Andromeda 2.0 is the nickname for an upgrade Meta announced on October 28, 2026, to the system that decides which ads each person on Facebook and Instagram might see. It changed three things under the hood, and Meta was unusually plain about what you have to do in response: nothing. In its own words, "There's no setting to switch on, no new campaign type to adopt, no migration to plan."
That has not stopped the videos. Within a day of the announcement there were tutorials telling owners to restructure their accounts, rebuild their ad sets, and sign up for a course on the new "2.0 structure." We said the same thing about the first Andromeda that we will say about this one: there is no switch, so there is nothing to rebuild around. We made that case in full in The Algorithm Changed Again. You Probably Should Not Do Anything.
But "do nothing to your account" is not the same as "nothing changed." Something real did change, and it changes what your ads need to look like. Meta's system now reads every ad you run, describes what each one is about, and gives brand-new ads a faster start based on what they say and show. The structure of your account barely matters to that. The ads themselves matter a lot. So this piece explains what the update actually is, in plain English, and then what a small business should do with its ads, using a local swim school as the example the whole way through.
Meta Picks Every Ad You See in Three Steps. Andromeda Is the First One.
Each time a person opens their Facebook or Instagram feed, Meta has to choose a handful of ads to show them out of tens of millions that are running. It cannot score all of those carefully in the split second before the feed loads, so it works in stages, like a hiring process that starts with a stack of a thousand resumes and ends with one hire.
Step one is Andromeda. Meta calls it the retrieval engine. Its job is to take that enormous pool and narrow it to a shortlist worth considering for this one person, right now. If your ad does not make the shortlist, nothing else that happens later can save it. It simply is not in the room.
Step two is ranking, and the brain behind it is a model Meta calls GEM. GEM looks at the shortlist and predicts how likely this person is to click, call, buy or fill out a form. It is trained on how people across Meta's apps actually move toward a purchase, then passes what it learns down to the models that score ads live as each feed loads. This is also where the auction happens, so your bid and budget come in here.
Running through both is GenRecs, the newest piece. It uses large language models, the same family of technology as the AI chatbots, to read each ad's text and images and describe what the ad is about in a kind of shared vocabulary. That description is what lets the system relate a brand-new ad, with no history at all, to ads it already knows well.
None of these is new as a name. Andromeda was introduced in 2024, GEM in early 2025. What the October announcement covers is how each one got better, and the honest version of that story is more interesting than the thumbnails.
The Old Andromeda Only Checked the Popular Shelves. The New One Reads Every Ad.
Meta's own analogy is a good one. The first Andromeda was a concierge who knew your taste, but who mostly checked the shelves they knew best, the popular ones. The new version, in Meta's words, will "read every product label on every shelf, in the same amount of time." Technically, Andromeda now sends each ad to specialized "expert" models and, per Meta, "scores every eligible ad" at over a billion ads per second.
Here is why a small business owner should care. If you have ever loaded five ads into an ad set and watched four of them sit at a few cents of spend while one takes everything, part of the reason was this step. An ad that never made the shortlist never got a chance to prove itself, no matter how good it was. We explained how that plays out in You Put Five Ads in the Ad Set. Meta Only Really Runs One. Meta's claim now is that "more of your ads, creative variations and products are considered for each person."
Considered is not the same as funded. Meta still picks favorites at the ranking step, and a weak ad is still going to lose. What changed is that an ad has a better chance of being looked at on its merits instead of being skipped because it was new.
Now the numbers, in proportion, because this is where the videos get loose. Meta says that in the second quarter of 2026, improvements across its whole ranking system drove 8.3% more ad clicks and 15.7% more conversions on Facebook. That 15.7% is the figure in most thumbnails, and it is not an Andromeda number. Andromeda's own share is much smaller: Meta credits its third-quarter upgrade with 1.37% more conversions on Facebook and 0.24% on Instagram Reels. GEM is the bigger engine, credited with up to an 18.9% conversion improvement since 2025. These are Meta grading its own homework, so treat them as direction, not a promise for your account. But the direction is the point. The part that got bigger is the machine's ability to read and match ads, not anything you control in your settings.
Meta Now Reads Your Ad Before Anyone Clicks It
GenRecs is the piece with the most practical consequence for an owner, so it is worth slowing down on.
Meta describes it this way: large language models "read each ad's text and images and describe the ad in semantic tokens that capture its key attributes." Meta's examples come from retail, things like a product's brand, color, fit and style. Those descriptions form a shared vocabulary, so the system can take an ad that went live this morning and say, in effect, "this looks like these other ads, and I know who responded to those." That is how a new ad gets what Meta calls "a faster track to performing" without weeks of trial and error.
Think about what that means for the ad itself. The system can only describe what is actually in it. If the reason someone should call you lives in your head, or on your website, or in the third paragraph of your caption, it is not doing much work at this stage. If it is in the image, in the headline, in the first line of the text, or said out loud in the video, the system can read it and use it.
Take a local swim school running an ad for lessons. Version one is a stock photo of a pool, the school's logo and "Now enrolling for fall." Accurate and almost useless to a system trying to work out who this is for. Version two shows a nervous five-year-old in the shallow end holding an instructor's hand, with a headline that says "Most of our swimmers started out scared of the water." That second ad tells Meta, without anyone clicking, that it is about young children, fear of water, patient instructors and parents making a decision. Meta can match that to the right parent far more easily than it can match a pool and a logo.
This is the same lesson we taught in July in You Don't Pick Your Meta Audience Anymore. Your Ad Does. What October added is the mechanism behind it, stated by Meta itself, and a stronger version of it: the system is now reading the ad's content directly, not just watching who happens to react.
Ten Versions of One Ad Look Like One Ad to Meta
Here is the trap the new system sets for owners who have heard "make more ads." If Meta describes every ad by what it shows and says, then ten ads that show and say the same thing get described the same way. To the system, they are one ad wearing ten name tags.
This is not our guess. Meta said it directly in December 2025: its ads system "analyzes creatives and groups together ads that share key visual and thematic attributes. When multiple ads look or feel alike, these are seen as variations of the same creative." Meta draws a line between iteration, which might be "two ads with identical visuals, but different text CTAs," and real diversification, which means "two distinctly different pieces of creative." Its advice is to "develop ads that are truly different in look, feel, storyline, and message."
So swapping the call-to-action button for a different one, or the headline color from blue to orange, or the photo for another photo of the same pool, does not give the system a new option. It gives it the same option again. Meta even flags the consequence in Ads Manager. If an ad's cost per result climbs above what your past ads cost, its Delivery column can show Creative limited, and at double the cost, Creative fatigue. Meta's prescribed fix for both is an ad with a "new image or video that is materially different from the original."
What does materially different mean for a small business? Different reason to buy, different moment, or a different worry answered. For the swim school, three genuinely different ads might be:
- The safety reason. A parent's real fear, stated plainly: drowning is a leading cause of death for young children, and lessons are how you lower the risk. Shot at the pool, with a lifeguard stand in frame and an instructor teaching a float.
- The scared kid. The child who cries at bath time, and the first lesson where they put their face in the water. Short video, real kid, real instructor, the small win on camera.
- The schedule and the price. The working parent comparing you with the Y. Saturday morning slots, small groups, the price per lesson spelled out, a quick look at the warm indoor pool.
Those are three ads a parent could tell apart in a second, and so can Meta. Three of those beat fifteen versions of "Now enrolling!" every time. If you want the math on how many ads a small budget can actually support, we laid it out in One Meta Ad Is Too Few to Test. Fifty Is Too Many. And if you are wondering which kind of ad to make first, start with the one that shows the service working, which we covered in Show the Thing Working. That Is the Whole Ad.
Meta Decides Which of Your Ads a Buyer Sees, and in What Order
The third change is the quietest, and it argues against the most common "2.0 rebuild" advice.
Meta's ranking system has, since at least early 2025, used what it calls sequence learning: modeling "the sequence of actions a person takes before and after seeing an ad," in order "to better infer sequence of ads for optimizing future conversions." In plain English, Meta is not just picking one ad for one person. It is learning which ad tends to come first, which comes next, and which one is showing when someone finally buys. The October upgrade made GEM read twice as much of each person's engagement history and pushed its learning across every surface at once.
For the swim school, a parent might first see the safety ad and keep scrolling. A week later they see the scared-kid video and watch it to the end. Then the schedule-and-price ad shows up on a Sunday night when they are planning the week, and they book a trial lesson. No owner planned that order. Meta did, because all three ads were in the same place where it could choose between them.
That is why splitting every stage into its own campaign works against you on a small budget. A "problem awareness" campaign, a "consideration" campaign and a "conversion" campaign each get a slice of a $40 daily budget, each learns from a third of the data, and Meta can no longer decide which of your ads a person should see next because the ads live in separate boxes. The setup that gives Meta the most room is the plain one: one campaign per offer, aimed at the result you actually want, with your distinct ads together inside it. We made the full case for that in Stop Building Complicated Meta Campaigns. Andromeda 2.0 did not change that advice. It strengthened it.
Your Meta Account Needs No Rebuild. Give It Different Ads, a Real Conversion Goal and a Week of Patience.
Meta closed its announcement with one sentence that tells you where your effort belongs: how much you get from these upgrades "still depends on what you give the system to work with: the outcomes you tell it to optimize for, the quality of the data you share... and how much of the optimization you let the models handle." That turns into three checks you can run this week, none of which involve restructuring anything.
1. Sort your ads by amount spent and look for near-copies. Open Ads Manager at the ad level, set the date range to the last 30 days, and sort by Amount spent. Look at the ads sitting near the bottom with almost no spend. If an ad looks and says nearly the same thing as an ad above it, Meta has most likely grouped the two and is backing the stronger one. Leave the winner alone. Instead of making another variation, write one new ad around a reason to buy you have not used yet. Check the Delivery column while you are there for Creative limited or Creative fatigue on the ads that are spending.
2. Confirm Meta is optimizing for the result you actually want, and that it can see it. Every one of these systems learns from the outcome you choose. If your campaign is set to traffic or engagement because those results looked cheaper, the smartest ranking system in the world is now very good at finding people who click and leave. For a service business that usually means leads, or calls if your phone is the real conversion. For a store it means purchases. Then make sure the result is actually being counted: a pixel that misses half your sales gives GEM half the picture. If you have not checked your tracking since Meta changed how its pixel counts customers, start with Meta's November 12 Pixel Change Only Counts Customers Who Log In.
3. Give each new ad about a week before you judge it. Meta says an ad set's learning phase usually ends after about 50 results in the week after its last significant edit, and that editing during that week resets it. A new ad that gets a slow first two days is not a verdict. Pausing it, tweaking it and relaunching it puts it back at the start of the line. If you want a fair read on a new idea, Meta's built-in creative test is the cleanest way to get one, and we walked through it in Meta Won't Give Your New Ad a Fair Shot. Here's How to Force One. When it is time to decide what to cut, Meta Already Turned Off Your Worst Ad. You're About to Turn Off the Wrong One. covers how not to kill the wrong one.
That is the whole update. A bigger, faster reader on Meta's side, and no new buttons on yours. The owners who come out ahead are not the ones who rebuilt their accounts in a panic the week after the announcement. They are the ones whose ads clearly show who they are for and why someone should buy, because that is now exactly what the system reads.
Working out those reasons to buy, and turning them into ads that Meta and your customers can tell apart, is most of what our Facebook ads management team does all day. If you would rather hand that part off, we are glad to take a look. If you are doing it yourself, the three checks above are a good Monday morning.




