What is retargeting, and does it actually work?

You looked at a pair of shoes once and now those shoes follow you across the internet for a fortnight. That is retargeting, and almost everybody has experienced it from the receiving end without ever being told how it works. From the advertiser's side it is sold as the cheapest thing in marketing: 10 times the click rate, half the cost per customer, free to set up. All three of those claims are true and all three are misleading. This is what is actually happening, how to do it, and why the arithmetic comes out against it.

What retargeting is in one sentence, and the thing that actually makes it different: you are not buying a new audience at all, you are buying a second conversation with traffic you have already paid for once

Ordinary advertising buys strangers. The industry word for that is prospecting, and it means exactly what it sounds like: paying to put your ad in front of people who have never heard of you. Retargeting buys a specific list of people who already did something: visited a page, watched a video, added to a cart, installed an app. Everything good and everything limiting about it follows from that one distinction.

• The audience is defined backwards. Normal targeting says "women aged 25 to 40 interested in fitness". Retargeting says "the 812 people who opened my pricing page in the last 30 days". It is a list, not a description.
• You are advertising to your own past traffic. Which means the list can only ever be as big as the traffic you already have. This is the fact that ends up mattering most, and it is the one that gets left out of every guide.
• "Retargeting" and "remarketing" are the same thing. Google historically said remarketing, everyone else says retargeting. Some people use remarketing to mean email specifically. There is no meaningful technical difference.

It comes in seven flavours, and they are not equally available to a small product.

How a platform knows the person scrolling Instagram is the same person who visited your website: a tag on your page hands your visitor's identity to a company that already knows exactly who they are

This is the part that feels like magic and is not. The short answer is that you tell them, and the company you tell is one that already has your visitor logged in somewhere else.

The sequence, for a Meta pixel, in the order it happens:

The same tag does a second job, and it is the one that matters even if you never retarget.

• It reports whichever page it is sitting on. On a product page it says "someone looked at this". On the page a customer lands on after paying, the one that says thank you for your order, it says "someone paid, $34.99". Same few lines of code, different page.
• The amount comes from one extra line you add on that page. The tag does not work out that a sale happened. You tell it. The confirmation page carries the same snippet plus a line naming the event and the value, and your shop software fills the number in. That is the entire difference between reporting a visit and reporting a sale.
• That is the only way Meta or Google ever learns you made a sale. The payment happens on your site, or inside Stripe, where they cannot see any of it. If you do not tell them, they never find out.
• Which is why they can tell you an ad worked. Meta sees the click on your ad, then the purchase from that same browser twenty minutes later, and joins the two together.
• And why they can go and find you more buyers. Once Meta knows which of your visitors actually paid, it can point your next campaign at strangers who resemble the payers instead of at people who merely like clicking ads. That is a tool for finding new people, and it has nothing to do with retargeting.
• A website tag cannot see inside an app. It is JavaScript running in a browser, so it cannot see one single thing that happens in an installed app, including the install itself.
• Apps use a kit instead, and you should install it for the same reasons. You add a small library to the app itself, Meta's SDK or Google's Firebase SDK, and it reports the same kinds of events from inside the app: the app was opened for the first time, someone reached the paywall, someone paid $34.99. Without it Meta and Google have no idea your app was ever installed, cannot tell you which ad caused it, and cannot aim your next campaign at people who install and pay rather than people who tap.
• On Android that works properly. On iPhones it is crippled. Android reports back within minutes. On iOS, unless the person taps Allow on the tracking prompt, and only 15% to 30% do, everything comes back through Apple's privacy system instead: a day or more late, in 64 coarse buckets, and sometimes not at all if too few people converted to clear Apple's threshold. That is why app advertising is measured so much worse than web advertising.

Which explains the two things people find most surprising about it.

• It is not listening to your microphone. It does not have to. You handed over the visit yourself, in exchange for a free analytics and ads product.
• It works across devices. Because the identifier is an account, not a browser. Visit on your laptop at work, see the ad on your phone that night. That only works inside a walled garden where people log in.

Outside the walled gardens it is much shakier, and the difference matters when you choose where to run.

The practical takeaway: retargeting inside Meta, Google or TikTok is a solved problem, and retargeting across the open display web is held together with tape.

Setting it up costs nothing and takes an afternoon, and you should do it even though retargeting is not worth funding, because the very same tag is what makes your ordinary advertising work at all

There is no product to buy and no minimum spend. The steps are the same everywhere:

• 1. Create an ad account. Free. You are not committing to spend anything.
• 2. Copy the tag into every page. One block of JavaScript in the <head>. On a static site that means pasting it into each file; on anything templated it is one edit.
• 3. Mark the pages that mean something. Pricing, checkout, signup, thank-you. These become the segments worth separating later.
• 4. Save an audience. An audience is just a saved filter: "visited any page, last 30 days". It starts empty and fills as people arrive.
• 5. Wait. This is the actual constraint. An audience is a rolling window, so it takes as long as the window to reach full size.

Step 5 is why the timing advice is so lopsided in one direction.

Retargeting is not a Google and Meta feature, it is a pixel feature, and nearly every ad platform ships a pixel. What actually differs between them is the smallest list each one will serve an ad to and how long it remembers a visit, and both of those thresholds changed recently while most guides still quote the old ones. The last column rates installing the free tag, which is worth doing. It is not a rating of spending money on retargeting ads, which this article argues against.

Which visitors are worth chasing, and the two settings that matter more than any bid you will ever place

• Depth beats breadth. "Anyone who visited" is the weakest possible list. Someone who reached your pricing page is worth several times someone who bounced off a blog post, and they should get different money and a different message.
• Recency decays fast. A 7-day audience will outperform a 180-day one on almost every metric. Long windows are for building volume, not for performance.
• Always exclude people who already converted. This is the single most common unforced error. Without an exclusion audience you spend real money advertising a product to the people who already bought it, and they are the most likely to click, which makes your metrics look better while wasting the budget.
• Segment by what they did, not by who they are. The whole advantage here is that you know the action. Throwing that away by lumping everyone into one list gives up the only thing retargeting is good at.

Why retargeting looks miraculous inside Meta and Google's own reporting: it clicks around 10 times more often and converts roughly twice as well as cold traffic, and both of those numbers are real

The statistics quoted in every retargeting pitch are, unusually for marketing, mostly accurate. It is what they are taken to mean that goes wrong.

Notice which row is different. Four of those are observational: they compare people who saw retargeting to people who did not, and those two groups were never comparable to begin with. Only the last is an experiment, and it is the only one that can tell you the ads caused anything.

The measurement trap that flatters retargeting more than any other channel in marketing: the ads are shown to exactly the people who were going to buy anyway, so Meta and Google take the credit for sales that would have happened without them

The mechanism is simple enough to state in one line, and it invalidates most published retargeting results.

• The audience is selected on intent. The people on your retargeting list are, by construction, the people who already came to your site. They were always the most likely to buy.
• The platform then claims the sale. If someone was going to buy on Thursday and they see your ad on Wednesday, the ad gets the credit. Nothing about that is fraudulent. It is just the wrong question being answered.
• It gets worse: they count people who never clicked. Many platforms count a conversion if the ad was merely displayed, without a click, within a day. On a small audience with a high frequency, almost every buyer will have technically seen something.
• So the only honest measurement is a holdout. Randomly withhold ads from a slice of the audience, then compare the two groups. Everything else is a guess dressed up as a measurement.

What a holdout typically finds, using the standard worked example:

The error runs in both directions, which is why "just trust the platform" and "just trust Google Analytics" are both wrong

Meta and Google taking too much credit is the famous story, and it is only half of it. Google Analytics makes the opposite mistake, and it can be just as large.

• Meta and Google claim too much. They count people who only saw the ad and never clicked it, and they assume everyone who bought was persuaded, so the number comes out too high.
• Google Analytics claims too little. In one controlled group experiment, Google Analytics credited remarketing with 1.7% of conversions while the measured lift against the control group was 18.5%, an 11x understatement.
• Both can be true on the same campaign. They are different errors: one counts influence that did not exist, the other refuses to count influence that did.
• Which means the range of "truth" is enormous and nobody can hand you the answer for your own account. You have to run the test.

How each measurement method fails, so you know what you are looking at.

The limit that decides everything else about retargeting: your audience can never be larger than the traffic you already have, and privacy defaults quietly delete about three quarters of it before you can reach anyone

Everything above is about how well retargeting converts. This section is about the ceiling on how many people it can convert at all, and for a small product the ceiling binds long before the conversion rate does.

Start with 10,000 monthly website visitors, which is more traffic than most independent products have, and watch what is left.

Each of those haircuts is separately documented, and they multiply rather than add.

• A US-heavy audience fares much better, because consent banners are mostly not required. The same 10,000 visitors leave roughly 5,000 addressable rather than 2,600.
• Hashed email lists are the one thing that survives all of it. If you have signups, uploading them is worth more than any pixel tuning, because a matched account does not care about cookies, browsers or ad blockers.
• Working backwards from the 1,000-person threshold: you need roughly 4,000 monthly visitors for retargeting to deliver reliably in a consent-heavy market. Below that it will technically run and will not do much.

Frequency and the creepiness cliff: the effect is concentrated in the first day and the first few impressions, and a small audience with an ordinary budget mathematically forces you straight past the point where the ad starts costing you goodwill

This is the failure mode that produces the shoes-following-you-for-a-fortnight experience, and it is almost always an arithmetic accident rather than a decision.

• The effect decays within days, not weeks. In the Sahni, Narayanan and Kalyanam field experiment, 33% of the first week's entire effect happened on day one, and the impact of retargeting fell the longer it had been since the person's original visit.
• Most of the value lands in the first two impressions. Roughly 80% of an ad's weekly impact occurs in the first two views. Conversions generally peak at three to five exposures.
• Past about 5 to 7 views, warm audiences start to sour. At 10 or more exposures, measured negative sentiment rises by about 16%. You are now paying to annoy people who liked you.
• Wear-out is wildly uneven. Across 2.8 billion impressions in 30 natural experiments, four campaigns wore out after one or two exposures while ten showed almost none after fifty. Your creative and category decide which you are.

Now put the small audience and the ordinary budget together, and see why over-exposure is the default outcome rather than a mistake:

What retargeting actually costs, and why its famously cheap cost per customer is partly an accounting illusion that disappears the moment you measure it honestly

On paper retargeting wins every cost comparison. Here is the whole comparison, including the rows the pitch decks leave out.

Throw away the buyers who were going to buy anyway and the cheap price disappears:

Work out what it would cost you specifically: the calculator below turns your traffic into the audience you can actually reach, the number of times your budget will force each of them to see the ad, and what one real customer ends up costing

Three questions decide this, and all three are arithmetic rather than opinion: how many people can you reach, how often will they see you, and what does one real customer cost. Set your own numbers, or use the presets.

What changes if your visitors arrived for free: an organic audience is genuinely additive rather than cannibalising your other advertising, which kills half the argument above, but a free audience still does not make an $80 customer cheaper than a $45 stranger

Everything above quietly assumes you paid for the visit. Plenty of people did not: they rank in search, they got a good thread on a forum, someone told a friend. That is a materially different case and it deserves a straight answer.

• The scaling objection disappears completely, and it was one of the two that mattered. The warning that moving money into retargeting shrinks the very list you are retargeting depends entirely on the list being built by paid traffic. Search results, word of mouth and forum posts keep arriving whether or not you advertise, so spending here takes nothing away from anything.
• The cost objection does not budge. The 27.5% is low because the audience is selected on intent: everyone on it already came to your site, so they were always the most likely to buy. Arriving by a search result rather than an ad click does not make them less likely to have bought anyway. $22 divided by 0.275 is still $80, and the same platform still sells you a total stranger for $42 to $48.
• A free input does not make an expensive output cheap. The audience costs nothing to have. The impressions you serve to it cost exactly what impressions cost.
• What decides it is your price, not where the traffic came from. At a real $80 a customer, the usual rule that a customer should return three times what they cost means that customer has to be worth about $240. The median subscription app earns $21.37 per payer in a year. For a normal consumer app that is a no by a factor of ten. For a $200-a-year business tool it is fine.
• You do get one thing the paid case never has: a free holdout. Nothing about the audience is at risk, so withhold ads from a random 10% to 20% and find out your own number rather than borrowing this article's. That is the only figure that counts and it costs you nothing but a month.

The harder version, where the page they read is about the subject rather than about your product: better odds that the ad did something, much worse odds that anyone buys

Take an accent training app running an article on how accents actually form. The readers are exactly the right kind of person and none of them came to buy anything. Retargeting them is a different proposition again, and it moves in two directions at once.

• Your share of genuinely caused sales should beat 27.5%, and that is the good news. That figure is low precisely because retargeting audiences are people already partway through buying, so the ad collects credit for sales that were coming anyway. Someone reading an explainer is in no purchase process at all and may not know your product exists. There is far less for the ad to take undeserved credit for, so more of what it produces is real.
• Your conversion rate will be far worse, and that is the bad news. The 3.8% retargeting conversion rate above was measured on people who looked at products. An article reader is several steps further back, so the cost per customer starts much higher than $22 before any correction is applied to it.
• The two effects pull against each other and nobody has published the net. Better causation multiplied by a worse conversion rate. Which one wins depends on your subject and your product, which is exactly why the holdout is the entire answer here rather than a formality.
• Do not show them the product they never looked at. Lambrecht and Tucker found specific product ads only beat generic brand ads for people whose preferences had already formed. Article readers are the opposite case by definition. Generic creative is the choice the research actually supports.
• Calling this retargeting flatters it. Someone who read an explainer and never saw your product is much closer to a warm stranger than to a cart abandoner. Judge it against what a stranger costs, not against the $22 benchmark, which was measured on people who were shopping.
• The audience floor bites hardest here. An article pulling 2,000 readers a month leaves roughly 520 you can reach after the privacy haircuts, well under the 1,000 needed for steady delivery. Google will happily run a list of 100. It will not do much with 520.

The three cases side by side, because they are usually discussed as if they were one thing:

Which points at the two things a free audience is actually good for, neither of which is retargeting:

• Use them as a seed to go and find strangers who resemble them. Hand Meta or Google your readers, and especially the handful who paid, and let it build an audience of people who look like them. That pool is the whole internet rather than the 520 people you already have, and it is the part of advertising that can actually grow.
• Take their email while they are on the page. Then reaching them costs nothing, forever, instead of $80 a customer. On an audience you did not pay for in the first place, paying Meta to talk to them is the strangest possible choice.

Every situation people claim retargeting is worth paying for, and what actually happens in each one: seven of the ten are a straight no, and the three that are not are tests rather than plans

These are the ten situations the industry names when it tells you to run retargeting.

If you are going to run it anyway, the setup that wastes the least money

• 1. Install the tag today, even at zero traffic. Free, and it is what lets Meta and Google measure and aim every other campaign you ever run.
• 2. Set every window to the maximum. 180 days on Meta, 540 on Google Display. Narrow later; you cannot widen backwards.
• 3. Email your list yourself before you upload it anywhere. Reaching your own signups costs nothing. Only pay Meta to reach them if emailing them has already failed.
• 4. Build exactly three audiences. Deep visitors at 30 days, all visitors at 180 days, and converters as an exclusion. More segments than that on a small audience just starves each one.
• 5. Cap frequency at 3 to 5 a week, and set the budget to whatever that many views costs for an audience your size, rather than to what you wish you were spending.
• 6. Keep at least 80% of the budget on finding new people. That is the part that actually grows you, and it is what creates next month's retargeting list in the first place.
• 7. Do not use dynamic product ads by default. The research says generic creative usually beats showing people the thing they already passed on.
• 8. Run a holdout before you believe any of it. Exclude a random 10% to 20% for a month. Whatever the gap turns out to be, that is your real number, and it will be well under what Meta and Google tell you. If it comes out worse than your ordinary advertising, switch it off.
• 9. Divide the reported cost per customer by the share your own test says the ads caused before you compare retargeting to any other channel. Otherwise you are comparing a measured number to a flattered one.

The honest bottom line: retargeting does not work, because it costs more than it brings in. The ads do move some people, and that part is properly measured: roughly a 15% lift in return visits in the best controlled study available. But Meta and Google report an effect about four times larger than the one they caused, because the ads are shown to people who had already decided. Strip those people out and the famous $22 customer really cost $80, where the same platform sells you a total stranger for $42 to $48. There is no version of that arithmetic where you come out ahead, and you cannot spend your way past it either, because the list only exists at all thanks to the money you spent on strangers. Install the tag on day one anyway. It is free, and it is what lets Meta and Google tell you which ad produced a sale, aim a campaign at people who buy rather than people who click, and find you strangers who resemble your existing customers. Then put the budget into finding those new people, and treat every retargeting figure Meta and Google show you as an upper bound rather than a result.