule of thumb, not a platform rule:

Put about 20% of monthly spend into testing new concepts.
Give each concept enough to buy about 10 conversions at your target cost before you judge it.

So: concepts you can test in a month = testing budget ÷ (10 × target cost per order)

A worked example with round, made-up numbers:

Monthly Meta budget: ₹3,00,000
Target cost per order: ₹500
Testing budget: 20% of ₹3,00,000 = ₹60,000
Budget per concept: 10 × ₹500 = ₹5,000
Concepts you can test: 60,000 ÷ 5,000 = 12 a month, or three a week

Now check the main campaign against the learning guidance. The remaining ₹2,40,000 is about ₹56,000 a week. At ₹500 an order that is about 112 orders a week, enough for two ad sets at 50 each, and not enough for five.

How to judge a concept once it has had its ₹5,000:

Result	Decision
Cost per order at or below target	Move it into the main ad set
Up to 30% above target	Keep the angle, remake the first three seconds, test once more
Far above target, or it would not spend	Stop it, and note which angle failed

With a small budget the same sum still works. At ₹60,000 a month and the same ₹500 target, the testing budget is ₹12,000, which is two concepts a month. That is fine. Two honest tests beat ten starved ones.

Tracking checklist

The system optimises towards whatever you report. Report the wrong thing and it will find you more of the wrong thing.

Meta Pixel and Conversions API both running, sending the same events from browser and server.
Duplicates removed. Check in Events Manager that browser and server events are being deduplicated, or every sale is counted twice.
Enhanced conversions on in Google Ads, so sales can be matched using hashed customer details.
Optimise for the real outcome. Purchases for a store. For lead businesses, send back which leads were qualified, not every form fill.
Purchase value passed correctly, net of discounts, in rupees.
One primary conversion per campaign. Not purchases and add-to-carts mixed together.
Landing pages that match the ad. With final URL expansion on, any page on your site can become a landing page. Slow or thin pages now cost you directly, which makes website development part of ad performance.

Four ways brands get this wrong

Counting edits as new creative. A new caption on the same Reel is not a new idea.

Keeping the old structure. Fifteen interest-based ad sets on ₹1,000 a day each will never leave learning.

Switching on AI Max and walking away. Read the search terms report weekly for the first month. Add negatives. Exclude URLs.

Judging on the platform's own return figure. Each platform claims every sale it touched. Compare total ad spend with total revenue from your store or CRM, month by month, and trust that number.

Where this fits

This is how we run performance marketing: a lean account, clean conversion data and a steady supply of new angles. The supply is the hard part, which is why media buying and content creation are planned together. And once an ad wins the click, the cheapest follow-up is often a message. See the maths on WhatsApp marketing costs.