Meta Andromeda Update: What Changed and How to Check Your Account

There is no single Meta Andromeda update. Meta announced Andromeda on 2 December 2024 as the ads retrieval engine behind Advantage+, its suite of automated campaign tools, and said it extended the system in Q4 2025. Retrieval narrows tens of millions of ad candidates to a few thousand. Searches for "Andromeda AI" and "Andromeda ads" point at this system.

Retrieval is the first step of a multi-stage ad system. Ranking, which scores the shortlisted ads, comes after.

Meta's July 2026 call does not mention Andromeda. Retrieval is the stage Andromeda covers, so we read the new model as related to it, but that link is our inference, not Meta's statement. For a buyer, the practical point is the same: plan for gradual shifts in delivery rather than waiting for a dated event.

Andromeda shortlists which ads are considered. GEM, Meta's Generative Ads Recommendation Model, is a large ranking model that passes what it learns to smaller models used at serving time. Lattice merges separate models into larger shared ones. None of the three generates your creative.

Meta does not run its largest models, GEM among them, on every ad request. They transfer knowledge to lighter models that run at serving time (Meta Q4 2025 call).

These are company-wide figures, not a forecast for your account. Meta's creative-generation tools, such as the Muse Image model announced on the Q2 2026 call, are separate products.

No source shows that the Andromeda update lowered ROAS across Meta accounts. One vendor study of its own customers, Confect's analysis of 3,014 ecommerce advertisers in 2025, found average return on ad spend (ROAS, revenue per dollar of ad spend) down 7%. Prospecting ROAS fell 13%. These figures are vendor-reported and show a decline in one sample, not its cause. The data is in the Confect Andromeda study.

The same study covered $834 million in spend and 1 million ads. Its other findings:

• Conversion rates fell 17%.
• Retargeting stayed relatively stable.
• The top-performing group's ROAS fell 31%, while the bottom group edged up.

Four cautions apply:

• Every advertiser in the sample uses Confect, a catalog-ads platform.
• Confect excluded 42.4% of ads for failing impression or performance thresholds, and 5.1% of advertisers as outliers.
• A strong group falling while a weak group rises is also what regression to the mean looks like (extreme results drifting back toward the average on their own). That reading is our inference.
• Do not assume your account is hit or spared. Test it.

Meta states that look-alike ads are grouped as variations of one creative, and that it is testing creative fatigue and similarity insights in Ads Manager. It has published no similarity percentage or creative quota, so those remain agency convention. Agencies publish counts such as 10 to 20 ads per ad set, which are not Meta guidance.

As of October 2026:

An optimization event is the result an ad set is built to produce, usually a purchase. The learning phase is the period when Meta's delivery system is still working out who responds to an ad set. The 50-event guidance is long-standing and not specific to Andromeda. Treat the convention rows as starting points to test, not rules.

Concepts you can judge per week = weekly budget ÷ cost per purchase ÷ 5. At a $50 cost per purchase, a $50-a-day account gets about 7 purchases a week, enough to read one concept. It takes about $350 a day to read nine.

A concept is one distinct idea: a different angle, offer, creator, setting or opening. A new caption or crop of the same idea is the same concept.

Our model, at a $50 cost per purchase:

The assumptions are:

• Purchases split evenly across concepts. Delivery is uneven in practice, so treat the table as a ceiling.
• Five purchases per concept is our floor for a directional read, not statistical proof.
• Cost per purchase stays constant.

The same arithmetic applies inside one ad set, using Meta's guidance of about 50 events a week. With 10 concepts, each gets about 5 purchases. With 15, about 3.3. With 20, 2.5. To read 10 concepts at 5 purchases each, one ad set needs 50 purchases a week, which is $2,500 a week at $50 per purchase.

Below those budgets, our recommendation is to run fewer, more distinct concepts and read them on earlier signals:

• Thumb-stop rate, the share of impressions where someone pauses on a video.
• Click-through rate, clicks per impression.
• Add-to-cart rate, the share of visitors who add a product to their cart.

Andromeda is the likely cause only when new-customer campaigns fall while retargeting holds and your tracking and edits are clean. Run these five tests in Ads Manager, comparing the last 28 days with the prior 28 (except test 4). The thresholds are our rules of thumb, so adjust them to your account's normal volatility.

Did the drop start on one day, or drift over weeks?

A single-day fall of more than a third in daily ROAS that then holds points at your account or tracking: check edits, the conversion setup and Meta's status first. A decline spread over three or more weeks with no account change is consistent with creative supply or delivery shifts.

Pass: no single-day fall of a third or more that holds. Fail: one such fall.

Is prospecting down while retargeting holds?

Prospecting means campaigns aimed at people who have not bought; retargeting means people who have. If prospecting ROAS is down 10% or more while retargeting is within 5%, creative supply is the likely cause. The Confect sample above showed the same shape: prospecting down 13%, retargeting stable.

Pass: prospecting is down less than 10%, or retargeting fell with it. Fail: prospecting down 10% or more while retargeting is within 5%.

Is one concept taking most of your spend?

Group live ads by concept, then check the last 14 days. If one concept and its variants take more than 70% of spend, you effectively have one live ad. Add a genuinely different concept before adding more variants.

Pass: no concept above 70% of 14-day spend. Fail: one concept above 70%.

Did your CPM rise more than the market's?

CPM is the cost per 1,000 impressions. Compare your CPM for the last 28 days with the same 28 days a year earlier. Meta's global average price per ad rose 12% year over year in Q2 2026 (Meta Q2 2026 call). That is only a rough company-wide reference, because Meta publishes no CPM benchmark.

Pass: your CPM is up 12% or less year over year. Fail: your CPM is up more than 12%.

Does Meta's purchase count still match your store?

Compare Meta-reported purchases with orders in your store for the same 14 days. If the ratio moved by more than 20% against the previous 14 days, fix tracking before touching creative.

Pass: the ratio moved 20% or less. Fail: it moved more than 20%.

Reading the result. If tests 1 and 5 pass and tests 2 and 3 fail, treat it as a creative supply problem and use the capacity table to decide how many new concepts to launch. If test 1 or 5 fails, fix the account first. If test 4 fails, your costs rose faster than Meta's market-wide price per ad, so check audience, placements and seasonality before blaming Andromeda. A test 4 pass does not clear creative: rising market prices can explain part of a drop alongside weak creative supply.

We plan it as a weekly decision. Our Creative Intelligence platform combines your ad-performance data, competitor activity, customer signals (reviews, comments, direct messages and Reddit discussions) and past ad learnings into a ranked test plan every Monday. The plan says what to test and what to skip, and why, with execution-ready creative briefs.

Because Andromeda draws from the ads you supply, the decision that stays yours is which distinct concept to make next and how many your budget can read. Use the capacity table to choose how many items from the plan to launch. You can browse examples in our trending Meta ads by category.

Pricing is tailored to your accounts and team. Ask for it when you book a demo with our founding team.