Custom Audiences: Build and Manage Them With Your AI Agent

What are custom audiences?

Custom audiences let you show ads to specific groups of people based on data you already have, instead of relying only on the platform's interest and demographic targeting. You can upload your customer list, reach people who visited your website, or follow up with people who engaged with your content. Then you build "lookalike" audiences that find brand new people who resemble your best customers.

This is some of the highest-leverage targeting in paid media. The problem has always been the setup. Customer data has to be cleaned, hashed, mapped to the right fields, uploaded in batches, refreshed as people churn, and then wired into each ad set. Do one step wrong and your match rate quietly collapses. Most marketers either avoid it or pay an agency to babysit it.

Xylo removes that work. Xylo is a connector (an MCP server) that plugs your Meta, Google, and TikTok ad accounts into the AI assistant you already use, like Claude, ChatGPT, Cursor, or Claude Code. You connect your accounts once, then you just talk to your agent. It does the real work across all three platforms and tells you what it changed.

So the rest of this guide is not about writing code. It is about what to ask for.

The four kinds of custom audience

The same handful of ideas show up on Meta, Google, and TikTok, with slightly different names. In plain terms:

1. Customer list audiences. You upload emails and phone numbers from your CRM, and the platform matches them to real accounts. (Meta calls these custom audiences from a customer list, Google calls it Customer Match, TikTok calls it a customer file audience.)
2. Website audiences. Retarget people who visited specific pages on your site. This needs a tracking pixel installed (the Meta Pixel, a Google tag, or the TikTok Pixel).
3. App and engagement audiences. Reach people who took an action in your app, watched your videos, opened a lead form, or interacted with your profile.
4. Lookalike audiences. Start from any of the above and let the platform find new people who behave like them. This is your main prospecting engine.

You do not need to memorize which platform calls which what. You describe the group of people you want, and the agent picks the right audience type on the right platform.

How it works in practice

Here is the whole flow.

1. Connect once. Sign in to Xylo, authorize Meta, Google, or TikTok through their normal OAuth screens, and pick the account you want. Your tokens are encrypted (AES-256) and the agent never sees your credentials.
2. Point your AI app at Xylo. Add the Xylo connector using the endpoint https://xylomcp.com/api/mcp. It works with Claude (on claude.ai or Claude Desktop), ChatGPT (any paid plan, in Developer mode), Cursor, Claude Code, Codex, and anything else that speaks MCP.
3. Ask in plain English. No dashboards, no CSV gymnastics, no field mapping. Xylo carries 300+ ad operations that cover essentially every Ads Manager surface, and the agent reaches for the right ones to do what you asked.

A safety note worth saying up front: anything that spends money is created PAUSED by default, so the agent can build a campaign and target your new audience without a cent going out until you look it over and approve.

Creating a customer list audience

Uploading a customer list used to mean exporting a CSV, hashing every email and phone number with SHA-256, lining up the columns in exactly the right order, and pushing it up in chunks. Get the column order wrong and the platform accepts the file anyway while your match rate falls through the floor.

With Xylo plus your AI agent, you never touch any of that. You point the agent at your customer data and describe the segment.

Create a Meta custom audience called "Top Customers Q1 2026" from the customers in this list who have a lifetime value above 500 dollars. Use their email, phone, and name so we get the best match rate, then tell me the estimated audience size.

The agent handles the hashing, the field mapping, and the batched upload for you. You can hand it the same list and ask it to build the matching audience on more than one platform at once.

Take this customer list and build a Customer Match audience in Google Ads and a customer file audience in TikTok with the same people, so I can retarget them everywhere.

Keeping audiences fresh

Audiences are not set-and-forget. People buy, people churn, and a stale list wastes spend. Normally you would write a recurring job to add and remove members. With an agent you just say what changed.

Add these 40 new customers to my "Top Customers" audience on Meta.

Remove anyone who refunded last month from the customer audience and confirm how many you took out.

For ongoing upkeep, you can schedule it in plain language and let the agent do the diffing and the batching behind the scenes.

Every Monday, sync my "Active Subscribers" audience on Meta and TikTok from this CRM export: add new subscribers and remove cancellations.

Large lists are handled for you. The platforms cap how many records go up per call, and the agent splits the upload into the right-sized batches so you do not have to think about limits.

Reviewing what you have

You do not need to dig through three separate Audiences tabs to see where things stand. Ask.

List all my custom audiences across Meta, Google, and TikTok with their type, status, and approximate size. Flag any that are too small to use.

You get one readable summary instead of three dashboards, with sizes and readiness side by side.

Creating lookalike audiences

Lookalike audiences are the most powerful prospecting tool on these platforms. They find new people who resemble a source audience you already trust, like your purchasers or your highest-value customers. (For a deeper walkthrough, see our lookalike audiences guide.)

Building one is a sentence.

Build a 1 percent lookalike of my "Top Customers" audience targeted at the United States, then tell me how big it is.

The "percentage" controls reach versus precision. A 1 percent lookalike is the closest match to your source and the smallest. A 10 percent lookalike is much larger but looser. You can ask the agent to build a few and compare.

Create 1 percent, 3 percent, and 5 percent lookalikes from my purchasers in the US, and a 2 percent lookalike in Canada. Name them clearly so I can tell them apart.

Putting audiences to work

A custom audience does nothing until it is attached to an ad set. The agent can build the campaign and the targeting in the same breath, paused and ready for your review.

Set up a retargeting campaign on Meta aimed at my "Top Customers" audience, 50 dollars a day, ages 18 to 65. Leave it paused so I can check it first.

Three targeting patterns are worth knowing, and you can ask for any of them directly.

Exclusion. Stop paying to acquire people you already have. Excluding current customers from a prospecting campaign is one of the fastest ways to cut wasted spend.

In my prospecting campaign, exclude my existing customers audience so we only pay for new people.

Layering. Combine a custom audience with demographics or geography to tighten who sees the ad.

Target my 30-day website visitors, but narrow it to women aged 25 to 44 in the US.

Funnel stages. Build a separate audience for each step of the journey and treat them differently. A simple structure looks like this:

Audience	Source	Use case
All website visitors (180 days)	Website pixel	Broad retargeting
Product page viewers (30 days)	Website pixel	Product interest
Cart abandoners (14 days)	Website pixel	High-intent retargeting
Purchasers (365 days)	Customer list	Upsell and cross-sell
Purchasers (365 days), excluded	Customer list	Keep out of prospecting
Lookalike of purchasers (1%)	Lookalike	Prospecting

You can describe that whole structure once and let the agent create the pieces.

Build out my full funnel on Meta: a 30-day product-viewer audience, a 14-day cart-abandoner audience, a 365-day purchaser audience, and a 1 percent lookalike of those purchasers. Leave any new campaigns paused.

Size, match rates, and timing

A few realities apply no matter how you build these audiences, so it helps to know them:

• Minimum size. An audience usually needs at least 100 matched people before you can target it. For a lookalike, the source audience should ideally have 1,000 or more matched people to produce a useful result.
• Processing time. After an upload, it can take from 30 minutes to a few hours for an audience to populate fully. If you check the size right away and it looks low, give it time.
• Match rates. For email-only customer lists, expect roughly 50 to 70 percent of your records to match real accounts. Adding phone numbers and names lifts that meaningfully, which is why the agent uses every identifier you give it.
• Retention. Website and engagement audiences drop people automatically after their retention window (up to 180 days for website audiences, 365 for engagement), so they stay current on their own.

You can simply ask the agent to sanity-check these for you: "Is my source audience big enough to make a good lookalike?" or "Which of my audiences are below the minimum size to run?"

Getting started

1. Pick your highest-value segments. Purchasers, high-LTV customers, recent website visitors. Start there.
2. Upload them through your agent. Hand over the list and let Xylo build the customer audiences on whichever platforms you advertise.
3. Turn on retargeting. Ask for paused campaigns aimed at your website visitors and cart abandoners, then approve them.
4. Build lookalikes from your best lists. Use your purchasers as the seed for prospecting.
5. Schedule the upkeep. Set a weekly sync in plain English so your audiences never go stale.

For how audiences fit into a fully agent-run ad account, read our guide on building AI agents for ads, or browse more example prompts to copy.

For developers (optional)

If you would rather call Xylo directly, there is a REST API behind the same capabilities. The base URL is https://api.xylomcp.com, you authenticate with an x-api-key header, and audience, sync, and targeting operations are available for Meta, Google, and TikTok (Meta endpoints also take an x-ad-account header). Budgets are in dollars, new campaigns default to PAUSED, and responses come back as { data, meta }. Full details, including every audience endpoint, are in the API documentation. Most people never need this layer. They just ask their agent.

---

Custom audiences are too valuable to leave sitting in a spreadsheet because the setup is tedious. Connect Xylo to your AI assistant, point it at your customer list, and let it build and maintain your audiences across Meta, Google, and TikTok while you focus on the offer.

Connect Xylo to your AI agent. The free tier needs no credit card.