Lookalike audiences explained: finding more of your best customers on Google and Meta

A lookalike audience is a model of your best customers

A lookalike audience is a group of people the ad platform thinks resemble a group you already have. You hand it your best customers, it studies what they have in common across everything it knows about its own users, and it finds more people who share those traits. The people it finds have usually never heard of you.

Facebook announced Lookalike Audiences in March 2013, as an addition to Custom Audiences, its tool for uploading your own customer data.¹ Google, LinkedIn, TikTok and others built their own versions since. The idea has survived a decade of privacy changes, but the controls have changed a lot, and several things the earlier version of this article said are no longer true. I say so in each section.

Throughout I use one invented business. Ormonde Candle Company is an online candle shop in Kilkenny, run by Megan, at ormondecandles.example. It has 9,000 customers, 1,200 of whom have ordered more than once, and it spends about €3,000 a month on adverts. Nobody real, and every number in the examples is made up so that every step can be shown.

Worked example: choosing what to clone

Megan could give the platform all 9,000 customers. She could also give it the 1,200 repeat buyers. The first list is bigger. The second describes the customers she actually wants more of: people who came back without being chased. A lookalike copies whatever pattern the seed contains, so the second list asks the right question. The rest of this article works with those 1,200.

Lookalikes sit between interest targeting and retargeting

There are three common ways to choose who sees an advert, and lookalikes fill the gap between the other two.

Interest targeting is where you pick demographics, interests and behaviours you believe your customer has. It is educated guesswork, and the guess is yours.

Retargeting shows adverts to people who already visited your site or engaged with your brand. It converts well, but the pool is only as big as your traffic.

A lookalike reaches cold people, like interest targeting, but the description of who to look for comes from real customers rather than from your guess. That makes it a prospecting tool. Typical uses are scaling a campaign that already works, entering a new region with your existing customers as the template, launching a product to people who resemble buyers of a related one, and building awareness among the right people rather than everyone.

An example: Megan wants to sell a new range of beeswax candles. Interest targeting would mean guessing at "home interiors" and "sustainable living". Retargeting would reach the 30,000 or so people who visited her site recently. A lookalike of her repeat buyers reaches people she has never met who behave like the ones who keep coming back.

How big a lookalike is: a share of a population

Both platforms size a lookalike as a share of the people they can reach in your chosen locations. The arithmetic is a percentage of a total:

$$N = r \times P$$

Here $N$ is the number of people in the lookalike, $r$ is the share you choose, written as a decimal (1% is $0.01$), and $P$ is the number of people the platform can reach in the target locations. The smaller $r$, the closer the people are to your seed, because the platform keeps only its top scorers. Foundations: percentages

Worked example: sizing Megan's lookalikes

Assume, for the example only, that the platform can reach 4 million people in Ireland. The real figure varies by platform and changes all the time, so Megan would read hers off the audience estimate in her account.

A 1% lookalike holds $0.01 \times 4{,}000{,}000 = 40{,}000$ people.

A 5% lookalike holds $0.05 \times 4{,}000{,}000 = 200{,}000$ people.

On Google, a narrow segment aims at 2.5% of users in the target location, so $0.025 \times 4{,}000{,}000 = 100{,}000$ people on the same assumption. Five times the reach for the 5% version, and each extra person is a little less like the seed than the last.

Setting up a lookalike on Meta

Meta's lookalikes have been around longest and are still the most flexible.

The seed, which Meta calls the source or origin audience, must have at least 100 members, and at least 100 of them from one country. If no country reaches 100, a setting called allow_international_seeds lets Meta look for the minimum in another country.² The seed can be any Custom Audience of that size, including a customer list, a website audience built from the Meta Pixel, or an app audience.²

Size is a ratio. Meta's developer documentation gives a range of 1% to 20% in steps of 1%, or two preset types instead: Similarity, the top 1%, and Greater Reach, the top 5%.² A new lookalike takes one to six hours to fill, and Meta refreshes its members every three days while it belongs to an ad set.²

Two corrections to the earlier version of this article. It said a lookalike takes six to 24 hours to be ready. Meta's documentation now says one to six hours. It also told you to pick the countries on the lookalike itself. Since version 27.0 of the Marketing API, Meta no longer accepts a country or location when you create a lookalike. The lookalike delivers to whatever locations the ad set targets, and an ad set without location targeting returns an error.² I have also cut two figures from the old text: that Meta recommends seeds of 1,000 to 5,000 people, and that a 1% lookalike in the United States is about two million people. I could not confirm either in Meta's documentation as it reads today.

There is a better seed than a plain list. A value based lookalike lets you attach a number to each person in a customer list, such as what they spent, and Meta uses it to look for people resembling your most valuable customers rather than any customer. The seed still needs 100 people, each record carries a LOOKALIKE_VALUE, and every ad account using it must accept Meta's terms for the feature.³

One newer limit: since 2 September 2025 Meta has been restricting custom audiences that suggest health conditions or financial status. Such audiences are flagged, cannot run adverts, and cannot be used as a lookalike seed.² A list called "customers who bought our debt advice guide" is the kind of seed that will now be refused.

Worked example: Megan's value based seed

Megan exports her 1,200 repeat buyers with their total spend. Sam, a customer since 2023, has spent €410. Priya has spent €95. In the upload, Sam's row carries a LOOKALIKE_VALUE of 410 and Priya's 95. The lookalike now leans towards people resembling Sam. In Ads Manager she opens Audiences, chooses Create audience, then Lookalike audience, picks the value based list as the source and chooses 1%. She sets the location to the United Kingdom on the ad set, not on the lookalike.

Advantage+ audience turns your lookalike into a suggestion

This is the change that matters most since the earlier version of this article, and it changes what a lookalike means on Meta.

Advantage+ audience lets Meta's delivery system search beyond the targeting you give it. Meta's developer documentation describes it as expanding the pool the system can search, subject to constraints it will not cross: location, minimum age, language and custom audience exclusions.⁴ Since version 23.0 of the Marketing API, a new ad set with default or relaxed targeting is opted in automatically, and anything else must set the option explicitly to on or off.⁵

"Relaxed" is the key word. With Advantage+ audience on, a lookalike or custom audience you include is marked as one Meta may expand beyond.⁴ So your carefully built 1% lookalike becomes a hint about where to start looking, not a fence. Exclusions stay firm, which is why excluding existing customers still works.

Worked example: keeping the fence or dropping it

Megan wants to test whether her 1% lookalike genuinely beats Meta's own search. She runs two ad sets. In the first, she switches Advantage+ audience off, so the lookalike is a hard boundary:

{
  "geo_locations": { "countries": ["GB"] },
  "custom_audiences": [{ "id": "<lookalike_audience_id>" }],
  "excluded_custom_audiences": [{ "id": "<all_customers_id>" }],
  "targeting_automation": { "advantage_audience": 0 }
}

In the second, she switches it on and marks the lookalike as relaxed, so it becomes a suggestion:

{
  "geo_locations": { "countries": ["GB"] },
  "custom_audiences": [{ "id": "<lookalike_audience_id>" }],
  "excluded_custom_audiences": [{ "id": "<all_customers_id>" }],
  "targeting_automation": { "advantage_audience": 1 },
  "targeting_relaxation_types": { "lookalike": 1, "custom_audience": 1 }
}

Both exclude her existing customers, because exclusions are one of the constraints Meta does not expand. If the second ad set wins, the lookalike was useful as a starting signal but not as a boundary. That is a perfectly good answer.

Google retired similar audiences in 2023

Google's history is shorter and messier. Its original feature, similar audiences, was generated automatically from your remarketing lists. Google stopped creating new similar audience segments in May 2023 and removed them from all campaigns in August 2023, keeping the historical reports.⁶

Google did not replace them with one feature. It pointed each campaign type somewhere different: optimised targeting for what were then Discovery, Display and Video action campaigns, audience expansion for Video reach and consideration campaigns, and Smart Bidding with your own data for Search and Shopping.⁶ Lookalike segments arrived later as a separate tool, which the next section covers.

Customer lists add one more gate on Google. Using Customer Match lists for targeting needs 90 days of Google Ads history and more than 50,000 US dollars of lifetime spend. Accounts below that, in good standing on policy and payments, can still use lists for exclusions and in observation.⁷ Google's lookalike pages accept customer lists as seeds but do not say how that threshold applies to them, so check in your own account before you build a plan around it. How the lists themselves are prepared and kept fresh is in triggering ads from email activity, and what the threshold means for a small shop's tool stack is in a unified campaign automation stack.

Setting up a lookalike segment on Google

Google's lookalike segments work in Demand Gen campaigns and, more recently, in Video brand campaigns. Used anywhere else, the segment shows as "Eligible, Limited" and does nothing.⁸⁹ The earlier version of this article said Demand Gen only, which was true when it was written.

Demand Gen runs on YouTube, including Shorts, on Discover, in Gmail and on Google Display Network placements.¹⁰ The seed lists together must hold more than 100 active matched people, you can use up to ten of them, and a mobile device ID list is not allowed. Seeds can be website visitors, app users, customer lists, Google Analytics audiences or YouTube audiences.⁸ The earlier article hedged between 100 and 1,000. Google's help pages and its API documentation both say 100 today.⁸¹¹

Reach comes in three levels. Narrow aims at 2.5% of users in the advert's target location, Balanced at 5%, and Broad at 10%, with Balanced the default.⁸⁹ A new segment takes about 72 hours to populate, Google recommends creating it two to three days before you want adverts to serve, and it refreshes automatically every one to two days.⁸

The steps: create a Demand Gen campaign, add an audience at ad group level, choose New audience, open the lookalike segment option and choose New segment. Name it, choose up to ten seed lists, enter the countries, set the slider and save.⁸ Put the seed, the country and the reach level in the name, because you cannot edit a segment's definition later from the audience list.

Worked example: Megan's first segment on Google

Megan's Google account is 14 months old and has spent about €22,000 in total. That is well under 50,000 US dollars, so she cannot target with customer lists, and she does not lean on her customer list as a seed. Instead she uses her Google Analytics audience of purchasers in the last 180 days, 2,100 people, comfortably over 100. She names the segment "LAL purchasers 180d, GB, narrow", sets the country to the United Kingdom and the slider to narrow. On the example's assumption of 40 million users, that aims at about one million people. She creates it on a Monday for a campaign starting on Thursday.

Since 2026 Google's slider is a suggestion too

Google has made the same move as Meta. During 2026 it is rolling out what it calls lookalike as a suggestion in Demand Gen. The narrow, balanced and broad slider no longer works as a targeting limit. It becomes a signal the system uses while it hunts for conversions, and the change is on by default.¹² Advertisers have been able to opt out since March 2026, through a form that takes about three days to process, with controls in the interface promised later in the year. Opted out ad groups keep the old behaviour, where the slider respects the 2.5%, 5% and 10% limits.¹² Lookalike segments in Video brand campaigns are not affected.¹²

Optimised targeting is the other dial. In Demand Gen it uses Google's models to look beyond the audiences you chose, which Google calls signals.¹⁰ When both are switched on, Google says most of the reported gains move to the lookalike row, but optimised targeting can still reach people beyond the suggestion's reach.¹²

That corrects a tip in the earlier version of this article, which said that turning off optimised targeting keeps your adverts inside the lookalike. In 2026 that is no longer enough. If you want a hard boundary, the ad group needs to be opted out of the suggestion mode as well.

Worked example: comparing cost per order

Megan runs a test on Meta first, because her Google account is small. Three ad sets, €1,000 each over four weeks, identical adverts, Advantage+ audience off. She compares cost per order, which is spend divided by orders:

$$\text{CPA} = \frac{S}{O}$$

Here $\text{CPA}$ is the cost per acquisition, in this case per order, $S$ is the spend on the ad set and $O$ is the number of orders it produced. It is a rate like any other, money per event. Foundations: ratios and rates

The 1% lookalike brings 36 orders: $\frac{1{,}000}{36} = €27.78$.

The 3% lookalike brings 32 orders: $\frac{1{,}000}{32} = €31.25$.

The 5% lookalike brings 25 orders: $\frac{1{,}000}{25} = €40.00$.

Each bar is one ad set, and the vertical axis is what one order cost in euro. In this invented test, cost rises as the lookalike widens, because each wider audience adds people less like the seed. The narrow one wins on cost, but it is also the one that runs out of new people first, so the right answer depends on how much Megan wants to spend. With 36 orders against 25, the gap is big enough to take seriously but small enough to deserve a second month before she bets the budget on it.

What lookalikes do well

They improve on cold traffic. People who share traits with your customers tend to convert better than people chosen by guesswork. The earlier version of this article quoted "some studies" showing a 26% lower cost per acquisition for 1% lookalikes. I could not find the study, so the figure is gone. Test it in your own account, as Megan did.

They take your guess out of it, finding combinations of traits you would never select by hand. And they scale in steps: a narrow lookalike that works can be widened while you watch cost per order.

They stay current. Meta refreshes members every three days while the lookalike is in use, and Google every one to two days.²⁸

They can be built per product. Buyers of one range may look nothing like buyers of another, so each can have its own seed and its own lookalike.

The limits you need to know about

Quality in, quality out. A seed of people with nothing meaningful in common gives the model nothing coherent to find. A focused seed of 1,000 engaged buyers usually beats a mixed list of 50,000 contacts.

Apple took away much of the data. Since iOS 14.5, an app must ask permission through App Tracking Transparency before it tracks a person across other companies' apps and websites, and until the person agrees, the device's advertising identifier reads as zeros.¹³ The earlier version of this article said about 88% of iOS users worldwide and 96% in the United States opt out. Those were early daily figures reported in May 2021. Flurry, whose data was behind them, reported in the same month that the opt in rate across all apps had risen from 11% to 15%, and that among apps which had actually shown the prompt it hovered around a quarter.¹⁴ Either way, most iPhone users said no, and Meta has far less cross app data about them than it had in 2020.

Third party cookies did not die. The earlier version said Chrome was following Safari and Firefox in removing them. It is not. Google has kept its approach of offering users a choice about third party cookies in Chrome, and in October 2025 it announced it would retire most Privacy Sandbox technologies, including Topics and Protected Audience, citing low adoption.¹⁵ Browser signals are still thinner than they were, mostly because of consent banners and Safari, but the cookie apocalypse was cancelled.

The model is a black box. You see the size and a little about who is in the audience, never the traits the model used. That makes poor performance hard to diagnose.

Audiences tire. The same people see the same adverts and respond less. Refresh the seed and the creative, and rotate between seeds.

Small businesses may not reach the minimum. One hundred matched people sounds low, but a list of 300 email addresses may match far fewer than 300.

Lookalikes copy bias. If your customers are skewed by how you marketed in the past, the lookalike copies the skew rather than helping you reach new kinds of buyer.

I have also cut a claim from the earlier version that audiences now need at least 500,000 people to get out of the learning phase. No platform documentation says this, and on Meta the question has largely moved to Advantage+ audience anyway.

Under the hood: a classifier trained on positives only

Google and Meta do not publish their lookalike models. What follows is how lookalike modelling works in general, in four stages.

Read from top to bottom. First, your seed is matched to the platform's users, and each matched person brings along everything the platform knows: demographics, devices, what they engage with, what they buy. Second, a model learns which combinations of those traits set your seed apart from everyone else. Third, every user in the target location gets a score for how much they resemble the seed. Fourth, the users are ranked, and the top share becomes your audience. When you choose 1%, you choose where the cut sits in that ranking.

The scoring model can be anything from a logistic regression to a neural network. Logistic regression is the simplest to see working, because it turns a weighted sum of traits into a probability:

$$p = \frac{1}{1 + e^{-z}}, \quad z = \beta_0 + \beta_1 x_1 + \beta_2 x_2$$

Here $p$ is the estimated probability that a person belongs with the seed, $e$ is the base of the natural logarithm, about 2.718, and $z$ is the weighted score. $\beta_0$ is the starting point when no trait is present, $x_1$ and $x_2$ are two traits, each 1 if the person has it and 0 if not, and $\beta_1$ and $\beta_2$ are the weights the model learned for them. The logistic function squeezes any score into a probability between 0 and 1. Foundations: regression

Worked example: scoring two strangers

Suppose the model learned, for Megan's seed, a starting point $\beta_0 = -4$, a weight $\beta_1 = 1.5$ for "has bought home fragrance online", and $\beta_2 = 0.8$ for "engages with interiors content". All made up.

Person A has both traits: $z = -4 + 1.5 \times 1 + 0.8 \times 1 = -1.7$, so $p = \frac{1}{1 + e^{1.7}} = \frac{1}{1 + 5.474} = 0.154$.

Person B has only the second: $z = -4 + 1.5 \times 0 + 0.8 \times 1 = -3.2$, so $p = \frac{1}{1 + e^{3.2}} = \frac{1}{1 + 24.53} = 0.039$.

Person A scores about four times higher. Rank 4 million people this way, keep the top 40,000, and you have a 1% lookalike. Real models use far more traits, but the principle is the same.

There is a catch built into the setup. The platform has confirmed positives, your seed, but no confirmed negatives. Nobody has told it who is definitely not like your customers. Statisticians call this learning from positive and unlabelled data, or PU learning, and it is a field of its own.¹⁶ It is what makes lookalikes easy to use, because a customer list is all you need. It is also what makes them easy to fool. If your seed shares a trait that has nothing to do with buying, the model will happily learn it.

An example: Megan ran a big Instagram giveaway in December, and 400 of her 1,200 repeat buyers first arrived through it. The model may learn "enters giveaways" as a defining trait and fill the lookalike with competition fans. The fix is in the seed, not the model: leave the giveaway cohort out, or give it its own seed.

Meta and Google learn from different things

Meta matches your seed against people's activity on Facebook and Instagram. The lookalike's membership refreshes every three days, and separately the delivery system keeps learning inside the audience, putting the adverts in front of the people most likely to respond.²⁴ That is why creative and offer still matter with a great seed. Since late 2024 Meta's ad retrieval runs on a system it calls Andromeda, built to pick from a fast growing number of adverts, many of them made with its generative tools.¹⁷ In practice, variety in creative now does part of the work that narrow audiences used to do.

Google builds lookalikes from seeds you control: your website and app audiences, your customer lists, your YouTube audiences and Google Analytics audiences.⁸ Because they live in Demand Gen and Video brand campaigns, they serve across YouTube, Discover, Gmail and display placements.⁹¹⁰

Worked example: intent based seeds for a car marketplace

The earlier version of this article told the story of a real marketplace I worked on. The names are gone, and the example below is invented, but the pattern is the one I built there: seeds made from what people did on a page, not just which page they saw. In that project the granular seeds did better than generic car buyer targeting on cost per lead and conversion rate.

Treaty City Cars is an invented used car marketplace in Limerick, at treatycitycars.example. It has about 40,000 visitors a month. Car buyers are not one audience. Someone looking at a €14,000 family hatchback is a different prospect from someone looking at a €45,000 estate, with a different budget and different worries.

Step one: events that mean something

A page view says little. So the site sends events that do mean something, and only for visitors who agreed to advertising tags, a point the consent section below returns to:

• a make view, when someone has looked at three or more listings of the same make,
• a deep read, when someone scrolls past 75% of a listing, through the photos, the specification and the finance section,
• the model and the price bracket of each listing read that way.

// Fires once per listing, only after the visitor accepted advertising tags
const listing = { make: 'Ford', model: 'Focus', priceBand: '12k-20k' }
let sent = false
window.addEventListener('scroll', () => {
  const depth = (window.scrollY + window.innerHeight) / document.body.scrollHeight
  if (!sent && depth >= 0.75 && window.consent?.advertising === true) {
    fbq('trackCustom', 'ListingDeepRead', listing)
    sent = true
  }
})

Step two: seeds by model and price

In a month, 1,800 visitors look at three or more Ford Focus listings, and 600 of them read at least one past 75%. Those 600 become the "Focus, deep read, €12,000 to €20,000" audience. That is over Meta's minimum of 100, and every person in it did the same thing with the same kind of car. A buyer of €45,000 estates is never mixed in.

Step three: two uses for one audience

Retargeting: the 600 see adverts for Focus listings that match their price band, not a generic "browse our cars". Prospecting: a 1% lookalike of the 600 sees the same Focus adverts. It reaches people who have never visited but resemble the ones who read a Focus listing to the end. The same pattern repeats for the most popular models, each with its own seed, its own lookalike and its own adverts.

Why it works

The seed is clean, because a deep read filters out accidental clicks. The seed is consistent, because nobody in it shopped for a different kind of car. And the advert matches the seed, so the journey from advert to listing makes sense. As traffic grows, the seeds refresh and the lookalikes follow.

For a shop, the same thinking works with different events: size selector used, colour swatch clicked, added to wishlist.

Consent and list uploads in Ireland

Two sets of rules sit under every lookalike, one for the browser and one for the list.

The browser side: in Ireland, regulation 5(3) of the ePrivacy Regulations 2011 says nobody may store information on a user's device, or read information already stored there, without the user's consent and clear, comprehensive information about the purpose. The only way round it is regulation 5(5): storage purely to carry a communication, or what is strictly necessary for a service the user explicitly asked for.¹⁸ Advertising is neither. The Data Protection Commission (DPC) says cookies related to advertising are not strictly necessary and must be consented to, and that the rule covers pixels and similar technologies too.¹⁹ So a pixel that builds website audiences, the raw material of most lookalike seeds, fires only after a yes. How to do that without losing measurement is in my article on event tracking and consent mode.

The list side: uploading a customer list is processing personal data under the GDPR, so you need a lawful basis for doing it. Most shops reach for legitimate interests in Article 6(1)(f), which Recital 47 says may cover direct marketing, provided the balancing test holds. And anyone who has objected to direct marketing must stay out, because once a person objects their data may no longer be used for it.²⁰ Meta's customer list terms make you warrant both, and say Meta deletes the matched data after matching unless you chose optional features.²¹ Keep your suppression list current and check it before every upload. The detail of the lawful basis, hashing, consent fields and refresh timing is in triggering ads from email activity.

Server side tracking helps with data quality, not with consent. Meta's Conversions API connects your server, app or CRM directly to Meta.²² It makes the data you are allowed to send arrive more reliably. It does not make data you are not allowed to send acceptable.

Habits that make lookalikes better

Start with your best customers, not all customers. A seed of your top fifth by lifetime value describes the buyers you want. Mixing high and low value customers blurs the pattern.

Keep seeds consistent. Separate lookalikes per category beat one lookalike of everyone who ever bought anything.

Test the size. Do not assume 1% always wins. With a small seed or a product that appeals widely, 5% or 10% can do better. Test it properly, with the methods in the complete guide to digital marketing testing methods.

Refresh the seed. Add recent buyers monthly and build new lookalikes from recent conversions.

Exclude people you already have. On prospecting campaigns, exclude existing customers and recent visitors. On Meta, exclusions hold even with Advantage+ audience on.⁴

Give it time. A new ad set's first day says almost nothing. Judge it on weeks, not hours.

Test the fence against the suggestion. On both platforms, run one version with the lookalike as a boundary and one where the platform treats it as a suggestion, as Megan did. More often than not the platforms want you to pick the second, so make them earn it.

Where lookalikes go next

The direction on both platforms is the same. Meta folds lookalikes into Advantage+ audience as one signal among many. Google has turned its slider into a suggestion and leans on optimised targeting. Both are betting that their models find buyers better than your boundaries do, as long as they get good data and varied creative. For the wider argument about working with these automated campaigns, see the shift to black box AI campaigns.

That does not make lookalikes useless. It moves the value from the targeting setting to the seed. A well built seed of your best customers is still the clearest thing you can tell a platform about who you want. What changed is that the platform now takes it as advice rather than an order. Give it high quality data about your best customers, and it will find more people like them, whatever the setting is called this year.

Footnotes

1. Emil Protalinski, Facebook announces Lookalike Audiences, lets advertisers target potential customers similar to current ones, The Next Web, 19 March 2013, https://thenextweb.com/news/facebook-announces-lookalike-audiences-lets-advertisers-target-potential-customers-similar-to-current-ones

2. Meta, Lookalike Audiences, Marketing API documentation, Meta for Developers, updated 7 October 2026, https://developers.facebook.com/docs/marketing-api/audiences/guides/lookalike-audiences

3. Meta, Value Based Lookalikes, Marketing API documentation, Meta for Developers, 2026, https://developers.facebook.com/docs/marketing-api/audiences/guides/value-based-lookalike-audiences

4. Meta, Advantage+ Audience, Marketing API reference, Meta for Developers, 2026, https://developers.facebook.com/docs/marketing-api/audiences/reference/targeting-expansion/advantage-audience/

5. Meta, Marketing API: Changes to Advantage+ audience behaviours, Meta for Developers blog, 13 June 2025, https://developers.facebook.com/blog/post/2025/06/13/marketing-api-changes-to-advantage-plus-audience-behaviors/

6. Google, Building more durable and effective audience strategies in Google's ads platforms, Google Ads Help, 1 November 2022, https://support.google.com/google-ads/answer/12929853

7. Google, Customer Match policy, Google Ads Help, 2026, https://support.google.com/google-ads/answer/6299717

8. Google, Use Lookalike segments in Demand Gen campaigns as a Targeting Constraint, Google Ads Help, 2026, https://support.google.com/google-ads/answer/16902001

9. Google, Use Lookalike segments in Video campaigns to grow your audience, Google Ads Help, 2026, https://support.google.com/google-ads/answer/17219920

10. Google, Demand Gen Frequently Asked Questions, Google Ads Help, 2026, https://support.google.com/google-ads/answer/14509385

11. Google, Lookalike audience segments, Google Ads API documentation, Google for Developers, 2026, https://developers.google.com/google-ads/api/docs/remarketing/audience-segments/lookalike-audiences

12. Google, Use Lookalike segments to grow your audience, Google Ads Help, 2026, https://support.google.com/google-ads/answer/13541369

13. Apple, User Privacy and Data Use, App Store, Apple Developer, 2026, https://developer.apple.com/app-store/user-privacy-and-data-use/

14. Flurry, iOS 14 Opt in Rate: Behind the Numbers, Flurry blog, 20 May 2021, https://flurry.com/blog/ios14-opt-in-rate

15. Anthony Chavez, Update on Plans for Privacy Sandbox Technologies, Google Privacy Sandbox blog, 17 October 2025, https://privacysandbox.google.com/blog/update-on-plans-for-privacy-sandbox-technologies

16. Jessa Bekker and Jesse Davis, Learning from positive and unlabeled data: a survey, Machine Learning, 2020, https://arxiv.org/abs/1811.04820

17. Meta, Meta Andromeda: Supercharging Advantage+ automation with the next gen personalized ads retrieval engine, Engineering at Meta, 2 December 2024, https://engineering.fb.com/2024/12/02/production-engineering/meta-andromeda-advantage-automation-next-gen-personalized-ads-retrieval-engine/

18. European Communities (Electronic Communications Networks and Services) (Privacy and Electronic Communications) Regulations 2011, S.I. No. 336 of 2011, regulations 5(3) and 5(5), Irish Statute Book, 2011, https://www.irishstatutebook.ie/eli/2011/si/336/made/en/print

19. Data Protection Commission, Guidance note: Cookies and other tracking technologies, DPC, 2020, https://www.dataprotection.ie/sites/default/files/uploads/2020-04/Guidance%20note%20on%20cookies%20and%20other%20tracking%20technologies.pdf

20. Regulation (EU) 2016/679, General Data Protection Regulation, Article 6(1)(f), Recital 47 and Article 21(2) and (3), EUR-Lex, 2016, https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32016R0679

21. Meta, Customer List Custom Audiences Terms, Meta, 2026, https://www.facebook.com/legal/terms/customaudience

22. Meta, Conversions API, Meta for Developers, 2026, https://developers.facebook.com/docs/marketing-api/conversions-api