UGC questions › Testing ads on Meta
TL;DR
Meta's Andromeda update changed the first step in how Meta picks which ads to show. Meta says it improves results and that varied creative now matters more than narrow targeting. But Meta has never said how many ads to run or how it treats ads that look alike. Most Andromeda "rules" you read online are marketers' guesses. In our data, running many copies of one ad didn't help it last longer.
Key findings
TL;DR: Andromeda is the system that makes Meta's first shortlist of ads for each person. A second system then picks the final ad.
Meta's engineering team announced Andromeda on December 2, 2024. It handles "retrieval", the first stage of Meta's ad system. For each person and each moment, it picks a few thousand candidate ads from tens of millions. Later stages rank that shortlist and choose the ad you see (Meta Engineering).
Meta said the new system could run models 10,000 times more complex than before, on NVIDIA Grace Hopper chips and Meta's own MTIA chips. It reported a 6% gain in recall, which measures how often the right ads make the shortlist, and an 8% gain in ad quality "on selected segments". It didn't say which segments or how quality was measured.
Andromeda isn't a setting you turn on. It runs behind every campaign, alongside other models Meta has built since, like GEM, its large ads model, and Lattice.
TL;DR: Meta has told us what Andromeda is and shared a few results. Its main advice is to make more kinds of ads.
| Date | What Meta said | Source |
|---|---|---|
| December 2024 | Andromeda launched as the retrieval stage. 10,000x model complexity, +6% recall, +8% ad quality on selected segments. Meta expects the number of ads to grow a lot with AI tools. | Meta Engineering |
| March 2025 | Andromeda, Lattice and GEM described together, with gains of up to 6% in conversions for Lattice and up to 5% on Reels for GEM | Meta for Business |
| April 2025 | The focus "has shifted from niche targeting to creative diversification". Meta urges advertisers to vary their creative. | Meta for Business |
| September 2026 | A Meta spokesperson called creative variety a "longstanding best practice" and declined to detail how its models treat individual ads | Marketing Brew |
| October 2026 | The next version of Andromeda handles over a billion ads per second and has added 2.4% more Facebook conversions since 2025 | Meta Engineering |
Every performance number in this table comes from Meta. We found no independent test of Andromeda's effect on advertisers' results.
TL;DR: Meta hasn't said how many ads to run, how it handles look-alike ads, or how fast ads wear out now. Anyone who gives you exact numbers is guessing.
These are the most common Andromeda claims, with what Meta has actually said about each.
| Claim | Who says it | Does Meta say it? |
|---|---|---|
| "Creative is the new targeting" | Many agencies and consultants | Close: Meta says focus shifted from niche targeting to creative variety |
| "Run 10 to 50 ads per ad set" | Ad tools and agencies, with numbers from 6 to 50 | No. Meta has given no number |
| "Similar ads are grouped into one 'Entity ID' and only one gets shown" | Ad tools and agencies | No. One tool vendor admits "Entity ID" isn't Meta's term. Meta declined to confirm how it treats similar ads. |
| "Variety beats volume" | Agencies and Meta | Partly. Meta says to vary creative, and also expects ad volume to grow. |
| "Ads wear out faster now" | UGC vendors and agencies | No published rate |
| "Use broad targeting only" | Some consultants | Meta pushes automation, but hasn't said "only" |
| "Hooks matter more than format" | Agencies | No |
| "UGC wins under Andromeda" | UGC sellers, including us | No. Meta publishes creator numbers only for partnership ads. |
We sell UGC ads, so the last row is a claim we'd like to be true. We can't find evidence for it, so we won't make it.
The "similar ads" idea is the one most worth understanding. It makes sense that a system choosing from tens of millions of ads would treat near-copies as one idea. But Meta hasn't published how it decides two ads are "too similar", or what happens next. Thresholds like "60% similar" have no Meta source that we could find.
TL;DR: No. In our data, ad ideas run as 10 or more near-copies didn't last longer than ideas run as a single ad.
We tracked 1,368 Meta video ads from 70 advertisers from launch, including ads that were later switched off. We grouped ads into "concepts": the same advertiser, the same opening copy and the same video length.
| Concept run as | Ads | Advertisers | Still running at 90 days |
|---|---|---|---|
| A single ad | 520 | 60 | 68% |
| 10 or more versions | 132 | 11 | 32% |
Within the same advertiser, there was no real difference (hazard ratio 0.95, 95% CI 0.64 to 1.42). Relaunching a concept later didn't help either (53% against 63% at 90 days). Our part 2 run-length study has the full tables.
So the idea of "make 30 versions of the same ad and let Meta sort it out" isn't backed by this data. That fits Meta's own advice to vary creative, rather than copy it.
TL;DR: Maybe a little. Advertisers with more different ideas kept ads longer, but the gap mostly went away once we compared within the same industry.
For this guide, we ran a new cut. We took the 43 advertisers in our launch data with 10 or more ads, counted how many distinct concepts each ran per 10 ads, and split them into thirds.
| Advertisers grouped by variety | Advertisers | Ads | Still running at 90 days | 95% CI |
|---|---|---|---|---|
| Least varied third | 15 | 476 | 52% | 29% to 74% |
| Middle third | 14 | 465 | 53% | 25% to 85% |
| Most varied third | 14 | 312 | 79% | 65% to 93% |
The most varied advertisers kept ads running 28 points more often (CI +1 to +55). Once we split advertisers into thirds within each industry, the gap fell to 16 points (CI -21 to +55), which is too uncertain to call. Variety in video length showed no pattern at all.
Read this as a weak lean in the direction Meta suggests, not proof. It's 43 advertisers, run length isn't profit, and our "concept" measure uses copy and length, not what the video looks like. All our ads also launched in 2026, after Andromeda rolled out, so we can't compare before and after.
TL;DR: Meta says partnership ads lowered the cost of each extra sale in its own tests. Meta ran those tests, and some popular numbers have no clear source.
Partnership ads run under both the brand's and a creator's name. Meta's numbers on them:
| Claim | Source | How solid |
|---|---|---|
| 45% lower median cost per extra sale and 35% higher extra return, when partnership ads were at least 30% of a campaign's ads | Meta white paper, October 2026, 20 lift studies in Europe, January to June 2025, at 92% confidence (PPC Land summary) | Meta-run; we read a summary, not the paper |
| 19% lower cost per result and 13% higher click-through | Widely repeated, credited to Meta | We couldn't find Meta's original, and accounts of where it came from conflict |
Lift studies compare people who could see the ads with a held-out group who couldn't, which makes them stronger than most ad stats. But 92% confidence is below the usual 95% bar, and Meta is grading its own product.
In our run-length data, partnership ads didn't last longer than brand-page ads (54% against 61% at 90 days, a gap our sample is too small to judge). Our guide to Spark Ads, partnership ads and whitelisting covers how to set them up, including ad codes and creator permissions.
TL;DR: Make different ideas, not more copies. Keep your setup simple, and don't trust anyone selling exact numbers.
TL;DR: No one outside Meta can see how Andromeda treats your ads. Everything else is careful guessing, including ours.
We'll compare a second snapshot of our ad data in December to see whether ads are being switched off sooner.
TL;DR: Our data shows how long ads run, not sales. All the ads we tracked started after Andromeda began.
On Google
Add us as a preferred source and Google shows our UGC ad research more often in Top Stories and AI answers.
Book a 30-minute call. We review your current ads and your offer and see if we're a fit. No retainer required.
Book a Call