how to crack the meta ads andromeda algorithm in 7 steps

if your meta ads suddenly stopped converting over the past year, got more expensive, and your proven system no longer works, nobody hacked your account.

meta replaced the engine that decides who sees your ads. the update is called andromeda. there was no email, no announcement in the ads manager, no push notification. the entire retrieval and delivery system changed underneath every advertiser on the platform, and most operators are still running campaigns like it's 2023.

i've been spending $1M+ a month on meta ads for 7 years. andromeda is the biggest infrastructure change meta has made in that entire period. the operators who adjusted their accounts to align with it are printing. the ones who haven't adjusted are watching their CPLs climb every month without understanding why.

here are the 7 steps to align your account with how andromeda works.

first: what andromeda actually is

think of andromeda as the bouncer at the door before the auction.

the old system: you built an audience, meta showed your ad to that audience, and the auction determined which ads won impression slots.

the new system: before your ad ever reaches the auction, andromeda scans tens of millions of ads and picks a shortlist of creatives that might be relevant to each individual user. the shortlist is based on hundreds of thousands of data points meta has on that person: what they've browsed, what they've engaged with, what their friends are buying, what content they dwell on, and what behavioral patterns predict a conversion for a specific creative type.

if your ad doesn't make the shortlist, your budget, your bid, and your targeting never even get a chance. the auction doesn't happen for you. your money sits there, and your reach shrinks because andromeda didn't think your creative was relevant to anyone in the first place.

this is why creative quality now determines everything. the targeting settings in your ad set are a secondary filter sitting on top of a system that's already decided whether your ad is worth showing to anyone at all.

step 1: consolidate your account structure

one campaign. one objective.

stop splitting budgets across 5 different campaigns. stop creating a new campaign every time you launch a new batch of ads. stop running separate campaigns for different audiences or different creative tests.

meta's algorithm works like an LLM. if you feed claude a thousand documents and ask a question, it searches through the information, identifies the relevant sections, and synthesizes an answer. the more organized the information, the better the output.

andromeda works the same way. it needs to understand your account structure to make good retrieval decisions. five campaigns with split budgets means five separate optimization loops, none of which accumulate enough conversion data to learn properly.

one campaign concentrates all of meta's learning into a single optimization loop. the algorithm gets clearer signal, exits the learning phase faster, and makes better retrieval decisions because the data isn't fragmented.

what to do:

• one campaign with the leads objective (or purchase, depending on your business model)
• one testing ad set for new creatives (ABO, broad targeting)
• one scaling ad set for proven winners
• separate image and video ad sets to avoid CPM bias from mixed formats

step 2: go broad on targeting

turn off every interest stack, every lookalike audience, every layered targeting combination you're running.

meta knows better who should see your ad than you do. this isn't arrogance on meta's behalf. it's math. meta has hundreds of thousands of data points on every user. you have a spreadsheet of assumptions.

the old model: "who should see my ad?" you built the audience based on interests and demographics.

the new model: "which ad should this person see?" andromeda matches creatives to users based on real-time behavioral signals. the creative does the targeting. the ad set restrictions are a handicap, not a filter.

why broad works: advantage+ broad targeting gives andromeda the maximum audience space to find the behavioral pockets that respond to your specific creative. restricted audiences shrink that space and force the algorithm to learn from a smaller, noisier signal pool.

the transition discomfort: the first 5-7 days of switching from interest targeting to broad will look worse. this is expected. the algorithm is exploring a much larger pool. by day 14-21, cost per qualified lead is typically lower than the interest-targeted baseline, with significantly higher volume.

if you want a walkthrough of how we structure the broad targeting transition for clients, DM me "META" and i'll show you the week-by-week framework.

step 3: optimize for the deepest conversion event you can support

stop optimizing for "lead." that tells meta to find people who fill out forms. form-fillers are not buyers.

optimize for the deepest standard event your volume supports:

• "schedule" (booked appointment)
• "complete registration" (finished application)
• "submit application" (qualified lead)

the deeper the event, the higher-quality signal meta receives about who actually converts. the algorithm stops optimizing for people who click buttons and starts optimizing for people who book calls and show up.

the volume requirement: meta needs 50 conversion events per week per ad set to exit the learning phase. if you're optimizing for "qualified lead" and only getting 8-10 per week, the algorithm is still guessing.

the fix: if your volume doesn't support the deepest event, start one level up. optimize for "booked call" instead of "qualified booked call." as volume increases, shift the optimization event deeper.

the critical distinction: send back all the data to the pixel so meta has full visibility. but restrict which events count as optimizable conversions to only the qualified ones. this is how you feed the algorithm clean signal without starving it of data.

step 4: build 25-30 creative angles per batch

this is where most operators fail under andromeda.

one creative concept is one signal to the algorithm. if you're running 5 creatives that are all talking heads with different hooks, andromeda groups them into one container and treats them as one signal. you're not getting 5 chances in the retrieval engine. you're getting one.

25-30 genuinely different creative angles means 25-30 different signals to the algorithm. each one reaches a different behavioral pocket. each one gives andromeda a different matching opportunity.

how to generate 25-30 buying triggers:

don't start with claude. you will think of the first 10-15 buying triggers better than any AI. sit with a 30-minute timer and write down every reason someone would buy your product:

• the problem they're experiencing
• the outcome they want
• the proof they need to see
• the process they want explained
• the offer mechanics (pricing, guarantee, timeline)
• the objections they haven't voiced yet
• the social proof scenarios (what their peers are doing)
• the fear they're avoiding (cost of inaction)
• the aspirational identity (who they want to become)
• the triggering event (what just happened that made them search for a solution)

after the first 15, use AI to fill the remaining gaps by cross-referencing your buying triggers against competitor ads, negative reviews, and ICP research data.

format diversity is equally important:

• founder talking head
• client testimonial / UGC
• AI UGC
• native formats (iMessage, apple notes, google search results)
• static data visualization
• text-on-screen
• green screen
• claymation / stop motion
• split-screen explainer

a batch of 25 ads across 5 format types gives andromeda 25 different signals across 5 different visual containers. the algorithm has maximum room to discover which format-trigger combinations resonate with which behavioral pockets.

step 5: lock your signal data (pixel + conversion API)

the pixel is the algorithm's memory. the conversion API is the mechanism for keeping that memory accurate.

conversion API setup: run the conversion API for all events deeper than page views. deduplicate against the web pixel to avoid double-counting. the conversion API sends server-side data directly to meta, bypassing browser limitations and ad blockers that degrade pixel data quality.

event match quality: check events manager. your event match quality score needs to be above 7.0. below that, meta cannot reliably match the conversion events you're sending back to the actual users who converted. the signal is noisy and the optimization degrades.

the qualified-only signal rule: feed all data to the pixel so meta has visibility across the full funnel. but only count qualified conversions as your optimization event. unqualified leads, bot traffic, and freebie seekers should not train the algorithm on what a "conversion" looks like.

we use an AI qualification layer on some accounts: the lead's application responses are scored against ICP criteria, and only leads scoring above a threshold get sent back as conversion events. this trains the algorithm to find more people who look like the qualified leads, not more people who look like every form-filler.

step 6: run the 72-hour hands-off rule

launch the batch. walk away for 72 hours.

the campaign looking expensive at hour 36 is doing exactly what it should. the algorithm is in exploration mode, testing which creatives perform against which audience segments. the data from the first 72 hours is the learning phase. touching the campaign during this period (pausing underperformers, adjusting budgets, changing targeting) resets the learning and forces the algorithm to start over.

evaluate at day 5-7, not day 2.

at day 5-7:

• identify the top performers by cost per qualified lead (not cost per click or cost per lead)
• move winners to the scaling ad set with 20% budget increases every 2-3 days
• pause the clear losers and log them in the "tested/failed" database
• leave middle performers running through day 14 before final evaluation

step 7: build the feedback loop

this is the step that separates operators who improve from operators who repeat.

after every batch (2-week cycle), aggregate performance by:

• hook category (AUTH, CONT, PAIN, PROOF, MECH, FOMO)
• format type (SELF, HEAD, STAT, UGC, NATIVE, CLAY)
• awareness stage (problem unaware, problem aware, solution aware)

the patterns that emerge across batches tell you which combinations consistently work for your specific offer and ICP. the aggregate data is more valuable than any individual creative's performance because it reveals the structural preferences of your audience.

feed the aggregate data into the next batch's creative brief. the angle brief should prioritize untested triggers, double down on winning format-trigger combinations, and avoid repeating combinations that have already failed.

the feedback loop compounds. each batch is smarter than the last because it's built on more data. by month 3-4, the account has accumulated enough pattern data to predict which creative directions will perform before they launch. that's the edge.

the recap: 3 jobs under andromeda

everything in this piece reduces to three jobs:

1. feed it clean signals. conversion API with deduplication. qualified-only optimization events. event match quality above 7.0. the algorithm is only as smart as the data you give it.

2. give it real creative variety. 25-30 buying triggers per batch. 5+ format types per batch. each genuinely different creative is a separate signal to andromeda's retrieval engine. more signals = more audience pockets discovered.

3. get out of its way. broad targeting. one consolidated campaign. 72-hour hands-off after launch. the operators who try to outsmart andromeda with manual targeting and micro-optimizations are the ones watching their CPLs climb.

andromeda is not a setting you toggle on. it's the engine deciding whether your ad gets seen at all. align your account with how it works, and the results follow.

if you want us to audit your service business and show you how we'd restructure your meta ads account to align with andromeda's retrieval logic, DM me "META" and we'll map out the full rebuild.

($10M+ generated for service-based businesses through meta ads funnels, fully done-for-you)

zack