Imagine you’re launching a new kitchen product—a sleek, ergonomic veggie chopper. How could it not fly off the shelves? You know your current customers: suburban moms in their 30s and 40s. So you decide to build a lookalike audience of... suburban moms in their 30s and 40s.

Seems logical, right? 

Well no. The campaign falls flat. But why did this happen?

Demographics alone don’t tell the whole story. What you really needed was a way to reach people who actually enjoy cooking—people with an affinity for trying new recipes, watching food content, or buying cookware online—not just the ones who check some of the same basic boxes as those who do.

Traditional lookalike audiences tend to mirror superficial traits. Faraday’s lookalikes go deeper—focusing on behaviors, affinities, and actions that actually drive engagement.

## What is a lookalike audience?

But let’s take a step back to define some terms. In marketing, a [lookalike audience](https://en.wikipedia.org/wiki/Lookalike_audience) is a group of people who share characteristics with an existing audience—usually one that’s performed well in a campaign. The goal is to find more people likely to convert based on their similarity to your known winners.

Meta (you know, the artist formerly known as Facebook) popularized the term with their [Lookalike Audiences tool](https://www.facebook.com/business/help/164749007013531), which matches new users based on shared traits with a seed audience. As they define it:

> “A Lookalike Audience is a way to reach new people who are likely to be interested in your business because they share similar characteristics with your existing customers.”  
> — [Meta Business Help Center](https://www.facebook.com/business/help/164749007013531)

But what exactly counts as a “similar characteristic”? 

That’s where the quality of your data—and your modeling—really starts to matter, and that is exactly where Faraday can make a critical difference.

## Out with the old, in with the smart

The old way of building lookalikes is all about surface-level demographics: age, gender, location, interests. That works... until it doesn’t. It can lead to broad audiences that seem right but don’t perform, or hyper-narrow audiences that limit scale.

Like in our veggie chopper example: targeting suburban moms might seem like a safe bet—but it only works if those moms actually cook. Without behavioral context, you end up advertising to people who might look like your best customers but have no interest in what you’re selling.

Smart lookalikes—like the ones Faraday builds—go deeper. Instead of focusing only on who people are, we focus on what they do. That means modeling on behavior, not just fixed traits.

We might start with a seed audience of loyal cookware buyers, and then build a lookalike based on behavioral signals like:

* Affinity for cooking content  
* Purchase patterns across category types  
* Life events like moving to a new home (hello, new kitchen)  
* Household makeup (like number of people they cook for)

This kind of modeling doesn’t exclude demographic data—it uses it for context, not constraint. The result is smarter, scalable targeting that finds the people most likely to engage because of what they do, not just what boxes they check.

## Powered by the Faraday Identity Graph

Faraday’s smarter lookalikes are powered by the [Faraday Identity Graph](https://faraday.ai/features/faraday-identity-graph) (FIG)—our massive, privacy-safe third-party data asset that covers over 300 million U.S. adults. But it’s not just about having a lot of data. It’s about the kind of data we have—and what we do with it.

Most traditional lookalike tools lean heavily on deterministic data: fixed traits like age, gender, or location, and basic interest categories. That can give you surface-level matches—but it misses the deeper signals that actually drive intent.

FIG goes beyond that by integrating thousands of behavioral, demographic, and psychographic attributes. That includes things like:

* How often someone shops online, and in which categories  
* What life events they may be experiencing (like moving, getting married, or having kids)  
* Media consumption patterns, brand affinities, and spending behaviors  
* Numerous other traits, [visit this page for a full list](https://faraday.ai/features/faraday-identity-graph)

Using advanced modeling, we analyze these signals to surface [*probabilistic*](https://www.collinsdictionary.com/us/dictionary/english/probabilistic) *traits*—things that aren’t explicitly declared, but strongly inferred from real-world behavior. For example, we don’t need someone to tell us they love cooking; their actions (recipe site visits, cookware purchases, grocery frequency) say it for them.

So instead of just identifying people who look like your best customers, we can find the ones most likely to engage in the same purchasing behaviors.

## Smarter lookalikes aren’t just for digital ads—they’re transforming offline outreach too.

This smarter approach to lookalikes isn’t just changing digital campaigns—it’s reshaping how brands think about offline channels like direct mail too.

In fact, one of our favorite direct mail partners has been seeing some massive benefits by integrating behavioral modeling into their systems to bring precision and scale to their print outreach campaigns. Traditionally, direct mail strategies have focused on small, high-intent segments—like cart abandoners or loyalty members. These audiences convert well, but they’re also inherently limited.

When it’s time to expand reach, brands often assume they’re stuck building generic lists: same ZIP code, similar age range, same general profile. But with Faraday, they can go far beyond demographics—finding new households that act like their best customers, even if they’ve never visited the site or engaged with the brand before.

That means:

* Reaching new, high-propensity households—not just the usual suspects  
* Scaling volume without sacrificing conversion rates  
* Delivering campaigns to people whose real-world behavior signals true interest

It’s a smarter way to tap into the strengths of direct mail—tangibility, trust, high response rates—without relying on guesswork or outdated list strategies.

Because let’s face it: even the most beautiful mailer won’t work if it ends up in the wrong hands.

If you’d like to learn more about how our predictive lead targeting works, check out this recent [customer story](https://faraday.ai/blog/cs-smarter-prospecting-at-scale) about our work with a different direct mail provider.

## Smarter audiences. Smarter campaigns.

Lookalikes shouldn’t just look the part—they should work. Faraday makes that happen with behavior-based modeling, best-in-class data, and predictions that get smarter over time.

Whether you’re mailing coupons for kitchen gadgets or marketing mortgages, the smartest lookalike is the one that converts.

So ready to see the difference? Reach out and say hi\!

Faraday builds smarter lookalike audiences by modeling real-world behaviors—not just demographics—so you can find the people most likely to convert, not just the ones who look the part.

Not your average lookalike: How Faraday makes finding your next best customers easier

You’ve got data—most likely lots of it. And if you’re the type of person who ends up on a Faraday blog, you’ve probably already considered the value predictive modeling could add to your brand. But not all predictions are created equally.

## When does your data require enrichment?

For some tasks, modeling on first-party data alone—like transactions or web events tied to known users—is enough. This works well for optimizing things like cart abandonment campaigns, existing customer retention, or upsells within your CRM. If you determine that’s all you need at your company, Faraday can absolutely model on just your first-party data.

However, first-party-only modeling comes with a few major limitations:

* It can’t help you reach new customers who haven’t engaged with you yet  
* It struggles with cold start problems when launching new products or campaigns  
* It’s limited by the scale and diversity of your existing audience  
* It can miss broader behavioral patterns visible only in third-party data

If your goals include reaching new audiences or launching into unknown territory, you’ll quickly find the cracks in a first-party-only strategy. That’s where Faraday’s data advantage comes into play.

We don’t just model on your existing data—we help you *expand* it. Through identity resolution, we enrich your records with insights from the Faraday Identity Graph (FIG), a privacy-safe dataset covering over 240 million U.S. adults. This allows us to bridge the gap between what you already know and the wider world of high-propensity customers you haven’t met yet.

Enrichment lets us go beyond who a person is, and start modeling what they’re likely to do.

But attentive readers will have noticed a word that might stand out: “**enriched**.” No, we’re not talking about flour, uranium, or even beer. In this context, enrichment refers to our ability to match the identities in your dataset to real individuals or households in FIG—so we can enhance your records with high-value predictive traits.

[![Image of "That's definitely enriched beer," meme](/images/blog/enriched-beer.gif)](/images/blog/enriched-beer.gif)


## So what is identity matching, anyway?

Identity matching is the process of taking the first party records you give us—names, emails, phone numbers, addresses—and matching them to real people and places in FIG. It’s how we go from “some person” to “this person, at this residence, with these behavioral signals.”

Faraday supports two levels of matches:

* **Person-level match**: We recognize the individual based on identifiers like name \+ address, email, or phone, and enrich with person-level traits (e.g. age, gender, purchase propensity) and property-level traits (e.g. home value, square footage).  
* **Residence-level match**: We recognize the household address, but not the individual. In this case, we can still provide property-level traits and aggregate averages for that address.

## How the Faraday Identity Graph fits in

As noted, at the core of this process is FIG—our proprietary dataset, which we’ve been building over the last 12 years. This set covers thousands of behavioral, demographic, and psychographic attributes on over 240 million U.S. adults and their households. When we get a match, FIG is what powers the predictive features we feed into your model.

The inclusion of this set—and third-party modeling more broadly—solves some critical problems that first-party data alone can’t address:

* **Cold start**: FIG helps predict outcomes even when you have little to no historical data on a segment  
* **Audience expansion**: You’re not limited to your CRM—you can model and target high-propensity lookalikes across the U.S.  
* **Signal density**: FIG adds new, behaviorally-relevant features that enrich your model’s inputs  
* **Long-term accuracy**: More diverse and representative data leads to more generalizable predictions

No match? No enrichment. And that means fewer signals to work with.

## Why match rate matters

Match rate—the percentage of records we’re able to enrich—is a direct lever on how well Faraday can perform. It governs:

* The **accuracy** of your predictive model  
* The **reach** of your campaigns  
* The **value** of our insights to your team

Here’s what typical match rates look like, depending on the identifiers in your dataset:

| Input type | Typical match rate |
| ----- | ----- |
| Name \+ address | 60–80% |
| Email only | 30–50% |
| Phone only | 5–10% |

If your list is heavy on phones or anonymized emails, that’s likely dragging things down. And if you’re seeing \<50% match overall, it’s time to investigate.

## Troubleshooting a low match rate?

One of the most common issues we encounter, and accordingly one of the biggest impediments we face when it comes to a good model is a low match rate. But sometimes diagnosing an issue isn’t as easy as you might expect. So here are a few common culprits to look for:

* Sparse or malformed identity data   
* Hard-to-match populations (e.g. younger adults, recent immigrants, underbanked)  
* Anonymized or junk emails (e.g. `email@email.com`, Apple relays, Doordash+ aliases)  
* Business names and commercial address (rather than personal records)  
* Missing names, flipped fields, or swapped first/last names  
* PO Boxes and apartment units with formatting issues

Sometimes it’s just formatting. Sometimes it’s about who’s in your list. But if your list looks like this, you probably have a guess where to start.

| First name | Last name | Email  | Phone number | Last outreach date |
| :---- | :---- | :---- | :---- | :---- |
| Rose, Benjamin | ben.rose@… | Faraday.ai  | 434 54🥕4 | Went great 👍 |

## What we do when match rate is low

When match rates dip below expectations, we kick off a diagnostic process. We look at the identifiers you provided, check for issues like junk emails or missing fields, and investigate how much of the dataset contains the high-value combo of name and physical address.

Depending on what we find, we might:

* Fix dataset mapping or formatting issues  
* Ask for additional identifiers (e.g. billing vs. shipping address)  
* Run a Match Boost through upstream vendors to attach new emails, phones, or addresses to help improve match

**Heads-up:** Match Boost isn’t always included in contracts, so we’ll confirm that before we proceed.

And if we’re getting solid residence-level matches but not person-level ones, it may just be that this population isn’t present in FIG. That’s a data coverage limitation, not a system failure.

## Why identity resolution is crucial for predictive modeling

Your predictive model learns from the signals in your data. The richer the signals, the more precise the model. If we can’t resolve identities, we can’t generate signals. And that means the model is essentially working with blind spots.

Match rate isn’t just a technical stat—it’s the foundation for how much value you can drive with AI. So if your match rate looks low, let’s dig in together. Because the better we know your audience, the better your predictions will be.

## Wrapping it all up

At the end of the day, your predictive model is only as strong as the signals it’s trained on. And assuming we’re not talking about limited, first-party only modeling, those signals depend entirely on whether we can recognize the people in your data—and enrich those records with meaningful traits.

That’s why identity resolution isn’t just a behind-the-scenes technical step. It’s the bridge between your raw data and the actionable insights that drive real outcomes. From expanding your audience and overcoming cold start challenges to increasing model precision and campaign ROI, enrichment through the Faraday Identity Graph is what unlocks the full power of predictive AI.

So if you’re seeing lower-than-expected match rates, or if you’re unsure whether your data is giving you the best possible results, let’s take a look together. Fixing a match rate issue might be as simple as a formatting tweak—or it might point to a strategic opportunity to boost your reach and improve your outcomes.

Because predictive modeling shouldn’t just be about what has worked before—it should be about what’s going to work next. And that starts with knowing who you’re modeling.

Your predictive model is only as strong as your data’s match rate—because if you can’t resolve who’s in your list, you can’t predict what they’ll do.

Identity matching makes or breaks your predictive strategy

You might’ve heard—the economy’s a mess. The stock market is up and down more than a toddler on a trampoline, tariffs are flying, inflation’s still hanging around like an uninvited guest, and budgets are tightening across the board. When that happens, marketing is often the first thing on the chopping block—especially experimental or “new” tools. You know the ones. The tools the old guard still side-eye.

But here’s the thing: in unpredictable times, the worst thing you can do is pull back and wait for calm seas. When the wind picks up, it might feel safer to slow down—to take a step back—but that’s when it’s most important to steer into the wind. Into the waves. The more chaotic the market, the more you need to grab the wheel and take control.

And if you’re thinking about cutting predictive support now, you might as well be throwing your oars overboard. It might feel like a short-term budget win—but we can predict (with high confidence) that it’s costing you future sales.

## Why predictive is more critical than ever

Let’s be clear: predictive modeling isn’t just about getting *more* customers. It’s about getting the *right* customers. The ones who don’t just buy once and disappear—but who come back, bring friends, and grow with your brand over time. That’s what makes predictive AI so powerful, especially when the market is unstable. You’re not just acquiring customers. You’re investing in relationships.

Here’s why cutting predictive tools now could come back to haunt you:

### 1️⃣ Markets are shifting fast—and so are consumer behaviors

Everything’s changing—fast. Economic swings, political shakeups, social trends—you name it. And when the world changes, so do your buyers. Relying on last year’s data (or even last quarter’s) isn’t going to cut it. You need modeling that adapts in real time, rooted in current behaviors and signals. That’s how you keep pace with who’s actually buying, and why, even as the landscape shifts around you.

### 2️⃣ Consumers are spending more selectively

In tighter economies, people don’t just spend less—they spend smarter. That means you should be spending smarter, too.

At two recent conferences, we kept hearing the same thing:

“Lead quality is going down and I’m spending a ton. I’d rather pay more for a few high-quality leads than waste money on cheap ones that don’t convert.”

That’s where predictive modeling stops being a nice-to-have—and becomes a necessity. Faraday helps you zero in on high-propensity buyers—the ones who convert *and* stick around. Less wasted spend, more long-term value.

### 3️⃣ Your competitors probably aren’t guessing

Let’s be real—your competitors aren’t sitting around waiting for the economy to stabilize. They’re adapting, and many are investing in predictive tools to stretch their budgets further and maintain growth. They’re using AI to find better leads, launch more efficient campaigns, and reduce churn. If you’re cutting predictive while they’re doubling down on it, you’re not just falling behind—you’re handing them market share.

### 4️⃣ When budgets tighten, every dollar needs to punch above its weight.

Predictive modeling helps you do more with less by targeting smarter and reducing wasted spend. Instead of blanketing your audience and hoping for a hit, you’re focusing on high-likelihood converters from the start. That’s how you protect your investment and keep performance high—even when you’re spending lean.

### 5️⃣ Goodbye gut instinct, hello actual foresight

In unpredictable times, instinct can only take you so far. The days of "let’s just try it and see what happens" are behind us. Predictive tools help you replace guesswork with clarity—using actual data to guide your decisions, not just vibes. With Faraday, you can see what’s likely to happen next, so you can respond proactively instead of scrambling to react.

## So what does Faraday actually do?

Faraday delivers AI-powered predictive models built on the [Faraday Identity Graph](https://faraday.ai/features/faraday-identity-graph)—our massive, privacy-safe dataset of 1500+ consented third-party attributes covering nearly every U.S. adult and household. This isn’t just data—it’s actionable intelligence. Rather than waiting around for someone to click on an ad or fill out a form, **Faraday AI agents help you understand what a consumer is likely to do before they do it.**

Here’s a few examples of what that looks like in practice:

### 🎯 Lead scoring

Your CRM is probably packed with leads—but not all leads are equal. Some are ready to buy tomorrow, and others might never convert, no matter how many emails you send. Faraday helps you tell the difference. Our predictive lead scoring ranks your leads based on real behavioral and demographic signals—not just engagement metrics—so you can focus your time and budget on the ones most likely to take action. That means more efficient sales cycles, stronger conversion rates, and fewer hours chasing dead ends.

### 👤 Personalization

Most brands personalize based on what they know after the fact—like someone clicked a product, or abandoned a cart. But with Faraday, you don’t have to wait for those signals. Our models can predict the kinds of messaging, offers, and channels most likely to resonate with each individual *before* they’ve ever interacted with your brand. That means more relevant campaigns from the start, better first impressions, and a smoother path to conversion.

### 🧠 Customer intelligence

You probably have a sense of who your best customers are. But do you really know what sets them apart? Faraday digs deep to uncover the traits—behavioral, demographic, and psychographic—that distinguish your highest-value customers from the rest. Then, we help you go out and find more people with those same high-propensity traits. It’s like lookalike audiences, but smarter, and based on what *actually drives loyalty and LTV*, not just surface-level similarities.

## The bottom line: predictive modeling improves your bottom line

Cutting predictive AI to save budget is like taking the headlights off your car to go faster. Sure, it might seem lighter—until you hit a wall. With Faraday, you’re not just marketing smarter. You’re building future-proof growth.

Uncertainty doesn’t mean take a step back and wait—it means take a step forward and figure out how to solve it first. The current economic turmoil we're all experiencing might look like a setback, but in reality, it could just as easily be an opportunity in disguise. An opportunity to differentiate, to show your customers how you handle a crisis, maybe even to redefine how you generate value. We all get to shape our own futures, so don't let the uncertainty scare you. Use it to drive even better outcomes.

Because when the storm hits, you don’t drift and hope for the best. You steer into it.

In uncertain markets, cutting predictive modeling might seem like a quick budget win—but it risks losing the high-value, loyal customers that drive long-term growth.

The economy is in flux, but cutting predictive modeling could cost your next lifetime customer

You’re likely here because you want to get more out of your lead acquisition—and if you’re using [ActiveProspect’s](https://activeprospect.com/) [LeadConduit](https://activeprospect.com/leadconduit/), you’re already ahead of the game. But that is only the first step. Imagine combining that real-time lead filtering power with the [predictive intelligence and lead scoring](https://faraday.ai/templates) that Faraday has on offer. That’s where things get interesting.

This integration gives you the best of both worlds: Faraday’s ability to assess how likely a lead is to convert, and LeadConduit’s ability to make an instant decision on whether that lead is worth bidding on through their ping-post functionality. The result? Smarter decisions, faster processing, higher conversion rates, and optimal use of your marketing budget.

## What is LeadConduit?

ActiveProspect is a leader in optimizing lead acquisition, offering tools that help businesses validate and qualify leads in real-time. Their product, LeadConduit automates the decision-making process of bidding on affiliate leads, including rejecting leads that don’t meet your criteria. If you're buying or routing a lot of leads, this kind of automation is critical.

## So why integrate with Faraday?

While LeadConduit handles the immediate pass/fail/bid criteria, Faraday adds predictive depth to your decision-making through one of our [AI agents](https://faraday.ai/templates/lead-rejection). Our models assess how likely a lead is to convert based on how closely they resemble your customers. Instead of working from a fixed rule set, you’re working from the probability of a lead converting.

By connecting the two platforms, you can automatically filter out low-potential leads before you even bid on them. That means your conversion rates improve, your budget goes further, your sales team spends less time chasing dead ends, and your funnel stays clean and efficient.

## How this works for Faraday customers

When a new lead enters your system, LeadConduit sends that lead’s basic identifying information—things like name, email, phone number, and address—to Faraday for evaluation. From there, the [Lookup API](https://faraday.ai/docs/features/lookup-api) handles two core tasks: [identity matching](https://faraday.ai/blog/identity-resolution) and [predictive scoring](https://faraday.ai/blog/predictive-lead-scoring-b2c).

First, we match the lead to a real individual in our consumer data universe. The more identifiers you include, the more likely we are to find a high-confidence match. Even partial submissions—like just an email and zipcode—can often yield results, though additional fields like full address or phone help improve match accuracy and reduce false positives.

Once a match is made, we instantly return a payload of predictive insights. This can include a [conversion likelihood score](https://faraday.ai/templates/lead-prioritization), [persona classification](https://faraday.ai/templates/thematic-personalization), and any other configured outcomes you’ve built in your Faraday target. These models are customized to your business goals—whether that’s subscription signups, purchases, loan applications, or any other high-value conversion.

All of this happens in milliseconds, a requirement for operating within ping-post. As soon as the lead hits LeadConduit, the Lookup API evaluates it, and LeadConduit uses that response to decide whether to accept or reject the lead in real time. That means you’re not waiting for enrichment or reviewing leads after the fact—you’re getting a go/no-go signal before any budget is committed.

It’s like embedding a predictive intelligence checkpoint directly into your acquisition pipeline. Instead of just filtering on static criteria like age or location, you’re making decisions based on real behavioral and demographic alignment with your highest-value customers.

For additional details on inputs, outputs, and setup options, check out our [Lookup API documentation](https://faraday.ai/docs/features/lookup-api).

## Why this integration matters for your broader strategy

When Faraday and LeadConduit work together, you're not just filtering more effectively—you're building a smarter, faster, and more resilient lead acquisition system that adapts as you grow.

This integration powers:

* **More automation** – Less manual lead review, more consistency at scale  
* **Smarter prioritization** – High-potential leads move forward, low-value leads don’t  
* **Scalable efficiency** – Whether you process hundreds or thousands of leads, the system just works  
* **More efficient use of your marketing budget** – Stops wasted spend on bad leads.

Because it runs on live data and your custom models, it evolves alongside your business. As your audience evolves, your predictions evolve with them. It also adapts in real time to new lead sources, new personas, or entirely new markets.

Perhaps most importantly, this setup takes the grunt work off your team’s plate. With automated evaluation and rejection, your team can focus on what really matters: building relationships, closing deals, and driving growth.

## Ready to put your lead acquisition on autopilot?

Faraday \+ LeadConduit makes it easy to scale smarter, spend better, and stay focused on what actually drives revenue. If you're ready to see what predictive, real-time lead filtering can do for your funnel, let’s talk.

LeadConduit is an extremely powerful tool on its own, but when combined with Faraday’s predictive AI agents, it becomes a real-time decision engine that helps you prioritize high-value leads before you ever place a bid.

Supercharge your lead acquisition with Faraday + LeadConduit

In every Faraday [Outcome](https://faraday.ai/docs/abstractions/outcomes) or [Recommender](https://faraday.ai/docs/abstractions/recommenders), you’ll find a list of traits—things like age range, buying habits, or location—that influenced the model’s predictions. This is called **feature importance**.

Each trait is given a percentage that represents how much it contributed to the model’s decisions. The list is sorted by level of impact, so the features at the top are the ones the model relied on most.

That percentage might seem small at first—especially if you’re seeing numbers like 2% or 3%. But those numbers can carry more weight than you’d expect. Let’s take a closer look.

## Imagine you're using our dashboard for the first time

You scroll to the Data section of your Outcome and see a list of features with importance percentages—maybe something like this:

[![Image of features of importance in a Faraday outcome](/images/blog/product/feature-importance.png)](/images/blog/product/feature-importance.png)

If you're new to predictive modeling, this might raise a few questions:

* Is 3% good, that seems low?  
* Should I be paying attention to this number?  
* Why are all the numbers so small?  
* How do I know what’s really driving the model?

These are totally valid questions. Let’s break it all down, step by step.

## Why 3% can be a big deal

To really understand why 3% can be a big deal, it helps to zoom out and understand how feature importance works in the first place.

When we train a predictive model, it starts by looking across a huge number of traits—over [1,500 possible features](https://faraday.ai/features/faraday-identity-graph) in our [Faraday Identity Graph](https://faraday.ai/docs/abstractions/traits)—to find patterns that relate to the outcome you're trying to predict (like conversion, churn, etc.). But not every feature is useful. Some are noisy, redundant, or irrelevant for your specific business case. So the model filters that list down to only the features that carry a useful predictive signal—on average, that’s around 200 features. 

Once the model is trained, we calculate **feature importance** to give you a sense of which of those \~200 features made the biggest impact. Each feature gets assigned a percentage that represents its contribution to the model’s predictions. All of those contributions are added up and scaled so they sum to 100%—that way, you can compare them directly.

Now here’s the key part: if every feature contributed equally in a 200-feature model, each one would have about 0.5% importance. But that’s not how models work. In practice, some features matter way more than others. So when one feature stands out with 3% or more, it’s doing a lot of heavy lifting.

It’s kind of like being the lead vocalist in a 200-piece choir. Sure, everyone’s contributing—but a few voices are probably carrying the melodies that most of the audience will notice. Those top features are the ones your model is leaning on to make decisions.

## What should I look for in the list?

Instead of just reading the numbers one by one, look for patterns:

* If your **top 20 features** are all clustered between 2–3%, it means the model is pulling from a wide range of signals. It’s a balanced prediction strategy, where no single feature dominates.  
* But if your **top 2 features** are 5% and everything else drops off quickly—say, under 1%—then those top traits are carrying a lot more weight in the model’s decision-making.

In other words, the **r**elative spread tells you more than any single number. The top features can help you understand what kinds of traits are making people more (or less) likely to take the action you're predicting.

## What about features working together?

This is where things get a little more subtle. Some features don’t just matter on their own—they’re powerful in combination with others.

For example, maybe *age* by itself is helpful, and *income level* is too—but together, they might form a much stronger signal. These kinds of relationships are called [interactions](https://christophm.github.io/interpretable-ml-book/interaction.html?), and our models are built to detect them.

That said, interactions don’t always show up clearly in a list of feature importances. A feature might look low on the list, but still be playing a key supporting role behind the scenes. 

## So what if a feature is too important?

In most models, a well-distributed mix of features is ideal. So if you see a single one standing out dramatically with a notably high importance score—we’re not talking about 6% as much as 40-50% here—that is likely a signal worth digging into.

What’s happening? Here are a few possibilities:

- The model may be over-relying on a single feature, which means it's leaning heavily on one signal while ignoring others. That can make the model less flexible and more vulnerable to changes in that one trait. If the feature drops out of future data or its relationship to the outcome shifts, the model may perform poorly.  
- It might be a sign of [data leakage](https://en.wikipedia.org/wiki/Leakage_(machine_learning))—where the feature is unintentionally revealing information about the outcome it’s supposed to predict. For example, if a feature like "last purchase amount" correlates almost perfectly with whether someone converted, the model may appear highly accurate during training but fail in the real world. This is sometimes referred to as "cheating," and it’s one of the biggest pitfalls in predictive modeling.

This doesn’t mean there is a problem, but it certainly is worth a closer look. If you see a feature with importance that looks *too* high, reach out to us—we’ll take a closer look and either help you interpret the results, or make a change if something is happening in your model that shouldn’t.  

## Directionality

Each feature also includes [directionality](https://faraday.ai/blog/faraday-model-reports), which shows whether specific traits had a *positive* or *negative* influence on predictions.

This means you’re not just seeing how much something mattered—you’re also seeing *how* it mattered. For example, you might find that high-income households were more likely to convert (positive direction), while certain lifestyle types were less likely (negative direction). That context can be incredibly useful when refining messaging or targeting strategies.

You can read more about this aspect of feature importance in our previous [blog](https://faraday.ai/blog/faraday-model-reports).

## So what can you do with this info?

Once you understand what’s influencing your model, you can start using that insight in your day-to-day work:

* **Messaging**: If homeownership is one of your top features, your messaging might speak more directly to homeowners.  
* **Targeting**: If younger age ranges rank high, that could shift who you prioritize in your outreach.  
* **QA**: Seeing a feature rank high that *shouldn’t* be relevant? That’s a good moment to raise a flag and chat with us—we want your model to reflect reality, not noise.

Importantly, feature importance isn’t a checklist of what to act on—it’s a lens into what the model “sees” as predictive. Use it to guide your thinking, not to replace your judgment.

## Still have questions about what you’re seeing in your Outcome?

 We’ve got you covered. Just reach out to support—we’re always happy to walk through it with you\!

Feature importance in Faraday Outcomes and Recommenders shows which traits—like age or buying habits—had the biggest impact on the model’s predictions, helping you understand not just what drove decisions, but how and why.

Making sense of feature importance: Why 3% matters more than it might seem

It's 2025, and direct mail isn’t just alive—it’s thriving. 

Despite the rise of digital channels, businesses continue to invest heavily in direct mail because it works. Consumers engage with physical mail in ways that digital ads simply can’t replicate. Seniors rank it as their second-favorite method of business communication, and even Gen Z—raised in the digital age—often find it a welcome reminder of the tangible world. In fact, over [one in ten](https://www.pymnts.com/credit-cards/2024/boomers-prefer-email-gen-zs-opt-for-apps-as-top-choice-for-reward-delivery/) Gen Zers say direct mail is their favorite way to hear about banking rewards, proving that a well-executed campaign can capture attention across demographics. So it’s no surprise, then, that companies are doubling down, with  a [staggering 82% of businesses](https://go.lob.com/rs/900-QJF-050/images/2024-Lob-State-of-Direct-Mail-Marketing.pdf) having increased their direct mail budgets in 2024, a sharp jump from 58% in 2023\. 

And the reason? Direct mail delivers. As our partner, [Lob](https://www.lob.com/state-of-direct-mail/2025-report)—one of the best direct mail providers in the business—[reports it,](https://www.lob.com/state-of-direct-mail/2025-report) nearly 4/5 of executives they surveyed last year said that direct mail was their best outreach channel. That’s how much value they see it driving for their brands. And as AI and automation tools continue to improve execution—now used by 56% of marketers, up from 40% in 2023—direct mail is becoming even more effective, making it an essential part of modern marketing strategies.

But here’s the catch: it’s not always easy to implement.

Firstly, it can come with a serious price tag. Marketers are finding themselves paying more and more each month to reach high-intent prospects via mail. The United States Postal Service (USPS) plans to raise stamp prices five times by 2027, with a significant increase expected in July 2025\. This follows a 5-cent hike in July 2024 and a series of price increases in recent years, [raising stamp costs by 36% since 2019\.](https://www.axios.com/2024/09/26/usps-stamp-price-increase-2025) Additionally, USPS has outlined plans for twice-yearly increases in 2026 and 2027, subject to approval. With rising costs adding up quickly, maximizing the efficiency of direct mail campaigns has never been more important. Secondly, even the best campaigns need to be smart to get good response rates. Even if your messaging is on point, your product is best-in-class, and your brand is well known, it’s very possible that many of those sends won’t get a glance if your campaign isn’t properly targeted

So you’re likely starting to get the picture. That despite the cost going up, direct mail isn’t going anywhere. 

But where does that leave the rest of us marketers? Those without unlimited budgets. Most of us need to carefully manage each and every dollar of marketing spend to ensure the greatest value for our brand, and that means direct mail is a fantastic but risky tool that needs to be used correctly.

## So what should we do about it?

One of the best solutions for managing the low response rate problem is comprehensive direct mail targeting. Highly targeted campaigns allow brands to send mailers to only the most promising prospects, improving conversion rates. However, this approach also has its challenges. First, fully customized direct mail campaigns can still be expensive, as they require detailed audience segmentation and personalized messaging. Second, they can be too complex for many brands, requiring significant time and resources to manage effectively. Third, they depend on high-quality data to ensure accuracy. Although Faraday customers don’t have to worry about many of these issues thanks to the [Faraday Identity Graph (FIG),](https://faraday.ai/features/faraday-identity-graph) which provides rich consumer insights on nearly every US adult and their households, for better targeting, [as well as our powerful AI agents](https://faraday.ai/templates), which handle personalization and optimization, automatically, this tool is still effortful and overkill for some campaigns. Finally, postage for first class mail is three to four times the expense of bulk mailing options.

On the other end of the spectrum, the USPS offers a more budget-friendly option: [Every Door Direct Mail (EDDM)](https://www.usps.com/business/every-door-direct-mail.htm). EDDM allows marketers to send mailers to every address along a specific mail carrier’s route, without the need for custom prospect lists or addressing individual recipients. This can significantly reduce costs, sometimes as low as 20 cents per mailer, making it an attractive option for marketers looking to cut down on expenses.

However, the trade-off is clear: while the cost savings are substantial, EDDM lacks the targeting precision that custom direct mail offers. Since it sends to all households in a given area, marketers can’t ensure that their mailers are reaching high-intent or relevant prospects. This lack of segmentation means that some of the mailers will inevitably go to recipients who may have little to no interest in the product or service being advertised. As a result, brands risk wasting valuable marketing budgets on less targeted customers, potentially diluting the effectiveness of the campaign. Additionally, without the ability to target specific demographics or behaviors, the impact of the message might not be as strong or relevant, leading to lower conversion rates.

So, marketers often find themselves stuck between two good but limited options. Both can be powerful, but in some cases, neither can meet their exact needs. Do they want:

* High-cost, highly targeted direct mail campaigns, or  
* Low-cost, more loosely targeted EDDM campaigns?

As is often the case, a middle ground option would often best serve the needs.

## Enter: AI optimized EDDM

That’s exactly what Faraday has provided with our new high propensity AI optimized EDDM service. With approximately 15 carrier routes in an average zip code, knowing which ones perform best could strike the perfect balance between cost and effectiveness. Instead of choosing between expensive, hyper-targeted direct mail and inefficient, low-cost EDDM, brands can optimize their spend by targeting the best routes available through EDDM striking a balance between the other options. 

Carrier route aggregation provides a new tool by which brands can identify the carrier routes most likely to contain high-converting prospects but this functionality isn’t entirely new. Instead it expands on other existing aggregations we offer, including those at the state and household level (among a range of other options). But this feature is uniquely designed to take advantage of the EDDM option provided by the USPS. For marketers looking to take advantage of this specific service is a more targeted and efficient way, carrier route aggregation can be the perfect answer. With it, Faraday empowers clients to:

* Identify high-propensity carrier routes  
* Deploy cost-effective EDDM campaigns with improved targeting  
* Reduce waste while maintaining strong performance

And it is important to note that we’re not suggesting that aggregating by carrier route level is some magic bullet that will solve all your direct mail problems. The reality is that knowing which targeting method is best suited to the situation in question is the most important question to answer for effective targeting. 

Luckily for our clients, we have a team of expert, forward deployed data scientists (as well as some pretty spiffy AI agents) who can help you answer those questions or just solve the problem for you. At Faraday, you don’t just get predictions, you get targeted implementation strategies that mean those predictions can actually move your business forward, quickly.

## Client example

Let’s unpack how this is already helping one of our home service clients. 

Before this new feature, this client was dropping \~10 million EDDM mailers a year relying on broad demographic filters like household income and age to choose carrier routes. After developing and deploying a [likely customer AI agent](https://faraday.ai/templates/lead-prioritization) across their service territory, they discovered that the top 10% of routes identified by the model delivered more than **3x** the profit per route compared to their previous average.

In practice, the client was able to maintain the benefit of EDDM’s lower cost while still ensuring that mailers reach the most likely converters, resulting in significantly improved marketing and sales KPIs. The bottom line? Smarter targeting makes direct mail more efficient, cost-effective, and impactful—helping brands win more customers while optimizing their marketing spend. 

## In conclusion

Direct mail isn’t going anywhere, but the way marketers use it needs to evolve. Rising USPS costs make efficiency more critical than ever, and blindly choosing between ultra-targeted campaigns or broad EDDM mailings isn’t always the smartest move. Instead, the key is flexibility—having the ability to adjust your strategy to get the best return on your marketing spend. Services like Faraday provide this flexibility and the expertise to utilize it effectively.

We’re not here to tell you that carrier route aggregation is the only answer—but we are saying that using the right tool for the job will make your direct mail campaigns more cost-effective and impactful. Sometimes, broad, untargeted EDDM campaigns make sense (like when you need to reach every homeowner on a street). Other times, hyper-targeted mailers are worth the investment. And now, with carrier route aggregation, you have a powerful middle-ground option to strike the perfect balance between reach and efficiency.

At Faraday, our goal is to equip you with the best possible tools for any situation. Whether you need pinpoint accuracy, cost-effective broad reach, or something in between, we’ve got you covered. So what’s stopping you? Hate making money? If not, let’s talk and see how Faraday can help you optimize your direct mail strategy today. [Sign up here](http://faraday.ai/signup) for a free trial (with a complementary predictive model validation).

Faraday’s AI-optimized EDDM strikes the perfect balance between high-cost, well-targeted direct mail campaigns and low-cost, low precision EDDM by identifying the most profitable carrier routes, maximizing impact while minimizing waste.

Finding the sweet spot in direct mail: Cost-effective targeting with AI optimized EDDM

Before committing to any predictive modeling platform, you need to know if their services actually work for you. That’s where a [validation test](https://faraday.ai/blog/cross-validation-technical-blog) comes in. Here at Faraday, a validation test (such as a holdout or cross validation) lets you see how well our AI agents can predict real outcomes—before you even sign up. By using your historical data, we can prove the accuracy of our agents and models without ever seeing your future results. Essentially, if the model is predictive, that means we're a good fit for helping you decide who to send direct mail to, how to accept and organize leads with lead scoring, and more. It’s the best way to validate AI-driven predictions—before you invest, and without any risk.

## How does it work?

A holdout test works by splitting your historical data into two sets: **one we see** (pulled from your recent conversion data, this is known as the test set) and **one we don’t** (pulled from the **most** recent conversion data, this is known as the holdout). First, we use only the training set to build a predictive agent. Once it is built, we generate predictions and compare them against the conversions that were ‘held out.’ If our agent accurately predicts those hidden outcomes, you can be confident in its ability to drive real results.

Cross-validation works similarly, but instead of holding out a single set for testing, it splits the data multiple times into different training and validation sets. Each subset takes a turn as the validation set while the rest are used for training, ensuring a more comprehensive assessment of model performance.

If you’d like a more detailed description of how this process works, take a look at this [blog](https://faraday.ai/blog/holdout-testing-with-faraday) from our COO, Robin Spencer, or for more detailed setup instructions, see this [blog](https://faraday.ai/blog/holdout-testing-evaluate-prediction-accuracy) from our CEO, Andy Rossmeissl.

Validation testing gives you a real-world view of the model’s accuracy, providing the reassurance that our predictions will stand the test of time—by making sure the AI isn't just memorizing past data, but identifying trends that matter going forward. But for the test to work, we need the right data—uploaded securely. Here’s what you need to know to get results as quickly as possible.

## What data do you need?

To set up a validation test, you must provide **names and dates** for both **leads and conversions**.

### Leads

We need to know when someone became a lead. Use the [Leads template](https://faraday.ai/data_solicitation/leads_template.xlsx) to provide this information, making sure to include the date for each lead. A leads file should ideally include the following fields:

* **First name**  
* **Last name**  
* **Address** (Street, City, State, Zip)  
* **Optional:** Email, Email Hash, Phone  
* **Date** (e.g., the date a lead was added to your system or when they received a mailer)

These personally identifiable details (PII) allow us to match leads to our demographic dataset. Without PII, we won’t be able to validate predictions, and the validation test cannot be run.

🚧 **Note on date formatting:** You can format the date in any standard way (e.g., YYYYMMDD, MM/DD/YYYY) and our system will automatically handle it.

### Conversion event data (transactions)

We also need the dates of individual transactions—along with optional details like product type and amount. Use the [Orders/Purchases/Transactions template](https://faraday.ai/data_solicitation/orders_purchases_transactions_template.xlsx) to submit this data. The conversion file should have the same structure as the leads file (first name, last name, address, etc.), but instead of lead dates, it should contain conversion dates.

### Conversions (customer-level data)

Customer-level conversion data (e.g., just the first or last transaction date) **is not ideal** because we need each individual transaction date. Although you can provide customer-level data using the [Customers template](https://faraday.ai/data_solicitation/customers_template.xlsx), we typically do not use it for a validation test.

## Uploading your data securely

To keep your data safe, always upload it through the **Faraday dashboard** in the Datasets section. You can also connect your platform to your existing data warehouse, but as this integration is more involved, this step would likely come after your successful validation test has been completed and you’ve decided to sign up with Faraday\!

🚧 **Do not email us any data.** The Faraday dashboard is the only secure place to upload files.

### Steps to upload:

1. Go to [app.faraday.ai](https://app.faraday.ai) and sign in.  

[![Sign in page](/images/blog/are-you-ready-holdout/sign-up.png)](/images/blog/are-you-ready-holdout/sign-up.png)

2. If you see a test mode warning, disable it before uploading your data.  

[![Make sure to leave test mode](/images/blog/are-you-ready-holdout/leave-test-mode.png)](/images/blog/are-you-ready-holdout/leave-test-mode.png)

3. Navigate to the **Datasets** section. 

[![Navigate to the Datasets section](/images/blog/are-you-ready-holdout/datasets.png)](/images/blog/are-you-ready-holdout/datasets.png)

4. Click **New Dataset** and select **CSV**. 

[![Select new dataset](/images/blog/are-you-ready-holdout/new-dataset.png)](/images/blog/are-you-ready-holdout/new-dataset.png)

5. Upload your **leads.csv** file and your **conversions.csv** file.

[![Choose CSV upload](/images/blog/are-you-ready-holdout/upload-csv.png)](/images/blog/are-you-ready-holdout/upload-csv.png)

## Want a little more help?

If you're still unsure how to set up your validation test, we've got you covered. [Watch this video walkthrough](https://www.loom.com/share/ab3750fdb50d47c2b6a0fc05e96e10dd?sid=aefe17c6-ba5f-458f-986f-73db3cf43877) with Faraday CTO Seamus to see the process step by step.

And if you still have questions, reach out to us at **support@faraday.ai**—we're happy to help\!

And that’s it\! Once uploaded, we’ll take care of the rest. Ready to get started? Upload your data today and see how Faraday’s AI agents can drive real results for your business.

Validate Faraday’s predictive accuracy risk-free with a holdout test—just upload your data, and we’ll prove our models work before you commit.

Are you ready for your pre-purchase AI agent validation?

Score aggregation is a method we use to enhance the accuracy and stability of predictions when individual-level data is incomplete. Instead of relying solely on aggregated traits—an approach that works great when appending data but can introduce volatility in model outputs when scores are produced with these aggregated inputs—we take the next step and aggregate scores themselves. This ensures that models can still generate high-quality predictions, even when full individual identity resolution isn’t possible.

By leveraging aggregated scores, we can create smoother, more consistent predictions across varying levels of enrichment. This builds on our existing strategy of using aggregated traits for data appends but tailors it specifically for improving modeling outcomes.

## A simple example

Sounds a little complicated? It’s actually easier than you think\! Let’s break this down with a simple scenario that unpacks two hypothetical individuals: 

#### John Doe – Fully identified

John lives at 456 Oak Street, and we have all the necessary PII to match him confidently in our system. Because of this, we can generate a highly individualized score based on his specific traits and history no problem. This is how our identity resolution process ideally works.

#### Jane Doe – Partially identified

Jane, on the other hand, lives at 123 Main Street, and we don’t have enough personal information to fully match her. However, we do know a lot about her household. Previously, we would have used aggregated traits from the household (or street, postal code, depending on how far out we’d need to go to find accurate information) to generate her score, but this could result in fluctuations.

Now, instead of scoring her based on those aggregated traits directly, we leverage an **aggregated score**—we average the **scores** of people in similar groups (in this case her household) rather than creating an average person based on traits and then scoring that, itself. This provides a more stable and accurate prediction, reducing potential variability and improving reliability.

[![Score aggregation diagram](/images/blog/score-aggregation.png)](/images/blog/score-aggregation.png)

## So how does it affect you?

The key impact of score aggregation is that it makes every model better. However, you might also notice changes to individual scores, and here’s why:

* **More reliable predictions:** Instead of introducing potential spikes from scoring based on aggregated traits, we now rely on aggregated scores, which reflect a more stable signal.  
* **More inclusive modeling:** Even when we can’t fully identify an individual, we can still provide a meaningful prediction using data from their household or surrounding population.  
* **Fewer gaps in data:** If a person’s identity is incomplete, we don’t discard the opportunity to score them—instead, we use the best available aggregated score to fill in the blanks.

### Key considerations

When implementing score aggregation, there are a few key considerations to keep in mind. First, using aggregated scores helps prevent sudden shifts in predictions that might occur when only aggregated traits are used. This approach provides a more stable and reliable signal. Second, our methods adapt based on the available data: when full PII (Personally Identifiable Information) is available, we prioritize direct individual scoring. However, when only partial data is present, aggregated scores serve as a reliable fallback. Lastly, if only household-level data is available, rather than estimating an individual’s score based on household traits, we derive a score from the averaged predictions of others in similar households, ensuring that the prediction remains accurate and meaningful despite incomplete individual information.

### In conclusion

By adopting score aggregation, we ensure that models remain robust and useful even when some data points are incomplete. This helps businesses make better decisions with confidence, regardless of data availability. If you'd like to learn more about how this—and the numerous other services and [AI agents](https://faraday.ai/templates) that Faraday provides—could benefit your business, [reach out](https://faraday.ai/talk-to-sales)\!

Score aggregation improves prediction accuracy and stability by averaging scores from similar groups rather than estimating individual scores from aggregated traits.

Noticed a change in your score reporting? Here’s how score aggregation is improving your results

When you’re ready to use Faraday to build [AI Agents](https://faraday.ai/features/customer-behavior-predictions), the first step is getting your data into our system. There are two primary methods to make that happen—using a data warehouse or uploading individual `.csv` files. Your best option depends on your team’s setup and resources. This guide covers how to ingest your data and best practices to make it usable for modeling as quickly as possible.

## Part 1: Locating your data

### Data warehouse vs. file upload

**Data warehouse:**  
If your data resides in a warehouse (e.g., [Snowflake](https://www.snowflake.com/en/) or [BigQuery](https://cloud.google.com/bigquery?hl=en)), integration is straightforward. Your team can provision a small, dedicated space within the data warehouse for Faraday, providing secure, efficient, and timely transfers without the hassle of repeated manual uploads. Our [connections](https://faraday.ai/docs/how-to/connections) console provides the necessary credential details for integration.

**File uploads:**  
If your data is managed as individual files, you can upload them directly via our datasets console. Although this method is simple, it is time-consuming to keep data fresh over time because of the level of effort required each time.
- `note:` Faraday is looking for `.csv` (comma-separated values) files - they can be exported from databases, various business platforms and spreadsheets alike.

  **Automation considerations:**  
  For brands using cloud storage (e.g., [Amazon S3](https://aws.amazon.com/s3/) or [Google Cloud Storage](https://cloud.google.com/storage?hl=en)), setting up automated data transfers is simple and efficient. These services can be connected to Faraday with minimal effort and can be nearly as effective as connecting directly to a data warehouse.

  An SFTP transfer option is equally suitable, but may require slightly more knowledge and access to resources within your organization.

---

## `.csv` file formatting requirements

To ensure smooth processing, follow these guidelines when preparing your CSV files:

- **Format:**  
  - ***Files:*** Most of these files are comma-separated and UTF-8 encoded, ideally with a header row listing the column names. If headers are missing, you can add them in the dataset configuration's 'Advanced options'.
  - ***Data warehouse*:** Sticking as closely to simple and common types of fields such as decimals, integers, boolean (true/false) and text (string/varchar) are most effective. Tables or views are both acceptable.

- **Size:**  
  - ***Files:*** Individual files can be up to 5GB in size, and generally should be fewer than 50 columns. Consider splitting larger files into smaller parts and-or removing any fields not related to *behavior*, *time*, *quality* or *identity*.
  - ***Data warehouse*:** Generally, there is no size limit for data warehouse tables, but large tables should be handled with care because the larger the dataset, the longer the processing time.

- **Structure:**  
  - ***Files:*** If providing updates to data, ensure consistent structure across all files in a dataset (cloud bucket/SFTP: "*prefix*") - column names, orders and the data types they are set to (if applicable).
  - ***Data warehouse*:** If using a table, ensure it remains up to date through a regular refresh process. Flattened tables (not nested) are preferred in most cases. Plan on establishing a set of fields to include and avoid changing them over time (such as adding or removing fields).

- **Delimiter consistency:**  
  - ***Files:*** Prefer commas as standard delimiters, but this setting is customizable at the dataset level (tabs, semicolons). If necessary, ensure all columns are quoted.

- **Additional flexibility:**
  - ***Files:*** We accept zip-compressed files such as .gz or .zip if you are batching and saving on storage.

---

## Part 2: Data essentials for accurate predictions

Once your data is uploaded, it must include the necessary elements for precise predictions. Faraday supports various data types, but certain formats yield better results.

### Types of data

- **Event stream data:**  
  This type of data is preferred for its detailed insights into customer behavior, where each row represents an event (e.g., account creation or purchase) - with a *timestamp*. This granularity is key for understanding customer journeys. (See our [Datasets documentation](https://faraday.ai/docs/abstractions/datasets) for more details.)
    
    Flavors:
    - email engagement data
    - website engagement data
    - app usage data
    - order data
    - sales funnel data
    - customer journey mapping
    - direct mail response tracking

- **Customer-level (summary) data:**  
  This type of data captures one row per customer with summary information such as: purchase history metrics, customer tier, original source or lifetime value. With a joining key (e.g., customer_id), it works best when combined with event stream data.
  - `note:` Summary type data may hold valuable insights on your customers' [first-party traits](https://faraday.ai/blog/first-party-features). It *can* also include event timestamps as additional columns when relevant and is frequently seen in CRM exports.

### Essential fields

To ensure accurate modeling, include these key fields:

- **Personally identifiable information (PII):**  
  First name, last name, address (street, city, state, postal code), email, and phone number. At least name + address or email are required at a minimum. More PII enables enhanced data enrichment through the [Faraday Identity Graph](https://faraday.ai/docs/abstractions/traits), and can unlock a world of value.

- **Event timestamps:**  
  Accurate timestamps are critical for tracking when events occur. "YYYY-MM-DD" is the preferred format, but other common selections are available via the app and API. 
  - `note:` Ensure all time-based fields follow the *same* format, whichever option you choose.

  - **Event properties:**
    Other values which are captured at the time of the event that may exist in your records. (Cost of goods, method of payment, unique products or services, result of event)
    
    In these cases, the additional data should answer the question: `How much value was generated for your business by this event?`

- **Unique identifiers:**  
  Each set of data should have a *primary key* (e.g., an order number) and a *reference key* to link said event to an individual. These datapoints deliver maximum flexibility to both you and Faraday through the process of updates.

### Identity resolution and matching

After your first-party data is connected, Faraday’s identity resolution software links individuals to additional third-party attributes with over 1,500 available traits and counting. Providing as much quality PII as possible improves matching accuracy and minimizes reliance on aggregated traits.

At a minimum, you must provide an email or a combination of first and last name with street address and postal code.

---

## Note on data hygiene

Organizing and labeling your data correctly is vital — poor quality input yields poor quality output every time. Clean, well-curated data leads to smoother processing and more accurate insights.

**Best practices include:**

- Maintaining consistent formatting: no type mixing across fields, names and addresses follow a similar structure, categories have fixed sets of standard values, and avoiding filler values for missingness (e.g., "None", "N/A")
- Verifying that phone numbers are a full 10 digits, accurate emails are provided, and names come with *full* physical addresses
- Deduplication
- Ensuring you do not have large gaps in field coverage or in time between groups of events
- Not providing more than you intend to use (aside from future use cases)

Investing time in even minimal data hygiene upfront saves time and reduces the likelihood of missing a problem that was lying in wait.

---

## What’s next?

Getting your data into Faraday is just the beginning. By following these best practices, your data will be structured and ready for predictive modeling—whether you’re predicting lead conversion or conducting market opportunity analyses. If you have trouble checking all the boxes above, don't worry - Faraday will help find a solution for your needs.

Let us know if you have additional questions as you prepare your data. You can connect with support at support@faraday.ai - happy modeling!

To get started with Faraday, proper data ingestion is essential. This article offers a technical overview of the processes and best practices for preparing and ingesting your data into Faraday's system, ensuring fast and accurate predictions.

Preparing your data for seamless ingestion - a technical guide

When brands partner with [Faraday](http://faraday.ai), the goal is clear: deploy predictive models as quickly and effectively as possible to [drive business ROI](https://faraday.ai/stories/financial_services_client). But before any predictions can happen, the foundation has to be set, and that means getting the right data into the platform.

For marketers without a dedicated data scientist on their team, this step can feel like a hurdle. What data does Faraday need? In what format? How can you ensure a smooth process that avoids issues with matching and model performance?

These are big questions, and answering them correctly could be the difference between actionable predictions in a week or two versus a much slower and less efficient process.

## Start with the big picture

Before diving into spreadsheets and data exports, take a step back. The first question we ask any new client is: ***What problem are you trying to solve?*** Your goal determines the data you’ll need. For example:

* Are you looking to improve conversion rates? You’ll need historical data on past conversions.  
* Want to identify high-value customers? Transactional data will be key.  
* Need to optimize outreach? Engagement metrics and response data should be included.

Knowing your objective up front ensures that the data you provide aligns with your business needs. 

So what data do we need? The best answer is probably to give us access to *everything—*or at least, as much as is feasible for your company. While some data points are necessary to make the right predictions, all of them can help us to build a better and more holistic understanding of your existing customer base and the best potential customers you could reach within the US population.

![Meme about how much we love data](/images/blog/welovedata-meme.png) 

## So more IS better, but formatting matters

*So we’ll just give Faraday all our data as is and let them figure it out?*

**So, the answer is both yes and no.** While more data is generally better, it’s not just about volume—it’s about data quality. Brands without an established system to ensure [data hygiene](https://www.indeed.com/career-advice/career-development/data-hygiene) may encounter issues like malformed addresses (e.g., missing state or zip code), missing customer details (e.g. no address at all), or improperly formatted columns (e.g., first and last names in one field). While Faraday’s team can spot these issues early on, working with clean, well-structured data from the start will drastically speed up the process, leading to faster, more accurate predictive models.

If you're facing challenges with data hygiene, don’t hesitate to reach out. In addition to our standard data cleanup, we offer more robust professional services upon request, with a variety of packages designed to help you wrangle your datasets. Even if your data isn’t perfect, we’ll help you identify and fix any issues, setting you up for quicker results.

## So what data should I include?

At a minimum, the following fields should be included to ensure proper matching and predictive accuracy:

* **Customer identifiers:** Name, email, phone, and address (as complete as possible)  
* **Leads:** Name and the date that someone became a lead  
* **Conversions (transaction-level preferred):** Dates of individual transactions, along with product details and amounts (if applicable)  
* **Engagement metrics:** Website visits, email opens, and interactions

#### Customer-Level vs. Transaction-Level Data

While you can provide customer-level conversion data (e.g., first or last transaction date), we usually do not need it and cannot use it for a holdout test. Instead, transaction-level data—with a date for each transaction—is preferred, as it enables more accurate predictive modeling.

And if you're unsure where to start, focus on conversion rates. Defining how your brand measures conversion—whether through completed purchases or booked appointments—helps us create more accurate predictive models and simplifies the entire process.

## Where does your data live?

Once you’ve identified the right data, the next step is understanding where it lives. Is it spread across multiple systems like a CRM, e-commerce platform, and customer support database? Do you have easy access to it and permissions to share it externally, or will you need internal approvals? For enterprise brands, pulling data may require specific requests within your organization. Understanding these logistics early on can prevent delays.

In the best case, your data lives in an organized data warehouse like [Snowflake](https://www.snowflake.com/en/), [Bigquery](https://cloud.google.com/bigquery?hl=en), or [Redshift](https://aws.amazon.com/redshift/). This makes ingress easy, as Faraday is a fully composable platform that can integrate with existing data warehouses, functionally sitting on top of them as a decisioning layer, and then automatically routing predictions to channels like [meta ads](https://aws.amazon.com/redshift/), [Shopify](https://www.shopify.com/), or many others. To learn more about our available integrations, take a look at [our integrations landing page](https://faraday.ai/integrations). 

![Faraday adds a decisioning layer on top of your data warehouse](/images/blog/data-warehouse-first-strategy.png) 

If your data isn’t stored in a warehouse, you can upload individual CSV files through our dashboard, but if you do make sure to pay attention to your data hygiene\!

## Data hygiene, what’s that?

Data hygiene is the practice of keeping your data clean, structured, and accurate to ensure the best possible outcomes when using predictive modeling and AI tools like Faraday. Poorly formatted, incomplete, or inconsistent data can lead to mismatches, errors, and weaker insights. Clean data doesn’t just make processing smoother—it directly impacts the quality of predictions and business decisions, **and this is the single most important thing to pay attention to when importing your data into Faraday\!**

The first step to good data hygiene is ensuring consistency in formatting. Datapoints such as names, addresses, and dates should follow a structured format, ideally with first and last names in separate columns and dates consistently standardized in a format like YYYY-MM-DD (I[SO 8601](https://en.wikipedia.org/wiki/ISO_8601)). Additionally, emails and phone numbers should be complete (no 7 digit phone numbers or filler values like “noemail@none.com”) to enable seamless enrichment with third-party attributes in the Faraday Identity Graph. Clean PII helps ensure the best possible enrichment rates for model performance.  
Properly labeled and complete fields prevent confusion and make it easier to match records correctly whereas missing or incomplete records can significantly weaken predictive accuracy. 

Despite the fact that data gaps can be somewhat mitigated by aggregated scoring, better results are always seen when providing a sound dataset from the get-go. If certain data points are unavailable, we can still make predictions, but providing as much information as possible—especially time-bound event data—will dramatically improve our ability to analyze your data quickly and accurately.

So in conclusion, taking the time to check, clean and structure your data before ingesting it into Faraday will save time, prevent headaches, and lead to stronger, more reliable predictions. If you have questions about how to prepare your data, don’t hesitate to reach out to your Strategic Account Manager or Faraday support. 

And speaking of our supportive team…

## Leveraging forward-deployed data science (FDS) for a faster start

Getting your data into Faraday doesn’t have to be a challenge—we offer multiple options to help you onboard seamlessly. Enterprise clients can access Forward-Deployed Data Science (FDS) for hands-on guidance, while all clients benefit from professional services, dedicated support, and self-serve resources to streamline data ingestion.

Our team helps eliminate bottlenecks by structuring and preparing data, optimizing predictive models, and providing strategic guidance. Whether you need help standardizing fields, refining lead suppression strategies, or aligning predictions with business goals, we ensure your models are built on a strong foundation.

For enterprise clients, FDS embeds our experts directly into your workflow, continuously refining models and adapting strategies to market shifts. No matter your level of support, we’re here to help you get started quickly and drive measurable impact from day one.

If you’d like to learn more about this available resource, take a look at our [recent blog post](https://faraday.ai/blog/FDS-changes-the-game) detailing our decision to forward deploy our data science team.

## Final checklist: Setting your team up for success

To make the handoff between your business and technical teams easier, follow this checklist: 

✅ Clearly define your business objective

✅ Identify where relevant data lives 

✅ Ensure core customer identifiers are included 

✅ Provide date fields for key events 

✅ Clarify how conversion rates are measured 

✅ Work with FDS to streamline the process

By following these steps, you’ll avoid common pitfalls, speed up deployment, and get to actionable predictions faster—maximizing the impact of AI-driven insights on your business.

## In conclusion

Getting your data into Faraday doesn’t have to be complex. With the right preparation and support, you can streamline the process, move quickly from data ingestion to predictive modeling, and unlock actionable insights faster.

If you’re ready to learn more or have any questions about preparing your data for Faraday, don't hesitate to reach out. Or if you're feeling prepared to go at it alone, get started for free today, and [sign up here](https://login.faraday.ai/u/signup?state=hKFo2SBTU2FQa0dIbmg4QzBfV0kwTDczbXdiVHc5Wm1mOVB5TqFur3VuaXZlcnNhbC1sb2dpbqN0aWTZIHk0S3JUSGlUMExSSFVPejduNDdBQnpFeHBjZVVjN00xo2NpZNkgMDF6SWFodVNaV203SmRVMFNtWmVNUHMzZ2ZHcklqWTI)\!

To deploy predictive models effectively with Faraday, clients need to start with clean, well-structured data that aligns with their business objectives, ensuring smooth ingestion and accurate predictions. This blog serves as a high-level introduction to this process.

Product