For roughly a decade, running Meta Ads Manager well meant knowing where the levers were. You picked the interests. You stacked the lookalikes. You built the exclusions, split the ad sets, and set the bids. The advertiser who knew the interface best usually won, because the interface was where the decisions lived.

That is no longer the arrangement. Meta has spent the last two years moving those decisions inside the machine, and in December 2024 its engineering team published the technical account of how far it went. The levers you used to pull are still visible in the interface. Most of them no longer do what you think.

Key Takeaways

• Meta runs six campaign objectives under ODAX, not the legacy eleven, and “Conversions” is an optimization event inside Sales rather than an objective of its own.
• Andromeda, Meta’s retrieval engine, decides which ads are even eligible for the auction. Your bid cannot rescue an ad that never gets retrieved.
• Because retrieval reads your creative, creative is now doing the work targeting used to do. Cosmetic variations of one concept are treated as one thing.
• Advantage+ Shopping was replaced by Advantage+ Sales in February 2025 and deprecated across all API versions in Q1 2026. You cannot build new ones.
• Dual Pixel and CAPI tracking is the entry fee, not an advanced tactic. Signal quality feeds the same system that decides your retrieval.
• Structure, signal, and creative volume are what you still control. Spend without them buys you nothing.

The practical consequence runs through everything below. Your job is no longer to tell Meta who to find. It is to feed the system data clean enough and creative diverse enough that it can find those people itself, and to keep your hands off the wheel long enough for it to learn. This guide covers the account setup that makes that possible, the tracking that makes it accurate, the structure that concentrates it, the creative that drives it, and the reporting that tells you whether any of it worked.

The engine underneath: what Andromeda changed

Start here, because the rest of the platform only makes sense once you understand this.

Meta’s ad delivery is a multi-stage pipeline. Retrieval is the first step, and it narrows tens of millions of ad candidates down to a few thousand relevant ones before larger ranking models decide what a person actually sees. Retrieval has always existed. It was just dumb.

Before Andromeda, Meta’s retrieval systems could apply only limited personalization, relying on isolated model stages and a large number of rule-based heuristics to manage the volume of ads. Rule-based heuristics is the important phrase. Those rules were largely your rules: your interests, your demographics, your lookalike seeds. You defined a gate, the system checked whether a user matched it, and matching ads went to auction. Targeting was a hard constraint because retrieval was too crude to be anything else.

Andromeda replaced that with a deep neural network. Meta reports a 10,000x increase in model capacity, a 6% recall improvement, and an 8% ads quality improvement on selected segments, running on NVIDIA Grace Hopper hardware and Meta’s own MTIA accelerators. The system now organizes ads into a hierarchical index whose layers are trained jointly with the retrieval models, rather than filtering by rules.

Why Meta built it matters as much as what it built. The stated pressure was volume: Advantage+ automation increased the number of eligible ads, and more than a million advertisers used Meta’s generative AI tools to create over 15 million ads in a single month. The old retrieval stage could not evaluate that many candidates under latency constraints. So Meta built one that could, and in doing so removed the reason your targeting inputs had to be hard gates.

Three things follow, and they are the reason this guide is structured the way it is.

Retrieval is a gate you can lose at silently. If your ad is not retrieved for a given person, nothing downstream matters. Your bid is irrelevant. Your budget is irrelevant. The auction happens without you, and Ads Manager will not tell you this happened.

Your creative is now the input retrieval reads. The system evaluates what is actually in the ad and matches it against behavioural patterns. Your audience settings inform that process, but they no longer command it.

Similar ads compete with each other rather than expanding your reach. The hierarchical index clusters semantically similar creative together. Fifty variations on one concept occupy roughly the position of one concept. This is the single most expensive misunderstanding in Meta advertising right now, and we’ll come back to it.

One honest note. Meta’s own numbers describe the system’s performance, not yours. Common Thread Collective’s analysis of 3,014 advertisers found ROAS declined about 7% on average during the rollout, with CPMs up 28% year-over-year in Q1 and click-through rates actually climbing 43% while conversion rates fell. More, cheaper clicks converting worse is what a retrieval system optimizing for engagement signals looks like when your creative and signal quality are not feeding it well. There are also credible practitioners who think the anxiety is overdone. Performance consultant Dara Denney has argued the algorithm matters far less than macro-economic conditions, and that consumers saving and trading down explains more of what advertisers are seeing than any retrieval change.

Both things are true. The mechanism is real and worth building around. It is not the only thing happening to your account.

Setting up: building the container properly

Business Manager at business.facebook.com is where everything lives: pages, pixels, catalogs, users, billing. Get the container right, because two of these decisions are permanent.

Currency and time zone cannot be changed. They determine when your daily budgets reset and how your billing thresholds trigger. Changing either later means a new ad account, and a new ad account means starting your account history over. Choose them as though you cannot revisit them, because you cannot.

Connect your Facebook Page and Instagram account so creative can run across every surface. Then complete business verification. Meta’s enforcement is automated and its appeals process is thin, so an unverified legal entity is both easier to flag and harder to rescue.

A word on Boost Post, since it’s usually the first question. Boosting is not a trap, it’s a scope decision. It fires your Pixel and it will report conversions. What it will not let you do is select the optimization event, and that single setting is what determines who the system looks for. Everything in this guide about signal and retrieval depends on choosing what you’re optimizing toward. Boost Post doesn’t offer the choice. That makes it reasonable for putting money behind an organic post already earning attention, and unsuitable for anything measured against revenue.

Signal: the Pixel, CAPI, and Event Match Quality

Here’s the connection most guides miss. Tracking is not just measurement. The conversion data you send back is training data for the system deciding which ads get retrieved. Bad signal does not merely give you bad reports. It gives the machine a distorted picture of who your buyers are, and the machine acts on that picture.

The Meta Pixel captures browser-side events. The Conversions API sends events server-side from your systems directly to Meta. You need both, deduplicated by event ID, because browser-side collection has been eroding since Apple shipped App Tracking Transparency with iOS 14.5 and has kept eroding through ad blockers and browser privacy defaults.

This is now table stakes rather than sophistication. CAPI adoption reached roughly 89% of active Meta advertisers by late 2025, up from 67% in 2024. If you are in the remaining tenth, you are not being conservative, you are competing against a system that has better information about your competitors’ customers than about yours.

Event Match Quality is the number to watch in Events Manager. It rises as you send more hashed customer parameters with each event: email, phone, name, city, IP, click ID. Higher EMQ means more conversions attributed to the ads that actually caused them. Treat a score below 6.0 as a defect requiring engineering attention, not a metric to note and move past.

Configure Aggregated Event Measurement too, prioritising your eight most valuable web events per verified domain.

There’s a broader pattern worth noticing here. The discipline of feeding structured, high-quality signal to a machine that decides your visibility is not confined to Meta. The same logic now governs whether AI systems surface your brand at all when someone asks a question in a chat interface rather than a search box, which is the shift Pixis examines in its work on multimodal AI and creative signals. Different surface, identical underlying problem: the systems that used to match keywords now read content, and the quality of what you feed them decides what they do with you.

Structure: concentrating the signal

Three levels, and the logic connecting them has inverted since 2022.

Campaign holds the objective and, with Advantage+ Budget, the budget. Ad set holds audience, placement, schedule, optimization event, and bid strategy. Ad holds creative.

The old instinct was to segment: separate ad sets per audience, per placement, per angle, so you could read the data and control the spend. That instinct was correct when you controlled targeting. It is now actively harmful, for a mechanical reason.

Segmentation splits your conversion signal into portions. Each portion has to independently reach the volume the algorithm needs to learn. Five ad sets generating ten conversions a week each are five ad sets that never exit learning, whereas one ad set generating fifty exits cleanly. Meta’s own testing found one ad set with 25 diverse creatives produced 17% more conversions at 16% lower cost than five ad sets with five creatives each.

The rule of thumb: fewer campaigns, fewer ad sets, more distinct creative inside them. If you are running more than three or four campaigns and none is reaching twenty conversions a week, you have a fragmentation problem, not a performance problem.

Objectives: six, not eleven

Meta runs the Outcome-Driven Ad Experiences (ODAX) framework. The six objectives are Awareness, Traffic, Engagement, Leads, App Promotion, and Sales. The old eleven-objective list was consolidated in 2023, new campaigns could not use legacy objectives from January 2024, and existing legacy campaigns had to be migrated by the end of that year.

If a guide tells you Meta has eleven objectives, it was written before 2023 or copied from something that was. “Conversions” is not an objective either. It is an optimization event you select inside Sales.

Funnel stage	Objective	Optimization event	Primary metric
Top	Awareness	Reach or ad recall lift	CPM, reach
Upper-mid	Traffic	Landing page views	Cost per landing page view
Mid	Leads	Instant form or website lead	CPL and downstream lead quality
Bottom	Sales	Purchase	ROAS, CPA
Retargeting	Sales or Leads	Purchase or lead	CPA, ROAS

The objective is the instruction the whole system optimizes against, and it is followed literally. Select Traffic while wanting purchases and Meta will reliably find people who click and do not buy. It is not misbehaving; it is doing precisely what you asked.

The harder decision sits inside Sales: which event to optimize for. The purist answer is the deepest event you care about, usually Purchase. The practical answer is the deepest event you can generate volume for. Optimizing for Add to Cart when you want purchases fills the funnel with browsers. Optimizing for Purchase when you generate twelve a week leaves you in learning permanently. Sustainability beats purity, and if your average order value is high or your sales cycle long, a higher-frequency proxy event is the honest choice rather than a compromise.

Targeting: from command to context

Everything in this section is downstream of Andromeda.

Core Audiences still exist and still narrow. Meta has been removing detailed targeting options and exclusions since March 2025, and in September 2025 introduced restrictions on lookalikes and source custom audiences built from health and finance data.

Custom Audiences remain genuinely useful, and their most valuable application is now exclusion rather than inclusion. Telling the system who not to serve is still a hard instruction it respects.

Lookalike Audiences have been downgraded rather than killed. The behavioural signals the system already holds generally exceed what a seed list can describe, so a lookalike constraint mostly limits delivery without improving match quality. The exception is genuinely small accounts with thin conversion history, where a tight lookalike still functions as early scaffolding.

Advantage+ Audience is the default, and the mechanic that matters is this: your inputs are suggestions, not constraints. The system may deliver outside your stated ages, interests, and locations when creative signals point somewhere better. If you require a hard boundary for regulatory or category reasons, Advantage+ Audience will not enforce it, and you need a different approach.

The workable pattern for most accounts is broad prospecting through Advantage+ Audience, Custom Audience exclusions to stop paying for people you already have, and separated retargeting only where the economics justify the fragmentation cost.

Creative: the part that is now targeting

This is where the work moved.

If retrieval reads your creative to decide who sees it, then every creative is a targeting decision. Not a message that gets delivered to a target, but the thing that determines the target. Meta’s data science team has put creative quality at roughly 56% of campaign performance outcomes, which is more than targeting, budget, placement, and timing combined.

The mechanical implication is the one advertisers keep missing. Because the index clusters semantically similar ads, variations are not diversity. Twenty ads that are one concept with different headlines occupy roughly one position in the index. Those twenty ads compete against each other for a single slot rather than earning twenty entries. Creative similarity above roughly 60% triggers suppression, and cosmetic differences that are obvious to you are invisible to a system reading semantic meaning.

Real diversity varies at the structural level: different hooks, different problems, different emotional registers, different formats, different people, different visual worlds. A useful discipline is to force each concept to occupy a distinct combination of who this is for, what outcome they want, and how aware they already are of the problem. If two ads answer the same question the same way, they are the same ad regardless of how different they look to you.

On volume, the honest position is that the specific numbers circulating are directional rather than precise, and mostly come from agencies with a product to sell. Brands testing 20 or more new ads monthly reportedly achieve around 65% higher ROAS than those testing fewer than ten, and UGC-style creative reportedly outperforms polished brand assets by roughly 48% on CTR. Treat these as evidence of direction, not targets to hit. The defensible version is simpler: you need more genuinely distinct concepts than you needed in 2022, and the constraint on most accounts is production capacity rather than strategy.

Fatigue has compressed too. Reported effective lifespans have fallen from roughly six to eight weeks down to two to four, which follows logically from better matching. A system that finds your optimal audience faster also exhausts it faster.

On placements: use Advantage+ Placements. Restricting placements narrows the pool the auction can buy from, which raises your costs to buy control you no longer meaningfully have. The exception is a placement producing genuinely low-quality results, where a value rule to lower bids is a lighter instrument than exclusion. Design mobile-first and vertical: Instagram Explore merged into Reels in January 2026, pushing still more inventory full-screen.

Budget and the learning phase

These are one topic, because the budget question that actually matters is whether you can afford to exit learning.

An ad set exits after approximately 50 optimization events in a rolling seven-day window. The arithmetic that follows: target CPA × 50 ÷ 7 = minimum viable daily budget. At a $40 target CPA, that is roughly $286 per day per ad set. Below that, you will not reach the threshold inside the window, and the algorithm never stabilises. This is the number to run before you build anything, because it determines how many ad sets you can afford to have at all.

Three details about learning that most guides bury:

Only your specific optimization event counts. A Purchase-optimized ad set gains nothing from Add to Cart or Initiate Checkout events, however often they fire. Advertisers routinely see a busy Events Manager alongside a stuck ad set and cannot reconcile the two. This is the reconciliation.

The window rolls. It is any seven-day window, not the seven days after launch. An ad set with 47 events in week one can still exit during week two. Do not panic-edit at day seven.

Resets are expensive and easy to trigger. Changing targeting, the optimization event, bid strategy, or budget by more than roughly 20% restarts learning. Each reset costs an estimated 5 to 15% ROAS the following week, and CPAs during learning typically run 20 to 50% above post-learning averages. The instinct to fix a bad day is the most reliably expensive instinct in Meta advertising.

Strategy	Use it for	Risk
Advantage+ Budget	Scaling ad sets that already work	Low
Ad set budget	Forcing spend onto a specific test	Medium
Highest Volume	Maximum results, CPA unconstrained	Medium
Cost per Result Goal	Holding an average CPA	Medium
Bid Cap	Hard auction ceiling	High
ROAS Goal	Value optimization with sufficient data	High

Note that ad set budgets are no longer an absolute lock: Meta can shift up to 20% of budget between ad sets toward the better performer even when you have set them manually.

If you cannot reach fifty events, you have four options and no others: raise budget, broaden the audience, consolidate ad sets, or move to a higher-frequency event. Learning Limited is not automatically failure. It means volatility, not badness. An ad set stuck in Learning Limited with a stable 30-day ROAS is fine, and the correct response to it is often nothing.

Once stable, scale in 15 to 20% increments. Doubling a budget resets what you spent two weeks buying.

Advantage+ in 2026: what exists and what doesn’t

This is where most published guides are simply wrong, so precision matters.

Advantage+ Shopping no longer exists as a creatable format. Meta renamed it to Advantage+ Sales in February 2025, and the old format was disabled for creation and modification across all API versions in Q1 2026. Legacy campaigns built before the cutoff may still deliver, but you cannot create new ones or make structural edits. Any guide recommending you launch an Advantage+ Shopping campaign is describing a button that is not there.

What actually changed, and it was more than a rename:

Ad sets came back. Advantage+ Shopping collapsed the ad set into the campaign and permitted exactly one. Advantage+ Sales removes the limit, with up to 50 ads per ad set. Given what we now know about creative diversity and retrieval, this is the most consequential of the four changes.

Targeting input returned. The loudest complaint about the old format was that you handed over budget and hoped. You can now supply audience suggestions and custom audience exclusions at ad set level.

The existing customer budget cap is gone. You recreate it with separate ad sets instead.

Scope expanded. Advantage+ automation now covers sales, leads, and app installs rather than catalog commerce alone. Meta expanded it to leads campaigns in 2025, which brought the same automation to service businesses that e-commerce had since 2023.

Meta reports 22% average ROAS improvement for Advantage+ Sales. Read vendor-measured lift on vendor products as directional.

Opportunity Score also became official: a 0 to 100 rating in Ads Manager. Be clear about what it measures. It scores your compliance with Meta’s recommendations, not your profitability, and most of its suggestions are to turn on more Meta automation. It is a useful diagnostic and a terrible target.

The sensible posture on all of it is parallel rather than wholesale. Run Advantage+ Sales alongside manual campaigns, validate against your own CPA before shifting budget, and keep manual structures for the creative testing and audience control that automation still handles poorly.

Reporting, attribution, and the two things you still control

Customise your columns. The default view foregrounds metrics that do not pay salaries.

Metric	What it actually tells you
ROAS	Revenue per dollar, as Meta measures it
CPA	Cost to acquire one conversion
CTR (link)	Whether creative earns the click
Conversion rate	Whether the page and offer hold up
Frequency	Fatigue, once it passes ~3 in prospecting
Cost per landing page view	Whether the page loads before people leave

Use the Breakdown menu to segment by placement, device, time, and demographic. That is where you find the segment quietly consuming a third of your budget at twice target.

On attribution, the important thing is not choosing correctly but choosing consistently. One-day click undercounts. Seven-day click plus one-day view flatters. Neither is truth. Pick one, hold it constant, and reconcile against your own backend, because Meta measuring Meta’s contribution is not a neutral referee. Blended measures like marketing efficiency ratio, total revenue over total marketing spend, are worth watching alongside platform ROAS precisely because they do not depend on Meta’s account of its own work.

Which brings us to the two places you still hold the wheel.

The first is measurement. The algorithm optimizes toward the event you tell it to. If that event is a low-value action, it will efficiently generate low-value actions forever, and the dashboard will look excellent while the business does not improve. Choosing what counts as success is not automated and will not be.

The second is policy, which is genuinely yours because the consequences are. Enforcement is automated and escalates from ad rejection to account restriction faster than most teams expect. The recurring triggers are misleading claims, personal attributes, before-and-after imagery, restricted goods, and sensationalised copy. Special ad categories still apply to housing, employment, credit, social issues, elections, and politics, though the restrictions that previously blocked algorithmic targeting for those categories were lifted during 2025, so Advantage+ Audience is now available to them. Review the policies before your creative team writes, not after Meta rejects.

Where this leaves you

The uncomfortable summary is that Meta took the parts of this job that rewarded interface fluency and automated them, then handed back the parts that require actual production capacity and judgement.

You cannot out-lever the system anymore. There is no interest stack, no audience insight, no bidding trick that beats a well-fed algorithm, because the levers that used to do that work now do something smaller than you think. What you can do is give it clean signal, a structure that concentrates rather than fragments that signal, and enough genuinely distinct creative that it has real options to match against real people. Then leave it alone long enough to learn.

That is a harder job than the old one, not an easier one. It just moved.

FAQs

How many campaign objectives does Meta Ads Manager have in 2026?
Six, under the ODAX framework: Awareness, Traffic, Engagement, Leads, App Promotion, and Sales. The older eleven-objective list was consolidated in 2023 and cannot be used for new campaigns. “Conversions” is an optimization event inside Sales rather than a separate objective.

What is Meta Andromeda and does it affect my account?
Andromeda is Meta’s ads retrieval engine, announced by its engineering team in December 2024 and rolled out across 2025. Retrieval is the stage that narrows tens of millions of ad candidates to a few thousand before the auction. It affects every account, because an ad that is not retrieved never reaches the auction regardless of bid or budget. The practical consequence is that your creative now does the work your targeting settings used to do.

Why did my Meta ads get worse without me changing anything?
The most common explanation in 2026 is that the delivery system changed underneath a setup built for the previous one. Accounts with many narrow ad sets, lookalike stacks, and a handful of near-identical creatives are structured for a system that no longer exists. That said, macro conditions affect conversion rates independently of any algorithm change, so check whether your landing page conversion rate fell alongside your ROAS before concluding the platform is the cause.

How many optimization events do I need to exit the learning phase?
Roughly 50 of your selected event per ad set in a rolling seven-day window. Only that specific event counts, so Add to Cart events do nothing for a Purchase-optimized ad set. Editing targeting, the optimization event, bid strategy, or budget by more than about 20% resets learning, at an estimated cost of 5 to 15% ROAS the following week.

How many creatives do I actually need?
More distinct concepts than variations, which is the part that matters more than any specific number. Because the retrieval index clusters semantically similar ads together, twenty variations of one concept function roughly as one entry rather than twenty. Vary the hook, the problem, the format, and the audience the ad implicitly speaks to. Published benchmarks suggest meaningful gains above roughly 20 new ads monthly, but treat those figures as directional, since most come from vendors selling creative production.

Is Advantage+ Shopping still available?
No. It was renamed Advantage+ Sales in February 2025 and deprecated across all API versions in Q1 2026. Legacy campaigns may still deliver but cannot be recreated or structurally edited. Advantage+ Sales supports multiple ad sets, audience suggestions, and custom exclusions, none of which the old format allowed.

Do I still need the Conversions API if my Pixel works?
Yes. Browser-side collection has been degrading since iOS 14.5 and continues degrading through ad blockers and browser defaults. CAPI adoption reached roughly 89% of active advertisers by late 2025, so running Pixel-only means competing against a system with better information about your competitors’ customers than yours. Signal quality also feeds the retrieval system, so this is an optimization issue rather than only a reporting one.

Should I use Advantage+ Budget or ad set budgets?
Advantage+ Budget for scaling ad sets that have already proven out. Ad set budgets when you need to force spend onto a specific test. Note that Meta can now move up to 20% of budget between ad sets even when you set them manually, so ad set budgets are a strong preference rather than a hard constraint.