Andromeda counts angles, not ads
How Meta's retrieval engine picks your ads, why a new reason to buy beats a new cut, and AG1's angle map vs IM8.
Meta's Andromeda decides which of your ads get a shot at each person. Here is why a new angle beats a new cut, AG1's angle map vs IM8, and a free Claude skill to map any brand's angles.
By Adam Raz
CRO · Ex-Meta · Published
Meta's ad retrieval engine decides which of your ads even get a shot at each person. Fifty versions of the same argument are, in practice, one ad. A new angle is a new reason to buy, and a new chance to reach a different person. Here is how Andromeda works, what AG1's angle mix looks like in Adge data, and a free Claude skill that maps the angles in any brand's ads.
Thousands
Ads kept per person, from millions1B+ / sec
Ads scored (2026 system)+2.4%
Facebook conversions since 202512
Angles in the skillWhat Andromeda changed
Andromeda is Meta's machine learning system for the first stage of ad delivery: retrieval. The 2024 version replaced rule-based shortcuts with a deep model and a hierarchical index of ad clusters. The 2026 version goes further: it scores every eligible ad, using a mixture of specialised user "experts" that a gating network routes each ad to.
Retrieval (Andromeda)
From every ad a person is eligible for, pick a few thousand candidates in milliseconds. If your ad is not picked here, ranking never sees it.
Ranking
Larger models (GEM on Meta's side) predict how likely this person is to act on each candidate and what it is worth.
Auction
Bid, predicted value and ad quality decide which ad is shown.
10,000x
Retrieval model capacity, Dec 2024+8%
Ads quality on selected segments, 2024100x
Faster than the 2024 system, Oct 2026With broad targeting and Advantage+, retrieval reads the ad itself (visuals, words, voice) to decide who it is relevant to. Your creative is now your targeting signal. If your ads all say the same thing, they all signal the same person.
Every eligible ad gets scored
Meta says the new system "scores every eligible ad" at near ranking-level quality. Near-duplicates are not thrown out. Each one is scored, but they score almost the same for the same people, so they compete for one audience instead of opening new ones Inference.
Many user experts, not one profile
The user side is an ensemble of specialised towers built from engagement history, cohorts, LLM-derived semantic features and cross-surface signals, aimed at interests that are "diverse, time-evolving, and contextualized". A gating network routes each ad to the most relevant experts. Distinct angles give it distinct things to route Inference.
The pool is flooded with variants
Meta credits generative AI tools with "hundreds of millions of new ads". When everyone can make 50 cuts of an ad, cuts stop being an edge. A different reason to buy still is Inference.
No advice in the post
The engineering post makes no creative recommendations; it reports +2.4% Facebook conversions since 2025. The advice to diversify creative comes from Meta's marketing guidance, below.
What Meta says vs what the industry says
A lot of Andromeda advice is confident and unsourced. This is the line between what Meta has said and what practitioners infer.
| Claim | Source | How to treat it |
|---|---|---|
| Retrieval narrows tens of millions of ads to a few thousand per request | Meta Engineering, 2024 and 2026 | Fact |
| The 2026 system scores every eligible ad, routing each to specialised user experts | Meta Engineering, Oct 2026 | Fact |
| Creative diversification gets the best results | Meta guidance in its Andromeda materials | Official guidance |
| Ads that lean into different pain points and personas let the system learn who responds to each | Meta business blog, Apr 2025 | Official guidance |
| Guidance is outcome based, not model specific: diversify, test, let the system optimise | Meta spokesperson to Marketing Brew | Fact |
| Similar ads share one "Entity ID" and one ticket to the auction | Practitioners and tool vendors | Useful model, not a documented Meta field |
| Above a 60% similarity score ads collapse into one entity | Third-party blogs | Unverified. Do not plan around a number |
Read the vendor claims as a mental model: near-duplicates compete with each other. The exact mechanics and thresholds are not public.
Angle, not ad
Most teams count ads. Some count concepts. Andromeda rewards a level above both: the angle, the single argument an ad makes for why this person should buy. Everything below the angle is execution, and execution changes rarely move an ad into new territory.
Who it is for
A GLP-1 user worried about losing muscle
The argument for buying now
Medication cuts appetite, so you need complete nutrition in one scoop
The story that carries it
A health coach shows her GLP-1 morning routine
Format, creator, hook, edit
9:16 UGC: "On a GLP? This is your new best friend."
| You changed | New angle? | Why |
|---|---|---|
| The hook line | No | Same argument, new first 2 seconds. Worth testing, but it is iteration. |
| The creator or actor | No | Same argument in a new face. Often clusters with the original. |
| Format or crop | No | Placement coverage, not a new reason to buy. |
| The problem it solves | Yes | Different pain, different person. |
| The moment it matters (GLP-1, travel, new parent) | Yes | A different situation makes it relevant to different people. |
| The proof (expert, customer, celebrity) | Usually | Counts if the proof is the point of the ad. |
| The offer | Sometimes | Offer-led ads are an angle. Ten welcome-kit ads are one angle. |
Write the ad's argument as one sentence with the product name removed. If the sentence is the same as another ad's, it is the same angle, however different the ads look.
Do different personas need different angles?
Mostly yes. Retrieval matches a person to an ad by what the ad is about. A bloated 30-year-old and a 55-year-old on GLP-1 medication are both AG1 prospects, but they are moved by different arguments. One angle can only be the best match for so many people.
When the reason to buy differs: a different problem, moment or objection. GLP-1 users need "protect muscle and nutrients"; travellers need "keep the routine when you leave home".
When the reason is universal and strong (a great offer, a simple routine). Vary the persona in the execution and keep the argument. AG1's "simplify the stack" ideas do this and last longest.
More angles is not automatically better. Each angle needs spend to learn. Ten thin angles on a small budget can lose to three strong ones.
AG1's angle map
We pulled AG1's Meta ads from Adge Nexus for the last 90 days and compared them with IM8, its closest challenger. Adge tags every idea with an angle description and a persona; Claude, connected to Adge, sorted those descriptions into the skill's 12-angle taxonomy.
3,006
AG1 ideas in 90 days3,429
IM8 ideas in 90 days6.1
AG1 effective angles5.3
IM8 effective angles33%
AG1 ads on value & offer57%
AG1 ads in top 3 angles12%
AG1 ideas with no clear angle6.5
AG1 ads per idea (sample)Share of ads by angle
Meta · 25 Aug to 22 Nov 2026 · 250 most-versioned ideas per brand| Angle (AG1) | Ideas | Ads | Share of ads | Ads per idea | Median days live | Live 30+ days | IM8 share |
|---|---|---|---|---|---|---|---|
| Value & offer | 84 | 541 | 33.3% | 6.4 | 17.5 | 42% | 36.9% |
| Life moment | 25 | 217 | 13.3% | 8.7 | 19 | 40% | 4.5% |
| No clear angle | 29 | 165 | 10.1% | 5.7 | 36 | 55% | 11.3% |
| Simplify the stack | 25 | 138 | 8.5% | 5.5 | 50 | 64% | 5.7% |
| Sensory & format | 19 | 125 | 7.7% | 6.6 | 15 | 42% | 1.1% |
| Identity & performance | 18 | 99 | 6.1% | 5.5 | 55 | 72% | 10.1% |
| Authority & science | 14 | 94 | 5.8% | 6.7 | 33 | 57% | 8.6% |
| Celebrity & ambassador | 13 | 76 | 4.7% | 5.8 | 38 | 62% | 0.6% |
| Us vs them | 9 | 73 | 4.5% | 8.1 | 21 | 22% | 4% |
| Problem & pain | 7 | 57 | 3.5% | 8.1 | 43 | 57% | 11.1% |
| Social proof | 7 | 42 | 2.6% | 6 | 28 | 43% | 6.1% |
A third of AG1's ads (33%) and 37% of IM8's argue on price, bundles and free gifts Data. A dozen welcome-kit ads are close to one ad to retrieval. The offer is not the problem; running it as the main angle is Inference.
Life moments (GLP-1, travel, military), taste and format, and celebrity carry 26% of AG1's ads vs 6% of IM8's Data. These likely reach people IM8's ads do not Inference.
IM8 puts 11% of ads on a specific problem (gut, energy, sleep) and 10% on performance identity. AG1 puts 3.5% on problem-led ads Data. That is AG1's clearest white space Inference.
Basis: Adge Nexus, Meta public ad library, 25 Aug to 22 Nov 2026. Sample: each brand's 250 most-versioned ideas (AG1 1,627 ads, IM8 3,102 ads), so ads per idea is higher than the brand-wide figure. Angle labels: Claude classification of Adge's per-idea angle descriptions; 10 of the 12 labels were detected (Desire & outcome and Objection buster need the full ad text). Effective angles = 1 / sum of squared ad shares. Days live is a proxy for a winner, not proof. No spend data exists for competitors.
Where AG1 versions hardest, ideas die fastest
The pattern the skill flags: angles where a brand makes many versions of each idea, but the ideas do not stay live. That is volume going into variants retrieval likely treats as one ad.
| AG1 angle slice | Ideas | Ads per idea | Median days live | Live 30+ days |
|---|---|---|---|---|
| GLP-1 and weight-loss medication | 11 | 9.9 | 12 | 0 of 11 |
| Bloating and digestion | 10 | 9.5 | 13 | 3 of 10 |
| All 250 sampled ideas | 250 | 6.5 | 28 | n/a |
| Simplify the stack | 25 | 5.5 | 50 | 64% |
| Identity & performance | 18 | 5.5 | 55 | 72% |
AG1 makes about 10 versions of each GLP-1 and bloating idea, then retires them in under two weeks. It makes about 5 versions of its simplicity and performance ideas, which stay live 50+ days Data. The timely angles are being versioned, not re-angled. The fix is new arguments inside the GLP-1 persona (protein and muscle, nausea and taste, cost vs separate supplements), not more cuts of the same one Inference.
AG1's persona spread
Adge tags each ad with the persona it speaks to. Across AG1's 2,994 tagged ads in the window, ten personas each get at least 5% of creative. One persona, the overwhelmed supplement skeptic, gets more than a fifth.
AG1 share of tagged ads by persona
Meta · 25 Aug to 22 Nov 2026 · Adge persona tagsAG1 runs a defined set of about ten personas Data. The gap is depth, not breadth: midlife energy (4%) and sleep (under 1%) are thin, while IM8 puts 7% of its ads on its "midlife adult fighting decline" persona alone Data.
IM8 shows 63 persona tags to AG1's 20, but persona labels are generated per brand, so the count partly reflects how finely they were split. Compare shares on matching personas, not raw counts.
The Creative Angle Analyzer skill
One Claude skill that does what you just saw, on any brand. Paste Ad Library ads or an export, or connect Claude to Adge, and it returns the angle mix, a diversity score, the versioning trap, a persona x angle grid and three angle briefs.
Answers: How many real reasons to buy is this brand giving Meta, who does each one reach, and where is it just making the same ad again?
The skill instructions (SKILL.md)
--- name: creative-angle-analyzer description: Classify every ad idea a brand runs on Meta by its creative angle (the argument for buying, not the format), map angles to personas, and score how diverse the mix is for Andromeda. Use when someone shares Ad Library ads, an ads export or a creative list and asks about angles, personas, creative diversity, Andromeda, Entity IDs, or "are we just making the same ad again". --- # Creative Angle Analyzer Andromeda, Meta's ad retrieval system, picks a few thousand candidate ads per person from tens of millions. Variants that look and say the same thing compete with each other for the same slot. A new hook, creator or crop on the same argument is a variant. A new angle is a new reason to buy, and it can reach a different person. This skill measures how many real angles a brand runs and how hard it works each one. ## Definitions (use these exactly) - **Persona**: who the ad is for (a person with a situation and a want). - **Angle**: the single argument the ad makes for why that person should buy now. One per idea. - **Concept**: the story or scene that carries the angle (a skit, a review, a founder talk). - **Execution**: format, creator, hook line, edit, crop, CTA. Changing only these never makes a new angle. Test for a new angle: if you wrote the ad's argument as one sentence with the product name removed, would it be a different sentence? If not, it is the same angle. ## Angle taxonomy (use exactly these labels) 1. Problem & pain: names a specific problem the product fixes (bloating, low energy, poor sleep) 2. Desire & outcome: sells the end state, not the problem (glow, calm, strength) 3. Simplify the stack: replaces many products, steps or decisions with one 4. Value & offer: the deal is the argument (discount, free gift, bundle, trial, guarantee) 5. Authority & science: an expert, study, ingredient or mechanism carries the claim 6. Social proof: real customers, reviews, numbers of users, before/after 7. Celebrity & ambassador: a known name carries the claim 8. Us vs them: compares against a named or implied alternative 9. Life moment: a specific situation makes it relevant now (GLP-1, travel, new parent, midlife, season) 10. Identity & performance: who you become or belong to (athlete, high performer, professional) 11. Sensory & format: taste, texture, ritual, form factor is the reason 12. Objection buster: answers a doubt head on (skeptic turned fan, "I tried to hate it") - No clear angle: product shown or unboxed with no argument. Count it; it is a finding. ## Steps 1. De-duplicate ads into ideas (same core visual or video and same concept = one idea). Keep each idea's number of ads (versions) and days live. 2. For each idea write one line: persona, angle sentence, angle label (one primary, at most one secondary), concept, execution. 3. Angle mix: per angle label count ideas, ads, share of ideas, share of ads, ads per idea, median days live, share of ideas live 30+ days. 4. Diversity score: - Distinct angles with at least 5% of ads. - Effective number of angles = 1 / sum(share of ads per angle squared). Under 4 is concentrated; 6+ is diverse. - Top 3 angles' share of ads. Over 70% is concentrated. - Share of ideas with No clear angle. Over 15% is waste. 5. Versioning vs staying power: flag angles with ads per idea above the brand average AND median days live below the brand median. That is volume going into variants Andromeda likely treats as one ad, on angles that are not sticking. 6. Persona x angle grid: rows = personas, columns = angles, cells = idea counts. Empty cells next to big personas are white space. 7. If a competitor is supplied, run steps 3 to 6 for it and show the differences in share of ads per angle. ## Output - Scope line: brand, platform, window, number of ideas and ads, source - KPI row: Ideas | Ads | Ads per idea | Effective angles | Top 3 angle share | No clear angle % - Angle mix table, sorted by share of ads - Versioning vs staying power: the flagged angles, one line each - Persona x angle grid - 3 angle briefs: a new angle (or an under-used one) for a named persona, each with the angle sentence, why now, and a first-3-seconds idea - 3 bullet read, each with a basis label: (Data) from counts, (Adge data) for platform norms, (Inference) for your read ## Rules - One angle per idea. If you need two, pick the one the first 3 seconds sells. - Never call a format, creator, or hook style an angle. - Days live is a proxy for a winner, not proof. Say so once. - Ad Library data has no spend or results. Never present share of ads as share of spend. - "Entity ID" is an industry term for how Andromeda groups similar ads, not a documented Meta field. Do not quote similarity thresholds as fact. - No em dashes.
Run it on your own account first, then on your top two competitors with the same window. The angles they own and you do not are your briefs. Re-run monthly: a falling effective-angle score is an early warning before CPA moves.
When more angles is the wrong fix
| Situation | Why angle diversity will not save it | Do this instead |
|---|---|---|
| Small budget | Each angle needs enough delivery to learn. Spreading thin starves all of them. | Run 3 to 5 strong angles, add one at a time. |
| Weak offer or product fit | New arguments for a product people do not want still lose. | Fix the offer and landing page first. |
| Every ad looks identical | Retrieval reads visuals and audio too. A new angle in the same template may still cluster with the old ones. | Change the concept and look along with the argument. |
| Angle is right, execution is weak | The argument never lands if the first 2 seconds fail. | Keep the angle, test hooks. |
| Chasing a competitor's angle | Their angle may be a low-spend test nobody switched off. | Use days live and versioning as signals, then test. |
Frequently asked questions
What is Meta Andromeda?
Meta's machine learning system for ad retrieval, the first stage of delivery. It narrows every ad a person is eligible for down to a few thousand candidates before ranking and the auction. Meta introduced it in December 2024 and published a next-generation version in October 2026.
Is an Entity ID a real Meta metric?
Not as a documented field. It is a practitioner term for how Andromeda appears to group near-identical ads. Meta's October 2026 notes say the new system scores every eligible ad, so "one ticket per cluster" is a simplification. The practical effect still holds: near-identical ads score alike and compete for the same people.
What is the difference between an angle and a concept?
The angle is the argument (why buy). The concept is the story that carries it (a skit, a review, a founder talk). One angle can run through many concepts; a new concept on the same argument is not a new angle.
How many angles should a brand run?
Enough to cover your main personas with their real reason to buy. In the skill, an effective-angle score under 4 is concentrated and 6 or more is diverse. AG1 scores 6.1 in its most-versioned ideas.
Can I run the skill without Adge?
Yes. Paste ads from the Meta Ad Library or an Ads Manager export and Claude will classify them. Connected to Adge, it runs on every idea a brand has launched, with personas and days live already tagged.
Does share of ads mean share of spend?
No. Competitor data comes from public ad libraries, which show creative, not spend or results. Read share of ads as where a brand puts its creative effort.
Sources
- Meta Engineering: Andromeda, next-gen personalized ads retrieval (Dec 2024)
- Meta Engineering: Andromeda, model-system co-design (Oct 2026)
- Jon Loomer: Meta Andromeda and creative diversification
- Marketing Brew: advertisers and Meta's creative diversification guidance
Data: Adge Nexus (Meta public ad library), pulled 22 Nov 2026.
Want this applied to your account?
20 minutes, no slides: your market, the first ads we'd make for you, and your pricing.
Get your ad ideas