Lookalike Audiences: How to Find New Customers in Plain English

What is a lookalike audience?

A lookalike audience is a simple idea. You give the ad platform a list of people you already love (your best customers, your recent buyers, your high value subscribers), and the platform goes and finds millions more people who behave like them. You stop guessing at interests and let real customer behavior point you toward your next customers.

Lookalikes are one of the most reliable prospecting tools in paid advertising. They tend to beat interest based targeting because they lean on what people actually do (buy, return, spend) rather than what they say they like.

The catch has always been the setup. Building a clean source list, waiting for it to process, choosing the right size, excluding existing customers, and rebuilding everything when your data goes stale is fiddly work. It usually meant a spreadsheet, a few trips through Ads Manager, and a careful eye on the details.

With Xylo, you skip all of that. You connect your ad accounts once, point your AI assistant (Claude, ChatGPT, Cursor, and others) at Xylo, and just ask:

Build a 1% lookalike on Meta from my top customers in the US, and exclude everyone who has already purchased.

The agent creates the source, builds the lookalike, sets up the exclusions, and tells you what it did. No dashboards, no exports, no code.

How it works with Xylo

There are three steps, and only the first takes any setup.

Connect in a minute. Sign in, authorize Meta, Google, or TikTok, and pick the account you want to work in. Your tokens are encrypted (AES-256) and the agent never sees them.
Ask in plain English. Tell your AI assistant what you want, the way you would brief a junior media buyer.
Xylo does the work. Behind the scenes, Xylo gives your agent 300+ ad operations that cover essentially every Ads Manager surface across Meta, Google, and TikTok. The agent builds the audience, reports back, and shows you what changed.

The MCP endpoint is https://xylomcp.com/api/mcp. It works with Claude (on claude.ai or in Claude Desktop), ChatGPT (any paid plan, in Developer mode), Cursor, Claude Code, Codex, and anything else that speaks MCP.

The part that actually matters: your source audience

The quality of a lookalike is decided almost entirely by the source you feed it. A great source produces a great lookalike. A sloppy source produces a sloppy one. Here is a rough hierarchy from best to worst.

Source audience	Quality	Why
Top 25% of customers by lifetime value	Excellent	Optimizes for your highest value behavior
All purchasers (last 180 days)	Very good	Broad, but still a real conversion signal
Add-to-cart users (last 30 days)	Good	High intent and a fresher signal
Website visitors (last 30 days)	Moderate	Broad, includes plenty of non-buyers
Page or profile engagers	Moderate	Engagement is not the same as buying intent
Video viewers (50%+)	Fair	Awareness level signal only

A common and costly mistake is using too broad a source. If you hand over your entire customer database (including one-time buyers who never came back), the lookalike optimizes for "anyone who might buy once" instead of "people like your best customers." Tighter is usually better.

You do not have to build any of this by hand. Ask:

Create a Meta customer-list audience called "Top Customers by LTV" from the file I am about to give you, then tell me how many were matched.

Or, if your buyer data already lives in your ad account:

On Meta, build a source audience of everyone who purchased in the last 180 days, and let me know the size before we make a lookalike from it.

How big should the source be?

Hard minimum: 100 matched people. Below this, the platform will not build a lookalike at all.
Practical minimum: around 1,000 matched people for a usable signal.
Sweet spot: roughly 2,000 to 50,000 matched people.
Diminishing returns: above 50,000, more data adds little.

If you are not sure where you stand, just ask your agent to check the matched size before building anything. It will tell you whether the source is strong enough.

Sizing a lookalike: 1% to 10%

Every lookalike has a size, expressed as a percentage of the population in your target country. This is the single most important dial to understand.

1% is the smallest and most similar. These are the people who most closely resemble your source.
10% is the largest and broadest. Far more reach, but a looser resemblance.

Here is how that plays out in the United States.

Size	Approx. US reach	Similarity to source	Best for
1%	~2.4 million	Highest	Conversion campaigns, bottom of funnel
2%	~4.8 million	Very high	A balance of reach and precision
3% to 5%	~7 to 12 million	High	Consideration, mid funnel
6% to 10%	~14 to 24 million	Moderate	Awareness, top of funnel

The rule of thumb: start narrow. For anything aimed at sales or leads, begin at 1%. If that 1% audience performs and you want more volume, widen to 2% or 3%. There is no need to decide all of this yourself. You can hand the strategy to your agent:

From my top-customers source, build three Meta lookalikes in the US at 1%, 2%, and 5%, each in its own ad set with the same creative, all paused so I can review before anything spends.

The agent builds all three, keeps them separate so you can compare them cleanly, and leaves them paused until you give the word.

Choosing a country

A lookalike is always tied to a specific country. The platform finds people in that country who resemble your source, accounting for local behavior. If you sell in several markets, you build a separate lookalike for each one.

Build a 1% lookalike from my best customers for the US, UK, Canada, Australia, and Germany, and name each one by country.

Each market is built independently, so a US lookalike and a German lookalike can look quite different even from the same source. That is the point.

Lookalikes across Meta, Google, and TikTok

Lookalikes are not a Meta-only idea, and Xylo speaks all three platforms.

Meta (Facebook and Instagram). Classic lookalike audiences, sized 1% to 10%, built from a custom or customer-list source.
TikTok Ads. TikTok lookalike audiences work the same way: pick a source, pick a country, choose a breadth from narrow to broad, and TikTok finds similar users for your campaigns, including for GMV Max and Smart+.
Google Ads. Google leans on Customer Match. You upload your customer list, and Google's optimized targeting expands delivery toward people who behave like that list. Xylo builds the Customer Match list and wires it into your campaigns.

You can run all three in one breath:

Set up prospecting on all three platforms from my best-customers list: a 1% lookalike on Meta and TikTok in the US, and a Customer Match list on Google with optimized targeting turned on. Keep everything paused.

When you want a single view of how they are doing, Xylo gives you true cross-platform reporting, one normalized report across Meta, Google, and TikTok, so you are not stitching three dashboards together.

Putting a lookalike to work

Building the audience is half the job. The other half is using it well.

Exclude people who already converted. Always remove your source and your existing customers from a prospecting campaign so you are not paying to re-reach buyers.

When you set up the lookalike ad set, exclude my source audience and my all-customers list so we only reach new people.

Layer demographics only when you have a reason to. If you know your buyer is, say, women aged 25 to 54, you can narrow the lookalike to match. But go easy. The lookalike already did the hard work of finding similar people. Pile on too many filters and you can shrink the audience until delivery struggles.

Add the 1% lookalike to a new ad set, narrow it to women aged 25 to 54 in the US, and warn me if the resulting audience looks too small to deliver well.

Everything starts paused. New campaigns and ad sets that Xylo creates are set to PAUSED by default, so nothing spends until you have looked it over and approved it. That safety net is on by design.

Watching performance

Once your lookalikes are live, you do not need to dig through reports. Ask for what you care about.

Compare my Meta lookalike ad sets over the last 7 days. Show cost per purchase, click-through rate, and frequency for the 1%, 2%, and 5% versions, and tell me which is winning.

A few signals worth knowing, so you understand the answer:

Cost per result by size. Your 1% should usually be the cheapest per conversion. If 2% is beating 1%, your source audience may be too small or too broad.
Click-through rate. A higher rate on the tighter sizes is a good sign that your source is high quality.
Frequency. If the same people are seeing your ads more than three or four times a week, the audience is getting saturated. That is your cue to widen the size or refresh.

You can even let the agent make the call for you:

If any lookalike ad set has a frequency above 4 this week, pause it and tell me what you paused and why.

Keeping lookalikes fresh

Lookalikes do not update themselves. When you add new customers to your source, the existing lookalike keeps using the old data. Left alone, it slowly drifts out of date.

The fix is a simple monthly habit: refresh the source with new customers, build a new lookalike from it, swap it into your ad sets, and retire the old one after a short overlap. You can hand the whole routine to your agent.

Once a month, rebuild my US 1% lookalike from the latest version of my top-customers source, swap it into the prospecting ad set, and archive the previous one after a week.

If your AI app supports scheduled tasks, this can run on its own and simply report back when it is done.

For developers (optional)

Everything above happens through plain conversation, so most readers never touch an API. If you are wiring Xylo into your own systems, there is also a REST API.

The base URL is https://api.xylomcp.com. Authenticate with an x-api-key: xy_sk_... header. Most Meta calls also need an x-ad-account: act_... header, Google needs x-google-customer-id, and TikTok needs x-tiktok-advertiser-id. Audience endpoints live under /v1/... for Meta, /v1/google/... for Google, and /v1/tiktok/... for TikTok. Responses come back as { data, meta, paging? }, budgets are in dollars (not cents), and new campaigns default to PAUSED, just like in the chat flow. Under the hood, Xylo runs on Meta Graph API v25.0, Google Ads API v23, and TikTok Marketing API v1.3.

Getting started

Connect your accounts. Authorize Meta, Google, or TikTok and pick an account.
Build a strong source. Ask your agent to create a source from your highest value customers.
Start at 1% in your biggest market. Let the lookalike run in its own paused ad set so you can judge it cleanly.
Expand gradually. If 1% works, test 2% and 3%, then add more countries and platforms.
Refresh monthly. Keep your source current so the lookalike stays sharp.

For more on the building blocks here, see our custom audiences guide, and to go deeper on hands-off automation, read building AI agents for ads. You can also browse ready-made example prompts to copy and paste.

Want to try it? Connect Xylo to your AI agent. The free tier needs no credit card, covers Meta, Google, and TikTok, and is plenty to build your first lookalike and see it work.