# What Meta Andromeda Changes for Ad Creative

> Meta Andromeda is the stage that shortlists ads before ranking. What Meta has said, what is only inference, and what it means for the ads you run for leads.

- Canonical: https://nask.ai/what-meta-andromeda-changes-for-ad-creative
- Published: 2026-11-07
- Author: Rodrigo Nask (https://rodrigonask.com/), Nask AI

---

## Key takeaways

- Andromeda is the retrieval stage: it shortlists ads for a person before the ranking stage scores them.
- Meta's guidance warns against near-duplicate ads, so count concepts, not files.
- Targeting still sets who is eligible, but a Meta executive says it matters less as ranking models improve.
- Judge the result by leads and booked calls, not by what Meta says about its own system.

Meta Andromeda is the retrieval system that narrows the pool of eligible ads to a shortlist for each person before a ranking stage picks one; for ad creative, what changes is that the shortlist can only be as varied as the ads you supply.

Everyone selling you something has an opinion on Andromeda. A course says it killed targeting. A thread says it killed ads with a talking head. A vendor says you need forty creatives by Friday. Meanwhile your ad account looks the same as it did last month, and your question is plain: what do I do differently on Monday?

This page answers that in order. What Andromeda is, according to the people who have reported on it. What Meta itself claims, labeled as Meta's. What is inference. Then what it changes for the ads you make when your business sells a call, not a product.

> **Figure (animated on the web version):** Three stages left to right. A pool of thirty grey ad dots narrows at the retrieval stage to five purple tiles, then the ranking stage picks one winner. The dots cut at retrieval fade out and never reach ranking.
>
> Retrieval is where an ad gets in, ranking is where it wins. **An ad that is not shortlisted never reaches the auction.**

## What is Andromeda, in plain terms?

Andromeda is the first of two stages that decide which ad a person sees. Trade press describes it as Meta's retrieval system: it takes a very large pool of eligible ads and narrows it to a shortlist of ads that person might like. A second stage, ranking, then scores that shortlist in more detail and picks.

Picture a recruiter with ten thousand applications. First someone cuts the pile to a few dozen. Then the hiring manager reads those few dozen closely. If you were cut in the first round, the hiring manager never sees you, however good you were.

That is the point of the stage. As [AdExchanger reported](https://www.adexchanger.com/ai/meta-has-a-new-ad-ranking-system-built-on-llms-plus-upgrades-to-andromeda-and-gem/), retrieval narrows the pile to a shortlist, and Andromeda got a hardware and software overhaul through Meta's partnership with Nvidia. The reporting from [PPC Land](https://ppc.land/meta-says-andromeda-lifted-facebook-conversions-2-4-since-2025/) adds the framing that an ad that retrieval does not shortlist never reaches the auction. Ranking is where an ad wins. Retrieval is where it gets in.

## What has Meta actually said about it?

Meta has said four things that matter to an advertiser, and most of what circulates is built on top of them. These are Meta's statements, relayed by the press, and Meta is the only source for each.

**It is a big rebuild.** [Jon Loomer's write-up](https://www.jonloomer.com/meta-andromeda/) quotes Meta calling Andromeda an end-to-end system co-designed across hardware, software and machine learning, and claiming a 10,000x increase in the complexity of models used for ad retrieval. In October 2026, as [ContentGrip](https://www.contentgrip.com/meta-andromeda-ad-retrieval-billion/) reported, Meta published an update saying the rebuilt engine evaluates more than a billion eligible ads per second.

**Creative diversification is the lever.** Loomer also quotes Meta saying that [creative diversification](https://nask.ai/meta-ads-creative-diversification-explained) is the best lever for finding relevant audiences. As he reports it, Meta's guidance warns against near-duplicates and different versions of the same idea.

**There is a conversion claim.** Meta attributes a 2.4% increase in Facebook conversions to Andromeda since 2025. [Relevant Audience](https://www.relevantaudience.com/meta/meta-andromeda-2026-ads-retrieval-explained/) points out that the post gives no baseline, method or time window, and PPC Land adds that an earlier investor figure was 1.4% for Q3, and Meta has not said how the two relate. It is a platform-wide average. It tells you nothing about your account.

**There is no new switch.** As far as the reporting shows, there is no Andromeda setting, opt-in or report to look for, and Meta frames the benefit as automatic. If a guide sells you a secret Andromeda toggle, close the tab.

> **Figure (animated on the web version):** A ledger of six claims about Andromeda. Four are stamped as things Meta or a Meta executive says, one with no baseline. Two, exact creative counts and per-creative paths, are stamped as not published by Meta.
>
> Sort every Andromeda claim by who said it. **If a number has no link back to Meta, it is a guess.**

## Why would a retrieval stage care about creative?

This section is inference, not something Meta publishes. It is the reasoning most practitioners use, and it is a working model, not a verified mechanism.

If the first stage picks a shortlist for a person from the ads you gave it, your ads are the raw material. A shortlist built from five ads that make the same point to the same buyer has one real idea in it. A shortlist built from five ads that speak to five different buyers has five. When a person arrives, the system has more chances to find an ad that fits them.

A Meta executive, speaking to AdExchanger, gave the other half of the picture: the importance of targeting is diminishing over time as ranking models get better at selecting who an ad would be relevant for. Put those together and the lever moves. Who sees an ad is increasingly decided by the system. What the ad says is decided by you.

Be careful with the details that circulate. Some guides describe the system reading visual features and assigning each creative its own path. Others give exact creative counts to hit. Meta has not published either. When a source states a mechanism or a number with no link back to Meta, you are reading a guess, however confident it sounds.

> **Figure (animated on the web version):** Two accounts compared. Account one has eight image files that differ only by color and all drop into a single concept lane, so the counter reads concepts: 1. Account two has four files with different ideas, each opening its own lane, and the counter reads concepts: 4.
>
> Alike ads add little, so count lanes, not files. **Eight recolors are one concept. Four ideas are four.**

## Why does Meta keep talking about concepts?

Because the people building the system describe it that way. In AdExchanger's report, Meta's VP of ads infrastructure said ads, campaigns and campaign structures change all the time, which is why the models track the underlying concepts rather than specific creatives. That is a statement about how the models read ads, and it is the most direct support for counting concepts instead of files.

It also explains why the October 2026 update matters less to you than the headlines suggest. A faster retrieval engine that evaluates more eligible ads per second, in ContentGrip's account of Meta's figure, changes how many ads the system can consider, not what makes one worth considering. The ad is still the thing you control, and a concept is still the unit the models appear to care about.

There is a practical edge to this. If the models follow concepts, then renaming a concept does nothing, and retiring an ad without retiring its idea does little too. When you read your results, read them by idea: which buyer, which reason, which first step. That is the grouping the system seems to use, and it is the grouping that tells you what to make next.

## What counts as a different ad?

An ad is different when it would persuade a different person, or the same person for a different reason. Four changes pass the test:

- **The angle.** The reason to act: the cost of waiting, a result someone got, the founder's story, the objection answered.
- **The offer frame.** What the reader gets for responding: a call, a free audit, a checklist, a short diagnostic.
- **The buyer.** Who the first line speaks to. An owner and an operations manager want different things from the same service.
- **The visual.** A different kind of picture, not a different shade: a face to camera, a screenshot of a result, a plain checklist, a photo of the problem.

Changes that mostly fail it: a new button color, the same headline reworded, a different stock photo of the same kind, another aspect ratio of the same ad. You still make the aspect ratios, because placements need them. You only stop counting them as extra ideas.

> **Figure (animated on the web version):** Two columns of gates lighting up one by one. On the left, a new angle, a new offer frame, a new buyer and a new kind of visual all pass the test of persuading a different person or the same person for a different reason. On the right, a new button color, a reworded headline, the same kind of stock photo and another aspect ratio all fail it.
>
> One test sorts the changes: **does it persuade a different person, or the same person for a different reason?**

The full test, with the Studio fields that produce each kind of difference, is in [how many ads per ad set on Meta](https://nask.ai/how-many-ads-per-ad-set-on-meta). This page does not repeat it, because the ad count was never the point of Andromeda.

## What does Andromeda not change?

Four things stay as they were, and each is a place where the Andromeda stories overreach.

**Your offer.** A retrieval stage can shortlist an ad, but it cannot make a weak offer convert. Ten angles on an offer nobody wants give you ten ads nobody wants.

**Your funnel.** Meta sees an impression, a click and a form. It does not see that the lead was a student, a competitor or a person who never picked up. If the signal going back is a form fill, the system learns to find more form fills. It learns about booked calls only if you tell it. That is what the nsk.js pixel and funnel stages are for, covered in the [Nask AI pixel setup guide](https://nask.ai/nask-ai-pixel-setup-guide).

**Your judgment.** More ads produce more data to read, not less. You still decide which angles are on brand, which claims are true and which ads should never run.

**The need for a boundary.** A boundary still matters in practice: the country you can serve, the people you must exclude. What shrinks, on the executive's account, is the value of narrow interest stacks.

## Does this help a service business or hurt it?

It helps if you can name more than one reason a buyer books. It hurts if you have one stock message and a template.

A store has a catalog. Every product is a new angle, a new picture, a new price. A service business has one offer, a few kinds of buyer and a handful of reasons to book. So the variety does not come from a product feed. It comes from crossing three lists: who the buyer is, what they are worried about, and what first step you ask them to take.

Take a consultancy that audits ad accounts. The reasons to book might be: spend is rising and leads are not, the in-house person has left, the reports are impossible to read, the owner wants a second opinion before renewing. The buyers might be the owner and a head of marketing. The first steps might be a call or a written audit. Four reasons, two buyers and two first steps make sixteen combinations. Most will be silly. A handful will be real ads.

Then comes the part that matters for a business that sells a call. A cheap lead who never books is worse than an expensive one who does. That is true with or without Andromeda, and it is why [tracking which ad produced a sale](https://nask.ai/track-leads-from-facebook-ads-to-sales) is the other half of this plan. Variety is how you give Meta more to match. Tracking is how you find out which of the matches were worth having.

## How do you audit your own ads for this?

Write one line per ad: who it is for, and why they would act. Then merge the ads that got the same line. What is left is your real count.

1. **List every active ad in the account.** Include the ones with almost no spend.
2. **Write the line.** "Owner of a small firm, afraid of paying for leads that never book." "Head of marketing, needs a second opinion before renewal."
3. **Merge duplicates.** Two ads with the same line are one concept in two costumes.
4. **Find the gaps.** Which buyers have no ad? Which reasons have none? Which first steps have none?
5. **Fill the gaps first.** Make the ads for the empty cells before you make a fourth version of the one that is working.
6. **Retire by reading, not by feel.** An ad that spends little but brings booked calls is doing a job. Keep it.

> **Figure (animated on the web version):** A terminal transcript of the audit. Six ads are each written as one line saying who it is for and why they would act. Three lines are marked as duplicates and merged, and the result reads: 6 ads, 3 concepts.
>
> Write one line per ad and merge the repeats. **The count that is left is the one Meta can use.**

If you find you have three real concepts across forty ads, that is the finding. It is not a crisis. It tells you where the next batch should go.

## When is Andromeda the wrong explanation?

Often. The word has become the answer to every bad week, which is a reason to be suspicious of it.

Jon Loomer says not to expect miracles, and that success is less about the number of ads and more about making the best ads you can. If results are poor, he argues, that is not proof Andromeda does not work: performance depends on the ads you supply.

Before you blame the system, check what you can see. Is the pixel firing, and are booked calls reaching Meta? Did the offer change? Did the landing page or form change? Did budget move? Is the audience boundary wider than the people you can serve? Each of those has a cleaner fix than "make more ads".

Variety is the right fix when the pattern looks like this: one ad carries almost all the spend for weeks, cost per booked call is rising, frequency is climbing for the same people, and all your recent ads say the same thing in different clothes. Once tracking checks out, that is a creative problem. Anything else is something else.

## Where does Nask AI fit?

In [Nask AI Studio](https://nask.ai/how-it-works) each group reads as a sentence: how many ads, for which buyer, at which awareness stage, leading with which angle, and you approve every ad before it ships. When you publish, pick the structure "Group → ad set" and each group becomes its own Meta ad set; the steps are laid out in [the guide on ads per ad set](https://nask.ai/how-many-ads-per-ad-set-on-meta).

## What should you do this week?

Do three things, in this order, and ignore the rest of the Andromeda noise.

First, run the one-line audit above on your active ads and write down the real count. Second, check that booked calls, not only form fills, are going back to Meta, because that is the signal the system is learning from. Third, brief the gaps: one new ad for each empty buyer and reason, not ten versions of your best performer.

Then leave it alone long enough to read it. Read cost per booked call by ad set first, spend concentration second, and frequency third. If the new ads get spend, they found people the old ones missed. If Meta ignores them all, they were probably copies.

## Frequently Asked Questions

### Is Andromeda a setting I can turn on or off?

No. Andromeda is part of how Meta delivers ads, and as far as the reporting shows there is no Andromeda setting, opt-in or report to switch on. What you control is the input: the ads, the offer, the targeting boundary and the conversion signal. This guide treats creative variety as the part of that input you can change this week.

### Did Andromeda change how many ads I should run?

It changed what an ad has to be to count. Jon Loomer notes Meta's older recommendation of no more than six ads per ad set disappeared in February, and Meta's guidance on creative diversification warns against near-duplicates. So count distinct concepts, not files. The separate guide on ads per ad set walks through the test.

### Do I still need to set targeting?

Yes, as a boundary. Reporting on the retrieval stage says targeting still applies. A Meta executive has also said targeting matters less over time as ranking models improve. A sensible reading is to keep the boundary honest, such as geography and any hard exclusions, and put your effort into the ads.

### Is a poor result after Andromeda proof the ads are not different enough?

No. Many things move results: offer, landing page, tracking, budget and seasonality. A flat ad set can have plenty of variety and a weak offer. Check the funnel and the measured leads first, then ask whether your ads say different things to different buyers. Variety is one cause you can fix, not the only one.

### Can I trust the numbers Meta reports about Andromeda?

Treat them as Meta's own measurements. Trade press has noted the conversion figure comes with no baseline, time window or method, and it is a platform-wide average, so it cannot forecast your account. Use the figures to understand the direction of travel, and judge your own ads by leads and booked calls.

## Give the shortlist something to choose from

If retrieval works the way practitioners infer, more distinct ads give it more to choose from. Write down the buyers you sell to and the reasons they book, and brief one ad for each pair that holds up. [Book a demo](https://cal.com/rodrigo.nask/60min) and see it done from one brief.

See Nask AI brief, generate and publish distinct ads for your own offer, with your approval on every ad. New campaigns publish to Meta paused by default.

- [Book a demo](https://cal.com/rodrigo.nask/60min)
