How the Meta Ads Algorithm Actually Works in 2026 (The 4-Step Auction)

Key Takeaways

Meta runs a four-step auction every time someone scrolls Instagram or Facebook: retrieval, light ranking, heavy ranking, and the auction itself. The whole thing happens in roughly 200 milliseconds.
Tens of millions of ads compete for a single impression. By the time the auction runs, that pool has been narrowed to a few hundred.
Your creative is your targeting. Since the Andromeda update, Meta reads your hook, format, on-screen talent, copy, and landing page to decide who to serve the ad to. Interest targeting and lookalikes barely move the needle anymore.
The equation that decides who wins is public. Total value = (Bid x Estimated Action Rate) + Consumer Value. That second half is Meta protecting the user experience.
The learning phase is overblown. Ads are always learning. Stop killing campaigns on day three because the dashboard is bouncy.
EMQ and Conversions API matter, but not as much as the gurus claim. We have seen ad accounts at a zero opportunity score that are wildly profitable.
Facebook does not always play by its own rules. Ban waves, ad rejections, and account suspensions correlate with Meta’s earnings cycle more than with anything you did.

If you run Meta ads long enough, you start to believe the platform is random. One creative scales to five figures a day, the next one identical to it gets killed in the auction. One ad account runs clean for two years, another gets banned for a creative that looked the same as the one beside it.

Meta is not random. It is a four-step auction with a public equation, run inside a roughly 200 millisecond window every time someone opens Instagram and scrolls. Once you understand how that auction actually works, almost every weird thing you have ever seen in your dashboard starts to make sense.

This is the version I wish someone had explained to me five years ago. The science behind the platform, and the bits that no one writes down because they are not flattering to Meta.

The four-step auction inside every impression

Every time a user opens Facebook or Instagram and an ad slot appears in their feed, Meta runs the same four-step process to decide whose ad fills it.

Step 1: Retrieval

This is the heaviest part of the whole pipeline because Meta is starting with tens of millions of ads and trying to decide which ones are even in the right ballpark for this user.

This is also where the Andromeda update fundamentally changed how targeting works. Before Andromeda, the adset was where you told Meta who to target. Interests, lookalikes, custom audiences. After Andromeda, Meta reads the ad itself to figure out the audience.

It looks at your hook. Your format. The demographic of the on-screen talent. If your ad features a 55 year old woman, Meta skews the targeting toward older women, in context with what the creative is actually saying. It reads your copy, your script, your thumbnail, even your landing page. Then it works out who this ad is trying to talk to.

The reason this update landed is simple. AI creative flooded the platform. The old sorting model could not keep up with the volume of variations advertisers were producing, so Meta retrained the system to read creatives as the signal of intent rather than rely on adset settings.

The practical consequence is everywhere in the accounts we audit. Interest audiences struggle to sustain budget. Lookalikes underperform broad. Most of the high-skill advertisers we talk to run almost no audience targeting at all. Maybe a purchaser exclusion. That is it. Everything else is broad, and the creative does the targeting.

The catch with broad: do not give Meta full control. The algorithm is good, but the suggestions Meta surfaces in Ads Manager (the auto-suggestions, the auto-expansion prompts, the Advantage+ defaults) will quietly drain your budget if you accept them all. Meta is a for-profit company. You are the customer. They want your spend.

Step 2: Light ranking

This barely deserves a paragraph. Light ranking takes the tens of millions of retrieved ads down to a few thousand by stripping out the long tail of clearly poor matches. It is a fast filter that does not need much explaining.

Step 3: Heavy ranking

Heavy ranking is where the equation kicks in. Meta takes the few thousand ads left and narrows them to a few hundred using a value score for each ad. The ad with the higher value score moves on. The ad with the lower one gets served a worse impression or no impression at all.

The equation looks like this:

Total Value = (Bid x Estimated Action Rate) + Consumer Value + User Experience

For a conversion campaign, the estimated action rate breaks down further:

Estimated Action Rate = Estimated CTR x Estimated Click-to-Conversion Rate

So in plain English, to win the auction you need to be willing to pay (bid), have a creative that gets clicked (CTR), and convert the clicks that come through (CVR). And you cannot make the user experience worse for the person scrolling.

Two things to take from this.

The first half is your soft metrics. Meta rewards ads with high CTR over time because click-through is the cleanest signal of relevance. We treat it as a culmination of every engagement metric: CTR, CPC, CPM, hook rate, hold rate, cost per three-second video view. If those are strong, your CPMs drop and your traffic quality goes up.

The second half is your funnel. Click-through is half the battle. The other half is how well you convert that click. If your landing page, offer, and checkout flow leak, your conversion rate drops and your total value score drops with it. The best creative in the world cannot outrun a broken funnel.

The consumer value and user experience layer is Meta protecting the platform from advertisers who would happily destroy it for short-term ROAS. Without that piece, the platform would be 50/50 scam ads instead of whatever it is today.

Step 4: The auction

By this point we are down to a handful of advertisers competing for one impression. This is where bid strategy actually matters.

Most of the accounts we work with run lowest-cost bidding (also called auto-bidding). Meta does the bidding for you and chases the cheapest possible conversion. We have seen ad accounts spending multiple five figures a day and even six figures a day on lowest cost alone, and they are clean, profitable accounts.

We have also seen accounts that only run bid caps and cost caps profitably. Both ends of the spectrum work. The right choice depends on what the actual constraint in the business is. If you cannot scale because your CPA is too high to support the LTV, a cost cap forces Meta to respect that ceiling. If you can absorb a wider range of CPAs and just need volume, lowest cost is usually fine.

What we will not say is “media buy your way from $1k a day to $10k a day.” There is a little room to move with bid strategy, but most of the lift from $1k to $10k comes from creative volume, funnel improvements, and offer work. Think like an entrepreneur, not like a media buyer.

The learning phase, and why we mostly ignore it

Meta will tell you that ads go through a learning phase, exit it, and then settle into a more predictable mature phase. That is the theory.

Reality: your ads are always learning. They never stop. There is no point at which an ad finishes learning and just performs forever. Everything dies eventually. It is the circle of ads.

Initial volatility is real. The first few days of an ad can look bumpy. Sometimes it settles. Sometimes it does not. Sometimes the answer is just to make better ads.

The rule we use with our team and our clients: if you are spending under $1k a day, day-to-day swings that look like a sine wave are completely normal. Zoom out to seven days. Look at the largest data window you have. Make decisions on that, not on yesterday’s number.

We have seen clients panic at “I was at 3x ROAS yesterday and now I am at 1.5x” while spending $100 a day. At that spend level, the variance is the game. Either zoom out or spend more.

If you want consistent day-over-day results, run Google Ads. That platform has its own problems, but consistency is one of its strengths. Meta is an emotional platform run on data generated by emotional humans. Expecting it to behave like a stock chart is the mistake.

Signal quality, EMQ, and Conversions API

Meta optimises against the data you send it. Better data, theoretically, means better delivery decisions and lower cost per result over time. So you want strong server-side tracking through Conversions API, and you want a high Event Match Quality score (EMQ) on your key events, ideally above 7.

Theoretically.

In practice, we have seen ad accounts with EMQ scores near 9.9 that struggle, and accounts at a zero opportunity score in Ads Manager that are spending five figures a day profitably. The correlation between “tick every Meta box” and “actually profitable” is weaker than the platform wants you to believe.

The version of this we run with clients:

Do install Conversions API properly. Server-side dedup, hashed email and phone, FBP and FBC, IP, user agent.
Do aim for high EMQ on purchase events.
Do not optimise your business around Meta’s in-platform diagnostics. Use third-party attribution (Triple Whale, Hyros, custom-built dashboards, or your own SQL on order data) to cross-check what Meta is reporting.
Take Meta’s CPA numbers with a grain of salt, especially after iOS 14. Most of what shows up as “CPA on Facebook” is a reconstruction.

CAPI is a hygiene baseline. It is not a strategy on its own.

Creative diversity (and why variations still work)

There is a narrative going around since Andromeda that you should only test net-new concepts and stop bothering with variations. That is not what we have seen.

Inside an adset, you might launch four ads. Ad one is a new concept. Ads two, three, and four are variations of that concept (different hooks, different opening shots, different copy framings). Ads two through four are variations, but they are still net-new at-bats against the concept.

In our test accounts, the variation often outperforms the original. The new concept turns out to be a loser, and one of the three variations is the actual winner. If you had declared variations dead and only ever launched fresh concepts, you would have killed the concept before the winning variation surfaced.

Run both. New concepts open up new angles. Variations let you refine the ones that work.

The unwritten rule: Facebook does not play by the rules

This is the bit no one writes down. Meta is a for-profit, publicly traded American company. They answer to shareholders, and their enforcement behaviour shifts with the earnings cycle.

The pattern we see across accounts in NZ and Australia (and from peers running accounts in the US, UK, and EU): ban waves, ad rejections, page suspensions, and BM bans cluster in waves. A month where everything is fine is followed by a week where every account is on fire. Then it goes quiet again.

That is not random. It correlates strongly with how Meta’s quarter is going. When the platform needs to tighten policy enforcement to clean up the optics for an earnings call, suddenly creative formats that ran fine for six months get rejected. When the pressure eases, those same formats get approved again. One advertiser posts a creative and it sails through. A different account posts an identical creative and it gets rejected. Why? Because policy enforcement is not the deterministic rule book the documentation pretends it is.

What do you do about it? Question everything. Just because Meta’s own documentation says “do X to improve performance” does not mean X will improve your performance. Just because their head of advertising says “the algorithm rewards Y” does not mean Y holds in your account. Run your own tests. Trust your own data. Be sceptical of anyone, including Meta, who tells you exactly what the algorithm wants.

The number of accounts we have seen accept Meta’s auto-suggestion to “lower your cost per result by 11 percent” only to watch CPA double the following week is not small. The suggestions exist because Meta is trying to push you into spending faster, not because they make your account healthier.

The three things that actually matter

Take the technical layer off for a second. If you strip out all the science, the equations, and the four-step auction breakdown, three things actually matter on Meta in 2026.

1. Your creative is your targeting. Run broad. Stop spending hours building lookalike stacks and custom audience exclusions. Pour that time into more creative tests. The retrieval step reads your ad to decide who sees it. Make the ad clearly speak to the customer you want.

2. Make creatives that stop the doom scroll. High CTR is half the equation. Easier said than done, but if your hook is strong and your format is platform-native, half the battle is already won.

3. Do not destroy the user experience. Bait-and-switch creatives, broken landing pages, dishonest offers. The third part of the auction equation is Meta protecting the platform from advertisers like that. If your ad piles up negative feedback signals, your delivery dies regardless of how good the bid and CTR look.

Three rules. Everything else is detail.

What this means for NZ and Australian businesses

A few specifics for the ANZ market based on the accounts we run and audit.

Broad targeting is more important here, not less. NZ and AU audiences are small. Layering interest stacks on top of small populations chokes delivery faster than it does in the US. We rarely set anything beyond age and gender on adsets for local clients.
Creative volume is the bottleneck for most NZ businesses on Meta. The accounts we see plateau at $200 a day are almost never plateaued because of bid strategy. They are plateaued because they have not tested enough creative angles. The fix is upstream of the auction.
CAPI matters, but most ANZ small businesses have it set up wrong. Half the audits we do find duplicate events, missing parameters, or no server-side setup at all. Fixing this is often the single biggest lever we pull on a new account.
Ban waves hit ANZ accounts on the same cycle as the US. If your account or business manager gets disabled, do not assume you did something wrong. Check the timing. If you peer search and three other people got banned the same week, it is the cycle, not you.

Frequently asked questions

Is creative really more important than targeting on Meta now? Yes, since the Andromeda update. Meta reads your creative (hook, format, on-screen talent, copy, landing page) to decide who to show the ad to. Adset targeting still matters at the margins, but most high-spending accounts run broad and let the creative do the targeting work.

What is the Meta ads auction equation? Total value = (Bid x Estimated Action Rate) + Consumer Value + User Experience. For conversion campaigns, estimated action rate further breaks into estimated CTR x estimated click-to-conversion rate. The ad with the highest total value wins the impression.

Does the Meta learning phase actually matter? Less than the platform claims. Ads are always learning. Initial volatility is normal. We tell clients to zoom out to seven-day or 14-day windows and stop making decisions on day-to-day swings, especially under $1k a day in spend.

Should I worry about my EMQ score? Aim for above 7 on purchase events, but do not treat it as the deciding factor. We have seen ad accounts at zero opportunity score that are wildly profitable, and accounts with near-perfect signal quality that struggle. EMQ is hygiene, not strategy.

Why do my Meta ad accounts keep getting banned for no clear reason? Meta enforcement runs in cycles tied to the earnings calendar. Ban waves cluster around quarter-end and policy-tightening windows. If your account gets disabled around the same time other advertisers are reporting issues, it is usually the cycle rather than something specific to your account.

Should I run lowest cost or bid caps? Either works depending on the constraint in your business. We have run six-figure-per-day accounts entirely on lowest cost and other accounts entirely on bid caps. If CPA needs a ceiling, use a cost cap. If you need volume and can absorb variance, lowest cost is usually fine.

Where to from here

The Meta algorithm is not random. It is a four-step auction with a public equation and a few unwritten rules. Once you stop fighting the platform and start working with how it actually decides which ad to show, the entire job changes.

Stop tinkering with audiences. Start pouring effort into creative volume. Audit your CAPI setup. Zoom out before making decisions. Question every Meta suggestion before you take it. Take the wins where they appear and do not blame the algorithm for things the algorithm did not do.

If you want a hand auditing your Meta account, fixing your Conversions API setup, or running a creative testing system that actually moves the needle, book a call with our team. We work with NZ and AU businesses on paid ads and the connective tissue around them.