Facebook Targeting Myths: What Meta Really Does

Almost everything advertisers repeat about Facebook targeting is wrong, outdated, or both. The myths survive because they sound logical and because forums recycle them faster than anyone checks. Meanwhile Meta has quietly rebuilt how targeting works, and the gap between folklore and reality is now enormous. This article lists the most persistent targeting myths and demolishes each one with what Meta actually says and what the data actually shows. No recycled advice, and no comforting forum folklore dressed up as strategy. Where a claim comes from an agency rather than Meta, it is flagged plainly so you can weigh it accordingly. The goal is simple: stop you wasting budget on rituals that quietly stopped working years ago, and replace them with what the platform itself now recommends.

Myth 1: Precise targeting is king

The oldest belief in the playbook: the tighter you target, the better you perform. It made sense when the auction was dumb and you had to do the matching yourself. It is now backwards. Meta has spent two years moving the matching job from you to the model, and the data follows. Analysis published by Lebesgue across many accounts found broad targeting delivered an average ROAS of 113 percent versus 76 percent for lookalike targeting, with comparable click-through rates. Treat that as agency-reported, not a Meta figure. But the direction matches everything Meta now ships: hand the model room and it usually beats your hand-picked precision. The era when a tight audience was an advantage is over, and clinging to it is the single most common way good budgets quietly underperform.

The reason is structural, not magical. A precise audience starves the learning phase. Every ad set needs roughly fifty optimisation events in seven days to stabilise, and a tiny hyper-targeted pool rarely produces them. So your clever audience sits in Learning Limited, delivering volatile, unreliable results, while a broad audience hits the volume the algorithm needs to actually find your buyers. Precision feels like control. In 2026 it is mostly self-sabotage. The advertisers winning right now give the model a wide field and let it narrow, instead of narrowing it themselves and wondering why nothing scales. Control over the audience was never the source of performance. Volume of signal always was, and a narrow audience strangles it.

Myth 2: Facebook listens to your microphone

This is the most viral targeting myth of all, and it is false. You talk about hiking boots, you see a boot ad, you swear the phone heard you. Meta has denied microphone-based ad targeting since a 2016 blog post, and Mark Zuckerberg repeated the denial under oath before the United States Congress. In October 2025 Instagram head Adam Mosseri told TechCrunch the company does not use your microphone to listen to you, adding pointedly that with the behavioural data it already has, it would not need to. That last line is the real explanation, and it is more unsettling than the myth. The truth is not that Meta cheats with a microphone. It is that it does not have to.

Why does the illusion hold so well? Because the real targeting is invisible and the coincidence is vivid. Independent academic audits and security researchers have found no technical evidence of continuous audio surveillance: it would drain batteries, burn data and trip detection tools. What actually happens is mundane and far more powerful. Your location overlaps with a friend who searched the product, your browsing trail, your purchase history and a million correlated signals let the model predict your interest before you voice it. You did not get an ad because you spoke. You got it because the model already knew. The phone did not hear you. The graph did, and the graph is quieter, cheaper to run and far more accurate than any microphone ever could be.

What the Active Listening pitch deck really proves

In August 2024, 404 Media published a leaked pitch deck from Cox Media Group promoting a product called Active Listening, claiming it could capture voice data from device microphones and listing Facebook, Google, Amazon and Bing as partners. The myth crowd treated this as the smoking gun. It was the opposite. Google removed CMG from its Partner Program. Amazon said it had never worked with CMG on the program and had no plans to. Meta clarified that the deck listed it as a general marketing partner, not a partner in that program. A marketing reseller overselling a fantasy in a sales deck is not proof Meta listens. It is proof a vendor lied to win clients, and the platforms publicly distanced themselves the moment it surfaced. If anything, the episode shows how strong the incentive is to invent a microphone story, and how fast Meta moves to deny one when its name is attached.

Myth 3: You should stack interests to refine the audience

The classic intermediate move: layer interest on interest on behaviour until you have carved out the perfect buyer. Yoga plus organic food plus eco-conscious plus high income. It feels surgical. It is mostly self-defeating. Each layer multiplies a restriction, and the audience collapses to a size that cannot feed the learning phase. Worse, stacked interests narrow the field exactly where the model wants room to explore. You are not refining the audience, you are blindfolding the algorithm and handing it a tenth of the data it needs. The surgical precision is an illusion built on a fragile assumption: that the interest labels themselves are accurate. They are not, as the next myth shows in hard numbers that should end the debate for good.

Myth 4: Interest targeting is reliable

Here is the fact that destroys the entire stacking ritual. A peer-reviewed NC State University study, Analyzing the Accuracy of Facebook Ad-Interest Inference, found that around 30 to 33 percent of the interests Facebook inferred about users were inaccurate or irrelevant. A third. The same researchers showed the system ignores sentiment: a user who posted a negative comment on a Harry Potter page was then tagged with interests in Harry Potter and Daniel Radcliffe. The platform could not tell hatred from love. When you stack three or four interests, you are not narrowing to a clean audience, you are compounding a one-in-three error rate across every layer until what is left barely resembles the buyer you imagined, and the audience you trusted is mostly mislabelled strangers.

Meta itself drew the conclusion the myth-believers refuse to. On 23 June 2025 it began consolidating thousands of detailed-interest categories into broader groupings, merging niche labels like EDM fans, SUV owners and vegan food into wider buckets. Campaigns built on the old granular interests before that date were allowed to run until 15 January 2026, then stopped delivering if they still used removed options. The most precise targeting tool in the platform is being dismantled by the platform. If interest targeting were as reliable as forums claim, Meta would not be retiring it. It is retiring it because the model finds buyers better without it, and because a third of those labels were noise in the first place.

Myth 5: A broad audience is wasted budget

The fear is intuitive: show your ad to everyone and you pay to reach people who will never buy. So advertisers narrow, narrow, narrow to feel efficient. But broad does not mean undifferentiated. With Advantage+ Audience and Advantage+ placements switched on, a broad audience is the model deciding, impression by impression, who is most likely to convert right now, then spending there. You are not blasting everyone equally. You are letting the auction route each euro to the cheapest available conversion across billions of people. Narrow targeting removes that freedom and forces spend onto a pool you guessed at, which is usually worse than the pool the model would have found on its own.

Meta is explicit on its own recommendation. Its guidance points advertisers toward audiences of roughly two to ten million people for most use cases, and warns that smaller, more specific audiences raise costs and trigger learning-limited delivery. That is the opposite of the wasted-budget fear. The platform is telling you, in writing, that going broad is the efficient move and that going narrow is what wastes money on an under-fed learning phase. The waste is not in reaching many people. The waste is in handcuffing the auction so it cannot reach the right ones cheaply, then paying a premium for the privilege of your own restriction, while a competitor running broad quietly undercuts your cost per result.

Myth 6: The ideal audience size is one to two million

This number is repeated like scripture, and nobody seems to know where it came from. One to two million, the Goldilocks zone, not too big, not too small. It is folklore. Meta does not publish a magic figure, and where it gives ranges it points higher, toward the two-to-ten-million band, precisely because the model needs room to optimise. The one-to-two-million rule is a relic of the manual-targeting era, when you were the optimiser and a smaller field was easier to reason about. The algorithm does not reason like you. A field it considers tiny is exactly where it underperforms, and the number you cling to is a comfort blanket, not a strategy.

There is a real nuance hiding under the myth, and it is the only part worth keeping. For a brand-new account with fewer than fifty weekly conversions, a niche market, or a budget under thirty euros a day, a smaller, more defined audience can outperform broad, because the model has too little data to self-optimise and needs a head start. So the honest rule is not a fixed size at all. It is a condition: small and specific while you lack signal and budget, then progressively broad as your conversion volume and Pixel data grow. The number was never the point. The signal volume always was, and once you have it, broad wins decisively.

Myth 7: Always exclude existing customers

It sounds like obvious hygiene: do not pay to advertise to people who already bought. So advertisers reflexively exclude all past purchasers from every campaign. Sometimes that is right. Often it quietly hurts you. Excluding customers makes sense for a one-time purchase, a single course, a wedding dress. It is a mistake for anything with repeat purchase, replenishment or upsell potential. Your existing customers are your highest-converting, lowest-cost audience: they trust you and the Pixel knows them cold. A coffee subscription, a skincare line, a pet-food brand, a B2B tool with seat expansion, all leave revenue on the table by reflexively locking out the people most likely to buy again.

The platform mechanics changed too, which makes the old reflex even less safe. Meta removed the existing-customer budget cap from Advantage+ Sales campaigns, the slider that used to guarantee, say, ninety percent of spend went to new customers. Now you control it with ad-set-level exclusions instead, and as Jon Loomer has documented, those exclusions do not behave the way advertisers assume: audience expansion and certain Advantage settings can still serve to people you thought you excluded. So blanket exclusion is now both strategically questionable and technically leaky. The defensible move is deliberate: exclude only when repeat purchase is genuinely impossible, and cap rather than ban existing customers when it is not.

Myth 8: Retargeting is your free performance lever

Retargeting always looks brilliant in the dashboard. Show ads to people who visited your site or abandoned a cart, and the ROAS comes back huge. So it gets labelled the safe, free win, the lever you pull for guaranteed performance. The dashboard is lying to you, and the lie has a name: incrementality. Retargeting reaches people already at the bottom of the funnel, many of whom were going to buy anyway: the cart was loaded, the intent was set, they were coming back. When your ad takes credit for that conversion, it is claiming a sale it merely witnessed. The ROAS is real on screen and largely fictional in your bank account, which is the most expensive kind of self-deception in the platform.

Incrementality testing exposes this repeatedly. Across studies and holdout experiments, display and social retargeting consistently show low true lift: the conversions would have happened without the ad. A familiar pattern is retargeting spiking during a brand or prospecting push, then taking the credit, when in fact the awareness media created the demand and retargeting simply harvested it. None of this means retargeting is useless. It means it is not free performance, it is demand capture, and it should be measured by incremental lift through a holdout, never by raw in-platform ROAS. Pour budget into prospecting, which tends to show the highest incrementality, and size retargeting to the demand it genuinely creates rather than the demand it conveniently claims. A small, well-measured retargeting layer is fine. A large one propped up by phantom ROAS is how accounts fool themselves into scaling spend that adds nothing.

Myth 9: Lookalikes always beat broad

Lookalikes were the darling of the last era, and the habit lingers: build a one-percent lookalike of your purchasers and you have the cream. In aggregate, that is no longer true. The same Lebesgue analysis that flagged broad at 113 percent ROAS put lookalike targeting at 76 percent, a meaningful gap in favour of broad for most accounts. The reason is the same theme running through every myth here: a lookalike is still a constraint you impose, and the modern model resents constraints. Given a broad field and clean conversion data, it builds a better, faster, ever-updating lookalike internally, without you freezing it into a static seed that ages the moment you create it.

Lookalikes are not dead, though, and the nuance matters. They still earn their place in one specific situation: when you hold data the Pixel has never seen. Offline buyers, loyalty members, high-lifetime-value email segments, CRM lists from channels Meta cannot observe, these make a value-based lookalike genuinely useful, because you are feeding the model signal it could not derive on its own. They also help at smaller budgets where broad lacks the volume to self-optimise. So the corrected rule: default to broad with clean Pixel and CAPI data, and reach for lookalikes when you have proprietary, off-platform signal worth seeding, not as a reflexive upgrade over broad that you apply out of nostalgia.

Myth 10: Audience overlap will cannibalise your campaigns

There is a real effect here, but the popular version is overstated and the prescribed fix often backfires. The myth says any overlap means your own ad sets bid against each other in the auction, inflating your costs, so you must obsessively isolate every audience. Meta has explained that overlap does not cause a literal self-bidding war: its auction is built to avoid showing one advertiser competing against itself for the same impression. The genuine cost of overlap is different and more boring: fragmentation. When several ad sets chase the same people, you split your conversion data into multiple under-fed learning phases, and none of them reaches the fifty-event threshold cleanly.

The fix the myth prescribes, slicing audiences ever finer to eliminate overlap, makes the actual problem worse, because finer slices mean more ad sets and more fragmentation. The correct move is the opposite: consolidate. Merge overlapping ad sets into fewer, broader ones and let the model sort the segments internally. One well-fed broad ad set beats five thin ones squabbling over the same people. This is why account consolidation is the dominant 2026 best practice: fewer campaigns, fewer ad sets, more creatives inside them. You solve overlap not by drawing tighter boundaries but by erasing the unnecessary boundaries you drew in the first place, then letting one strong ad set carry the volume.

The thread running through every myth

Notice the pattern. Every myth here is a leftover from the era when you were the optimiser: stack interests, narrow the audience, isolate every segment, exclude this, retarget that. Each assumes that more manual control equals more performance. The single biggest change in Meta advertising is that this assumption flipped. The model now does the matching, and your manual cleverness mostly gets in its way. The skill that pays in 2026 is not audience engineering. It is feeding the model clean conversion signal through the Pixel and CAPI, giving it a broad field, and pouring your creativity into the ads themselves, because creative is the lever you still fully control.

So before you tune one more interest stack or build one more granular exclusion, ask the better question: can the model actually see my conversions, and have I given it enough room to work? Nine times out of ten, an underperforming account is not a targeting problem, it is a signal problem or a creative problem wearing a targeting costume. Kill the myths, fix the Pixel, go broad, and spend your real energy on offers and creative. That is not where the forums tell you to look, which is exactly why it is where the wins are. The advertisers still chasing the perfect audience are losing, quarter after quarter, to the ones who stopped chasing it and started feeding the machine instead.

Sources

NC State University, Analyzing the Impact and Accuracy of Facebook Ad-Interest Inference (interest inference 30 to 33 percent inaccurate); Meta 2016 blog post and Mark Zuckerberg congressional testimony on microphone denial; TechCrunch, Adam Mosseri statement (October 2025); 404 Media, Cox Media Group Active Listening pitch deck (August 2024) plus Google, Amazon and Meta partner responses; Lebesgue analysis (broad 113 percent vs lookalike 76 percent ROAS); Meta Business Help Center (Advantage+ Audience, recommended 2 to 10 million audience, learning phase); Meta detailed-targeting consolidation and exclusion removal (23 June 2025, deadline 15 January 2026); Meta removal of the existing-customer budget cap in Advantage+ Sales; Jon Loomer Digital (existing-customer exclusion leakage); incrementality testing guidance from Measured, Cometly and fusepoint on retargeting lift. Agency figures are reported by their authors and not independently audited.