MON - FRI : 09:00 AM-18:00 PM

Meta ads in 2026: Andromeda, Advantage+ and creative strategy

October 15, 2026
Tin Le

Updated on October 19, 2026: reviewed and updated figures.

In December 2024, Meta’s engineering team announced Meta Andromeda, a system that does exactly one job: out of tens of millions of live ads, it picks the few thousand that suit a specific user, before the ranking models choose which ad to show. According to Meta, the new model is 10,000 times more complex than the old one at this stage, lifting recall (the ability to find relevant ads) by 6% and ad quality by 8% on the segments measured.

Nearly two years later, tactics that used to work, such as splitting 12 ad sets by interest or duplicating one video into 20 copies with different headlines, now tend to fragment the data and push costs up. In Vietnam, where Facebook ads reached about 79.0 million people at the end of 2025 according to DataReportal, this change touches almost every business that sells online.

What is Meta Andromeda?

Meta Andromeda is the ads retrieval system Meta announced in December 2024. It runs on the NVIDIA Grace Hopper Superchip and filters tens of millions of ads down to the few thousand candidates most relevant to each user, before the ranking and auction steps. Andromeda does not pick the winning ad. It decides which ads make it into the final round.

If an ad is not retrieved for a given person, its bid and budget no longer matter for that person. Meta also explained why it built Andromeda: the volume of creative was growing fast thanks to the AI ad creation tools in Advantage+, and the old system could not keep up.

How do Andromeda, Lattice and GEM differ?

System Role Announced Results reported by Meta
Andromeda Retrieval: selects a few thousand candidates from tens of millions of ads December 2024 Recall +6%, ad quality +8% on the segments measured
Lattice Ranking: one prediction architecture for many objectives, replacing many small models that ran separately May 2023 Shared learning across objectives, useful when measurement data is limited
GEM A foundation model trained at a scale comparable to an LLM, which passes its knowledge to the ranking models November 2025 Conversions +5% on Instagram and +3% on Facebook Feed in Q2 2025

Advertisers cannot tune any of these models. What you control is the input: creative, the conversion signals you send back to Meta, account structure and budget.

What has changed in Facebook advertising in 2026?

Advantage+ is now the default

In 2025, Meta merged the manual campaign creation flow and the Advantage+ flow into one, and renamed Advantage+ shopping campaigns to Advantage+ sales campaigns. For Sales, App or Leads objectives, Advantage+ is switched on by default. A campaign keeps its Advantage+ status when three levers are all automated: campaign-level budget, an audience limited only by location or set to Advantage+ audience, and placements with no exclusions.

On Meta’s Q3 2025 earnings call, the company said the annual run-rate of revenue flowing through its end-to-end AI-powered ad solutions such as Advantage+ had reached USD 60 billion. Advertisers running leads campaigns with Advantage+ saw a cost per lead 14% lower on average than advertisers who did not use it.

Broad targeting: your inputs are now only suggestions

With Advantage+ audience, age, gender, detailed targeting, custom audiences and lookalikes are only suggestions. The system prioritizes people who match those suggestions, then expands when it predicts better performance elsewhere. The only hard limits are location, language, minimum age and exclusion lists. As a result, splitting ad sets into “skincare” and “makeup” no longer separates two groups of people the way it once did.

Creative has become the targeting signal

Once the audience is broad, the ad content decides who stops to watch. A video about oily, acne-prone skin aimed at students pulls in a completely different group from a price comparison video for gift buyers. Jon Loomer, in his breakdown of creative diversification examples, sums up Meta’s message: invest less in targeting and put your effort into making many ads that are genuinely different.

Ads that look alike are likely to be counted as one

Meta recommends using creatives that differ clearly in format, visuals and message. Meta has not published the details of the mechanism, but practitioners have observed that ads with near-identical visuals and themes are often treated by the system as variations of the same creative. The common term in the industry is “Entity ID” (not an official Meta term): 20 ads sharing one product photo, differing only in headline, may be counted as just one. Since late 2025, Meta has been gradually adding creative similarity and fatigue indicators to Ads Manager, and they may not be available in every account yet.

Advantage+ creative enhancements

Meta offers more than 20 automatic creative adjustments, such as brightness tweaks, adding music, expanding images and generating text variations. The list and the names change often. Preview each option before you switch it on, especially with Vietnamese-language text and in industries with tight content rules, such as health supplements.

Conversions are counted differently too

Since March 2026, click-through attribution only counts link clicks. Conversions after a like, share or save have moved to engage-through attribution with a fixed 1-day window. During 2025, Meta also gradually rolled out an incremental attribution option in Ads Manager: the campaign optimizes for conversions that Meta estimates were caused by the ads, so reported conversions may be lower than under the standard method.

How should your account structure change?

Fewer campaigns, fewer ad sets, and enough data in each ad set to learn. According to Meta, an ad set needs about 50 optimization events within 7 days to exit the learning phase, and that number is counted per ad set.

With a target cost of VND 250,000 per order, each ad set needs about VND 12.5 million a week to get 50 orders. An account with 12 ad sets would need VND 150 million a week just to get all of them out of the learning phase, a level most small and medium-sized businesses do not have.

A suggested structure for most sales accounts:

  • One Advantage+ sales campaign holding most of the budget, containing the creatives that have already proven themselves.
  • One test campaign for new concepts, so new creatives do not have to compete right away with ads that already have a strong history.
  • A separate campaign for existing customers only when the message is genuinely different (member offers, repeat purchases).
  • For apps, Advantage+ app campaigns follow the same logic, optimizing for valuable in-app events rather than installs alone.

Meta creative strategy: how to diversify the right way

Creative diversification means making ads that differ at the idea level, each one speaking to a different group of people or a different need. There are four axes of difference:

Axis Real difference Not different enough
Concept Problem and solution, customer reviews, before and after, product demos The same idea with reworded copy
Persona Students with oily skin, new mothers, office workers, gift buyers The same model with a different background color
Format Vertical video, static images, carousels, clips cut from livestreams, self-shot KOC videos The same video exported in several aspect ratios
Hook A question, a result, a counterargument, a product-in-use scene A different font or sticker in the first second

How many concepts are enough? Meta gives no general number. Work it out from your budget, because each concept needs enough spend to be judged. If you spend VND 5 million a day and run 30 concepts, most of them will barely get delivery. At that level, 4 to 6 genuinely different concepts, each in 2 to 3 formats, make the results easier to read.

The first few seconds of a short video decide whether people keep watching, so hooks need to be tested on their own. AI speeds up production: writing scripts for each persona, generating hook variations, voiceovers (see tools and examples in our article on using generative AI in marketing). Distinct ideas still have to come from understanding your customers.

A 4-step creative testing framework

  1. Build a matrix of 3 to 4 personas by 3 to 4 angles, and only pick the cells with a clear business reason.
  2. Run new concepts in the test campaign for 7 to 10 days, without editing them midway.
  3. Read cost per result, 3-second view rate and click-through rate first, and revenue metrics only once there is enough data.
  4. Move winning concepts to the main campaign, keeping the same post ID so you do not lose engagement. Write down why the losing concepts lost.

How do you measure when Meta optimizes on its own?

Meta’s models learn from the signals you send back. Wrong data means optimizing in the wrong direction. Start by running the Meta Pixel and the Conversions API (CAPI) side by side, with event deduplication, and track Event Match Quality in Events Manager. In Vietnam, a large share of orders are closed over Messenger, Zalo or a hotline. If you do not send closed orders from your CRM back to Meta, the system only sees who sent a message, not who actually bought.

Next, Meta’s attribution model only sees touchpoints on Meta’s own platforms. Businesses that also run Google, TikTok and marketplace ads should cross-check against multi-touch attribution or total revenue.

Last comes measuring true lift (incrementality). Meta’s Conversion Lift splits users into a group that sees the ads and a group that does not. If you do not yet qualify for a lift test, you can switch ads off in a few provinces for 2 to 4 weeks and compare sales with similar provinces. Our article on optimizing the full marketing funnel with AI shows how to fit these measurement layers into the funnel.

How should you allocate budget?

An easy starting split: 70% for the main campaign, 20% for concept tests, and 10% for bigger experiments such as new formats or lift tests. When frequency keeps rising and click-through rate keeps falling, the testing share needs to grow.

Avoid raising or cutting budget too sharply in one go, because a big change can send an ad set back into the learning phase. Judge ROAS over 7-day periods instead of day by day. During 11.11, 12.12 and Tet (Lunar New Year), CPM usually rises sharply, so peak-season creative should finish testing a few weeks in advance. Once the season starts, the only job left is to scale budget behind the winning concepts.

Illustrative example: restructuring a cosmetics brand’s account

This is a hypothetical scenario for illustration, not data from a specific client.

A skincare brand spends VND 300 million a month (VND 10 million a day), split evenly across 12 ad sets targeted by interest and age, with four videos duplicated into every ad set. At a target cost of VND 250,000 per order, each ad set gets about VND 833,000 a day, or roughly 23 orders a week, less than half the 50-event threshold.

Item Before After
Structure 12 ad sets by interest and age 1 Advantage+ sales campaign (80% of budget), 1 test campaign (20%)
Targeting Detailed interests Nationwide, minimum age, excluding buyers from the last 30 days
Creative 4 duplicated videos with different headlines 8 concepts for 4 personas, each with a vertical video, a static image or a KOC livestream clip
Measurement Pixel only, orders closed in the inbox not recorded Pixel and CAPI, closed orders pushed from the CRM back to Meta
Main metrics Daily ROAS per ad set Weekly account-wide cost per order, share of new customers, a lift test after 6 to 8 weeks

The main campaign now has about VND 8 million a day, enough for more than 200 orders a week at the same target cost. Actual results depend on creative quality and data, so track it for 3 to 4 weeks before drawing conclusions.

Common mistakes Vietnamese advertisers make

  • Splitting ad sets by interest, age and gender when Meta already treats these inputs as suggestions.
  • Calling a change of headline, background color or sticker a “creative test”, when the system most likely sees these versions as one.
  • Switching ads off after 1 to 2 days or adjusting budget every day, so the ad set never finishes learning.
  • Optimizing for messages without sending closed orders back to Meta, so the system finds more people who like to message, not people who buy.
  • Relying on one KOC or one video style for every campaign, with no backup concepts when creative fatigue sets in.

Meta ads checklist for 2026

  • No more than 2 to 3 live campaigns for the same objective.
  • Each ad set can reach about 50 optimization events in 7 days on its current budget.
  • Targeting keeps only the hard limits: location, language, minimum age, exclusions.
  • The main campaign has at least 4 concepts that differ in angle or persona.
  • A test campaign is running, with new concepts every 2 to 4 weeks.
  • Pixel and CAPI run side by side, and orders closed through Messenger, Zalo and the hotline are sent back from the CRM.
  • Reports account for the March 2026 attribution change, with at least one incrementality measurement per year.
  • Each Advantage+ creative enhancements option is reviewed before it is switched on.

Adtek runs performance advertising and creative strategy for retail, cosmetics and app brands. If you want to review your Meta account structure for the way the system works now, talk to the Adtek team.

Sources

Leave a Comment

There are currently no reviews yet

Latest news