Meta Ads 2026: With Andromeda, the Creative Becomes Targeting

By Million MarketingReading time: 8 min

On Meta Ads in 2026, audience settings no longer decide who sees your ads: the content of your creatives does. Since the roll-out of Andromeda, Meta’s ad retrieval engine, a first stage of the system narrows tens of millions of eligible ads down to a few thousand candidates for each user [1]. The image, video, copy and sound of your ad therefore determine whether it even enters the race.

For the marketing team of a mid-sized company or a large enterprise, this changes three things: account structure (fewer campaigns, more budget per campaign), creative production (genuinely different angles, at volume) and measurement (looking beyond the ROAS shown in Ads Manager). This article covers each point, with the figures Meta has published and our 90-day roadmap.

What Andromeda, GEM and Advantage+ have changed

Delivering a Meta ad happens in several stages: retrieval (which ads deserve consideration), then ranking (which one to show, and at what price). The two models that matter today act at different levels.

Andromeda is the retrieval system. It filters tens of millions of candidate ads down to a few thousand. On 2 December 2024, Meta announced a 6% recall improvement in the retrieval system and an 8% gain in ad quality on selected segments, with inference throughput more than tripled [1].
GEM (Generative Ads Recommendation Model) is a foundation model that improves the ranking models. According to Meta (10 November 2025), it delivered a 5% increase in ad conversions on Instagram and 3% on Facebook Feed in Q2 2025 [2].
Advantage+ covers Meta’s automated campaigns. Meta reports that advertisers using image generation see 7% more conversions (December 2024 announcement) [1].

The practical takeaway: the more powerful the AI system upstream, the more the quality and variety of the signals you feed it, that is your creatives and your conversion data, drive performance.

Why creative diversity replaces audience segmentation

Meta’s systems analyse the content of ads. A video that opens with « CFOs, what happens to your working capital in 2026? » does not send the same signal as a static product shot in a luxury setting. The algorithm matches each ad with the profiles who stop on it, then adjusts delivery. The message selects the audience, not the other way round. Practitioners call this principle « creative diversification » [5].

This requires different angles, not cosmetic variations. Five versions of the same visual in five colours send a single signal. We work with four families of angles:

Proof and results: client testimonial, key figure, case study.
Status and desirability: refined art direction, essential for luxury.
Education: demonstration, motion-design explainer, « how to ».
Native formats: talking-head and creator content, at the pace of Reels and Stories.

Each family attracts different profiles. You « target » several personas without ticking a single demographic box.

Recommended account structure in 2026

A fragmented account dilutes signals: each ad set stays too small to exit the learning phase. Our starting recommendation, to be adapted to your conversion volume:

One main campaign in Advantage+ (sales or leads), with broad targeting, carrying most of the budget. As an order of magnitude, we aim for 70 to 80% of the Meta media budget on this campaign.
One dedicated creative-testing campaign, to validate new concepts before feeding them into the main campaign.
An optional retargeting campaign, only if your messaging requires it (abandoned basket, B2B cross-selling).

Consolidation brings budget liquidity: instead of spreading budget across ten audiences, you let the algorithm fund the winning creatives. It also reduces competition between your own ads in the auctions. For B2B advertisers, always compare this set-up with your LinkedIn set-up: our analysis of B2B LinkedIn Ads in 2026 helps you split budget between the two platforms.

Producing creatives at scale with AI, without compliance risk

Feeding the algorithm with different angles means producing a lot, fast. Generative AI is used to vary settings, adapt formats (16:9, 9:16, 1:1), test hooks and speed up editing. That is the role of our AI Studio; we detail the return on investment of this approach in our article on AI-generated visuals.

This volume comes with obligations. As set out in our analysis of the AI Act and AI-generated advertising content, the visible labelling obligation under Article 50 has applied since 2 August 2026; machine-readable marking benefits from a technical grace period until 2 December 2026 for systems already on the market. Fines can reach €15 million or 3% of worldwide annual turnover [3]. Four operating rules:

Keep the master files and their provenance metadata: social networks may strip them on publication.
Label realistic visuals generated or manipulated by AI, using the platform’s labelling tools.
Have a human approve every asset (logo, materials, proportions), with a written record of the approval.
Protect the brand image: for luxury, AI produces drafts and variations; the art director decides.

Measuring: Conversions API, incrementality and MMM

Steering Meta on Ads Manager ROAS alone amounts to marking your own homework. Three layers of measurement complement each other:

The Conversions API (CAPI) sends events from the advertiser’s server to Meta, without relying on the browser. For B2B, connecting the CRM lets you optimise for qualified leads rather than completed forms. Monitor deduplication and event match quality.
Incrementality tests (geo tests, conversion lift studies) answer the real question: how much revenue would disappear if you switched Meta off tomorrow?
Marketing Mix Modelling allocates budget across channels using aggregated data, with no personal data. Meta offers Robyn, an open-source project [4]. Our guide to MMM and incrementality compares the tools and sets out a triangulation method.

The consent context makes this even more necessary: see our analysis of B2B search advertising after cookies for the signal issues on the Google side.

B2B, mid-sized companies and luxury: applying these principles

Here are two typical scenarios, deliberately without result figures: they are working frameworks, not case studies.

B2B mid-sized company (software publisher, manufacturing). An Advantage+ leads campaign with broad targeting across France, fed by 25 to 30 short videos on three angles (business problem, client proof, demonstration). The CRM only sends Meta the leads that reached the « discovery meeting » stage, to train the algorithm on quality rather than volume. The winning creatives then feed LinkedIn.

Luxury house. A multi-country campaign with a global budget, product visuals shot in the studio then varied by AI in settings approved by the art director, labelled in line with the AI Act. Measurement relies on an incrementality study that includes in-store sales, so Meta is not judged on e-commerce alone.

90-day roadmap

Days 1 to 30: foundations. Account audit, campaign consolidation, Conversions API check (deduplication, match quality), CRM connection for B2B.
Days 31 to 60: creatives. Angle workshop, first wave of creatives with compliance check and brand approval, launch of the testing campaign.
Days 61 to 90: scaling and measurement. Winners moved into the main campaign, gradual budget increases, first incrementality test, Meta data integrated into the MMM.

Common mistakes to avoid

Multiplying ad sets and audiences, which fragments the signal.
Confusing cosmetic variations with genuine creative angles.
Publishing AI visuals without labelling or a record of human approval.
Optimising for low-value events (clicks, forms) rather than conversion quality.
Judging Meta solely on reported ROAS, without an incrementality test.

Our approach at Million Marketing

Our « Remix » method orchestrates SEO, SEA, AIO, Paid Social and the AI Studio around a single performance objective. On Meta, we restructure the account, produce creatives with the AI Studio within a compliant framework, and measure true incrementality before increasing budgets. An effective Meta campaign also generates brand searches on Google: that is why we align Paid Social with SEO and SEA rather than running each channel in a silo.

Explore our Paid Social offer →

Sources and references

Frequently asked questions about Meta Ads, Andromeda and creative

What is Andromeda on Meta Ads?

Andromeda is Meta’s ad retrieval system: it operates at the first stage of the recommendation chain and narrows tens of millions of eligible ads down to a few thousand candidates before final ranking. In December 2024, Meta reported a 6% gain in recall and an 8% gain in ad quality on selected segments [1].

Should you still target precise audiences on Meta Ads in 2026?

In most cases, no. Broad targeting combined with strongly differentiated creatives lets the algorithm find the right profiles. Audience restrictions remain useful for legal, geographic or brand constraints, but they no longer replace the message as the targeting lever.

How many creatives should you produce each month?

There is no magic number. Our starting point for a mid-sized or enterprise account: a wave of 20 to 30 creatives across 4 to 5 distinct angles, then continuous renewal based on the winners. What matters is the diversity of angles, not the number of variations of the same visual.

Can you use AI-generated visuals in Meta ads in France?

Yes, under conditions. Since 2 August 2026, Article 50 of the AI Act requires visible labelling of content generated or manipulated by AI when it could be taken as authentic, with fines of up to €15 million or 3% of worldwide turnover. Keep the master files, document human approval and apply the required disclosure.

How do you know whether Meta Ads really creates additional sales?

The ROAS shown in Ads Manager measures attribution, not causation. To settle the question, combine the Conversions API (a reliable signal), an incrementality test (test group and control group) and a Marketing Mix Model to allocate budget across channels.