Most marketing teams aren't starting from zero. You have CRM records, campaign data, transaction history — probably more than you can act on. But having data and having the _right_ data are two different things. And the gap between them is where budget gets wasted, campaigns underperform, and your team ends up doing work that better data would have made unnecessary.

The right question isn't "how much data do I have?" It's "do I have the data that lets my team stop compensating and start executing?"

## What goes wrong when the data is wrong

The same patterns show up for consumer brands across industry verticals — sometimes as a campaign that outright fails, more often as the quiet drain of time your team spends manually compensating for gaps the data should have filled. Here’s how that might look for you:

- A brand builds their entire acquisition strategy around a customer profile that turns out not to reflect their most valuable buyers — and spends years optimizing for the wrong person.
- A paid media campaign misfires because the targeting is built on assumption-based profiles instead of actual purchase signals — the right message reaches the wrong people, and someone spends two weeks in spreadsheets trying to figure out why the numbers look off.
- An inbound sales team wastes hours on leads that were never going to convert, because there's no signal to distinguish high-value prospects from low-value ones until a rep is already on the phone.
- An outbound sales team can't reach half their customer list because the records are missing valid emails, phone numbers, or addresses — and somebody has to go fix that before anything else can happen.

In every case, the brand had data. They just didn't have the data that would have prevented the wrong decision.

## What the right data actually looks like

The right data isn't a single thing. It's three layers, each doing distinct work:

**Verified identity data.** Clean, foundational fields — name, phone, email, address. These make sure you can actually reach your customers, eliminate duplicates, and build a complete view of each household. Everything downstream depends on this being right.

**Customer profile data points.** Curated context like household income, homeownership, Credit Score Proxy, and industry-specific essentials packages (see our offerings for [home services](/blog/homeservices-essential-data), [retail/ecommerce](/blog/retail-ecom-essential-data), and [financial services](/blog/finservices-essential-data)). Use these to segment audiences, tailor messaging, and prioritize outreach based on signals that have actually been shown to drive value in your vertical.

**Custom predictive signals.** Predictive scores trained on your specific outcomes and customers — [likelihood to convert](/templates/lead-prioritization), [best product recommendation](/templates/next-best-offer), [persona assignment](/docs/abstractions/persona_sets). Because they're built on your data, not generic benchmarks, they reflect your business and get sharper over time. Use them to tell your team exactly where to focus: which leads to call first, which product to surface, which customers to retain.

The key is knowing which layer solves which problem. You don't throw all of it at your CRM at once and see what sticks. An informed data strategy means identifying which specific categories of data — and which specific data points within those categories — actually drive value for your business right now. When you get that right, here's what becomes possible.

## What this actually changes

Once you have the right data foundation, the same problems that were draining time and budget become solvable:

- **Know your real ICP.** Instead of building an acquisition strategy around an assumed customer profile, you have the data to understand who actually buys from you and where to find more of them.
- **Personalize at scale.** Instead of generic messaging that treats every customer the same, you have the context to tailor offers, timing, and creative to the specific people most likely to respond.
- **Score and prioritize leads.** Instead of burning hours on leads that were never going to convert, you have predictive signals that tell your team exactly where to focus their time.

## Ready to get the right data?

Most brands we talk to are a few targeted data points away from campaigns that actually work. Explore what's available at [Faraday Pro](/signup) — or [talk to a context consultant](/talk-to-sales) if you want help figuring out exactly which ones move the needle for your business.

Faraday gives B2C brands the verified identity data, customer profile attributes, and predictive signals they need to stop patching data gaps and run campaigns that actually work.

Get the B2C data your brand needs to stop patching gaps and start executing

**See our complete guide:** [Powered by Faraday — embedding predictive AI in your platform](/blog/powered-by-faraday-complete-guide-embedded-predictive-ai).

---

Even the most powerful model in the world can’t drive real value if it doesn’t know who it’s talking to. As our CEO, [Andy Rossmeisl](https://www.linkedin.com/in/rossmeissl), [put it](/blog/ai-needs-consumer-data-to-give-context): “Without the specific identity and history of the customer it’s talking to the world's most advanced AI can do little more than offer a polite, generic greeting."

That's not a model problem. It's a context problem.

## The real issue isn't your model — it's what your model knows

Most AI agents start every conversation blind. Without history or clear signal about who's actually on the other end. So they default to generic responses. Not because they're poorly built, but because they have nothing to work with.

The instinct is to throw more data at the problem. More attributes. More integrations. More inputs. But raw data doesn't tell your agent anything useful about this specific person, right now. You can have a thousand attributes about someone and still not know whether they're price-sensitive, in-market, or a high-value prospect worth a personalized offer.

[As Andy puts it](/blog/ai-needs-consumer-data-to-give-context): "Context is the guiding light that turns a cold interaction into a warm conversion. It is the vital distinction between having a list of names and possessing the clarity to know exactly who a prospect is, and how to best engage them."

That shift — from dumping data at your model and hoping it figures things out, to feeding it curated signals — is what separates agents that plateau from agents that keep getting smarter.

## What Faraday does

Faraday enriches incoming identities against the [Faraday Identity Graph](/catalog) — 1,400+ consumer attributes across 240 million U.S. adults. You send us PII. We return a rich consumer profile, appended to your customer record, ready to pass directly to your agent or model.

For you, it’s as simple as a clean API call that turns an anonymous identity into someone your agent actually knows.

## What this means for what you're building

**Your agent personalizes from message one.** No more [cold-start](/stories/dtc-retailer) problems. The moment an identity hits your system, it can be enriched with the context your agent needs to skip the generic greeting and get to the relevant conversation. This is exactly what [Hazel](/stories/hazel) did with our data — their agentic marketing platform went from knowing nothing about a user to having real customer context on every interaction, with a working prototype in a single day.

**Your agent acts on signal, not noise.** Curated traits — household composition, lifestyle indicators, behavioral factors — slot directly into your agent's reasoning. Platforms like [Heatseeker](/blog/how-dtc-brands-build-living-personas-faraday-heatseeker) work with us to turn their customers' flat transactional records into rich consumer profiles overnight to power smarter personas, better targeting, and a shared language across the org.

**Your roadmap moves at API speed.** What would take a multi-year build internally — sourcing data, hosting it, modeling against it, maintaining it — becomes a few API calls. There's no infrastructure overhead, no six-figure procurement cycle, and no contract to renegotiate as you grow. Brands like [Hazel](/stories/hazel) and [SalesRabbit](/blog/salesrabbit) collapsed timelines from months or years down to days — and scaled on the same infrastructure they prototyped with.

## The bottom line

Your agent isn't underperforming because it's poorly built. It's underperforming because it doesn't know who it's talking to.

Faraday fixes that with a single API call — enriching every incoming identity with 1,400+ consumer attributes before your agent says a word. No data sourcing, no infrastructure build, no long procurement cycle. Just context, from day one.

If you’re ready to accelerate your roadmap and get the context your agent needs today, [start here](/signup) — or [talk to a context consultant](https://faraday.ai/talk-to-sales) about what enrichment looks like for your use case.

Faraday's API gives AI agents the customer context they need to personalize from day one — no data sourcing, no infrastructure build, no long procurement cycle.

Stop sending your AI agent into conversations blind: get customer context via the Faraday API

If you've been following our FIG v2 series, you already know about the [overhauled data catalog](https://faraday.ai/blog/fig-v2-data-catalog) and the broader [announcement of what's coming](https://faraday.ai/blog/now-live-faraday-identity-graph-version2). Today I want to go deep on what I think is the most technically significant upgrade in the entire release: historical data.

## What "historical data" actually means

The [Faraday Identity Graph](https://faraday.ai/catalog) has always given you an up-to-date snapshot of every US adult—what we know about them _right now_. FIG v2 changes that fundamentally. Instead of a single current value for each attribute, we have the full history of values going back many years, each one timestamped and annotated with precision and source metadata.

Think of it this way: rather than knowing that someone's income is \$85,000, you now know that their income was \$62,000 five years ago, \$74,000 three years ago, and \$85,000 today. Each of those data points has a timestamp, a confidence level, and information about where it came from. For you, this means every data attribute isn’t just a point-in-time number, but instead a timeline of historical events.

This is possible because Faraday has been a careful steward of consumer data for over a decade. We've been collecting, normalizing, and preserving these observations the whole time. FIG v2 is the first time we've been able to surface all of that accumulated history in a structured, ML-ready format.

## Why this matters most for modeling

The biggest immediate payoff from historical data is better predictive models. To understand why, it helps to think through how model training actually works.

Let's start with a propensity-to-buy model (i.e., a lead score model). When you build that model, you feed the algorithm historical examples: leads that converted and leads that didn't. The algorithm finds patterns—the characteristics that consistently differentiated the converters from the non-converters—and encodes those patterns into a model. Later, when new leads come in, the model applies those patterns to score them based on which is most likely to buy.

Here's the problem that a model without historical data has: if it’s trained without historical data then it is actually training on what the person looks like _after_ they purchased, NOT the actual time of purchase. Take a home remodeling brand trying to predict which prospects are likely to purchase new windows. To train that model, you feed it historical examples: people who bought windows and people who didn't. But here's the catch—a person looks very different in the data _after_ buying windows than they did _before_. After the purchase, there's a contractor transaction on their record, a home equity line of credit may have appeared, and their home improvement spend has spiked. Those signals weren't there when they were still a prospect. If your training data captures people as they look _after_ converting rather than just before, the model learns the wrong patterns entirely—and the prospects it finds will be people who match a post-purchase profile, not a pre-purchase one.

Until now, this was a problem we could only partially solve. If a brand had leads going back three years, we could append current FIG data to those leads, but "current" data for a lead from three years ago is not the same as that person's data _at the time they became a lead_. A lot has changed in three years.

With FIG v2, we can press rewind and assemble a training dataset based on what they looked like just _before_ they submitted the lead form and what they looked like after they purchased. The model learns what a genuine pre-purchase prospect actually looks like—and gets dramatically better at finding more of them.

## A first-of-its-kind resource for data science teams

Better internal modeling is the most immediate benefit, but FIG v2's historical dataset also opens up something that hasn't really existed before commercially: a comprehensive, historical consumer dataset that's fully ML-ready out of the box.

Every historical observation comes with the kind of metadata that data scientists actually need—data types, statistical types, directionality guidance, coverage breakdowns, and more. If your team is building your own internal models and wants to incorporate rich consumer context that goes back in time, this is the dataset to do it with.

## What's coming next: change detection

We're also excited about something that historical makes possible further down the road: monitoring when a person's profile changes in a meaningful way.

The most valuable signal isn't always the value of an attribute—it's the fact that the attribute just changed. Someone who _just became_ a new homeowner, or _just had_ their first child, or _just crossed_ an income threshold is a fundamentally different prospect than someone who has been in that state for years. Life events create moments of opportunity, and right now, most brands have no reliable way to know when those moments arrive.

Having a full historical vector for every attribute makes this kind of change detection possible. We'll be sharing more about how we're building that capability in Q2.

## Try it this month

FIG v2 is in beta now and releasing this month. If you want to get your hands on the historical data and see how it affects your model performance, reach out to your account management team. And keep an eye out—we have more deep dives coming before the full release.

FIG v2 adds historical timelines to consumer data, enabling far more accurate predictive models by training on pre-purchase profiles instead of post-purchase ones.

Deep dive: FIG v2 historical data and what it means for your models

[![diagram showing how marketing stacks have changed in 2026](/images/blog/bowties-diamonds/bowties-diamonds1.png)](/images/blog/bowties-diamonds/bowties-diamonds1.png)

Something very interesting has happened with the overall topology of a cutting-edge brand’s martech stack. Over the past year or so the pattern I’m seeing among leading brands (including many of our own clients) is a shift from “bowtie”-style stacks, like you see at the left, to “diamonds” like on the right. Here’s how and why this revolution is underway.

## The 2010s martech stack: Hub-and-spoke data warehouse models

Throughout the 2010s it was fashionable (and useful) to imagine any technology-forward organization’s stack as a “hub and spoke” with the data warehouse in the middle. Every component would pull the data it needed from the warehouse and, in return, supply the warehouse with data that other components might need.

In practice this was never quite so easy. Not every software product a brand needed to actually engage its customers natively talked to the warehouse. Sometimes this was for straightforward reasons—it’s hard to build and maintain countless native connectors to every “source of truth” across every vertical—and other times the end product saw the warehouse as a threat: they wanted to own the workflow _and_ the data.

[![diagram showing hub and spoke style marketing stack](/images/blog/bowties-diamonds/bowties-diamonds4.png)](/images/blog/bowties-diamonds/bowties-diamonds4.png)

## The early 2020s: ETL tools and the bowtie marketing stack

In the end, the market spoke, and robust Extract Transform and Load (ETL) tools like [Fivetran](https://www.fivetran.com/) became the crucial connector to feed data into the warehouse. Complementary Reverse ETL tools like Hightouch served the natural inverse role: delivering data from the warehouse back to the engagement tools marketers actually use.

It’s easy to see how central this ETL/RETL role became in the stack’s topology. So much so, in fact, that the “mental model” would often place it, rather than the warehouse itself, in the middle. That’s how we got the bowties of the early 2020s.

[![diagram showing bowtie style marketing stack](/images/blog/bowties-diamonds/bowties-diamonds2.png)](/images/blog/bowties-diamonds/bowties-diamonds2.png)

## 2026 martech stacks: Middleware and the diamond data model

The centrality of the ETL/RETL position did not escape its occupants’ notice. In many cases it was the _only_ tool with visibility on 100% of all of the organization’s data. The incumbent ETL/RETL vendors started to spread their wings, offering value-added services within this prime location—see Hightouch’s decisioning product and its latest agentic offering as examples. Other vendors like [OfferFit](https://www.braze.com/resources/articles/what-is-offerfit) (now a B[raze](https://www.braze.com/) product) sprung up in this “middleware” zone, leveraging data from the warehouse to orchestrate workflows within the engagement platforms.

Faraday, of course, thrives here as well, adding context to an organization’s data on its way to the engagement tools. Products like [cube](https://cube.dev/) are starting to imagine the middleware zone itself as multi-layered, offering normalized semantics that make its neighbors’ jobs easier.

The “source of truth,” meanwhile, at the bottom of the diagram, has crystallized in my mind to a single point. As Scott Brinker so eloquently [put it](https://newsletter.chiefmartec.com/p/the-saas-moats-are-crumbling-but-the-opportunity-is-bigger), once you “strip away the features, the interfaces, the proprietary data lock-in, the ecosystem lip service,” these systems essentially resolve to the same thing: the organization as represented by _data_.

[![diagram showing modern marketing stack](/images/blog/bowties-diamonds/bowties-diamonds3.png)](/images/blog/bowties-diamonds/bowties-diamonds3.png)

## The future of martech: AI agents and constellation structures

As the agentic paradigm deepens its hold on our industry, this diagram—already a far from perfect reflection of reality—will inevitably shift again. Where it goes is anyone’s guess, but I do believe it will be far less linear. Certainly agents will orchestrate everything, but I imagine their purview will extend beyond unidirectional customer engagement. A “constellation” structure is likely to emerge, with agents gathering data from every interaction and disseminating learnings throughout the stack.

The result may look eerily familiar — a modern recreation of the hub-and-spoke structure from a decade ago. Can’t wait to see where we go from there.

Martech has evolved from warehouse-centric “hub-and-spoke” systems to ETL-driven “bowtie” stacks and now to middleware-powered “diamond” models, with the next shift likely toward AI agent–orchestrated, more fluid “constellation” architectures.

Bowties to diamonds: 2026’s new marketing stack

**Update, May 2026:** This feature has been deprecated. The ability to interact with your Faraday personas and predictions is now available in a more powerful form through our [MCP server](https://faraday.ai/docs/features/mcp), which gives AI agents real-time access to your Faraday data and predictions directly. We recommend exploring the MCP as the successor to this feature.

---

In 2026, we all know that effective marketing requires personalization.

But this wasn’t always accepted as fact. Years ago, one of our earliest clients—[a major home goods brand](https://www.bollandbranch.com/)—wanted to confirm the premise for itself. So they ran a rigorous experiment to find out just how much value Faraday’s predictive [Personas](/docs/abstractions/persona_sets) could generate. They committed to a strategy that most marketing teams of that era only dreamed of: instead of sending one generic email to their entire list, they created unique content variants for every single campaign. They had their team write a generic control version, as well as three hyper-tailored versions specifically crafted for the distinct vocabularies and motivations of their unique Personas.

The results were undeniable: [A 30% lift in conversion](/stories/boll-and-branch).

By every revenue metric, the experiment was a massive success. Yet, despite the clear upside, they didn’t stick with the new system very long.

Why would a business walk away from a 30% lift? The answer is simple: because the manual labor required to understand what makes each Persona special, and then write, edit, and stage unique versions of every single email, was simply too hard and tedious to sustain. The friction of the process outweighed the lift in revenue.

This is the paradox of personalization: The more you tailor your messaging, the better your results—but the more work you have to do.

## Talk to your Personas

We built a new chat interface to solve that tedium.

Faraday has always been great at helping you understand _who_ your customers are. We have a range of tools—from [AI-generated summaries to salience trait detection](/blog/new-release-trait-salience)—that help you investigate exactly how they differ. But at the end of the day, you still have to figure out how to apply those insights and actually _talk_ to them.

Now, you actually can. Our new chat feature provides a familiar AI interface, but we preload the chatbot with all the in-depth details about your Personas. You don't have to teach the AI who your customers are. It already knows!

However, the agent does _not_ natively know anything about your brand. This is by design—it allows you to give the chat whatever context, guidelines, or quarterly goals you want so you get highly relevant outputs. You just chat!

[![Chat with personas feature in Faraday's dashboard](/images/blog/persona-chat/persona-chat-unfilled.png)](/images/blog/persona-chat/persona-chat-unfilled.png)

## Power up your marketing

This feature allows you to move from passive analysis to active execution immediately. Here is a sneak peek at a few ways you can start using it today:

- **Match Personas to campaigns:** Describe a specific product launch or holiday promotion and ask the chat which Persona in your set is most likely to respond.
- **Brainstorm tailored offers:** Identify an underperforming segment and ask the AI to suggest specific promotions or incentives that align with that Persona’s unique motivations.
- **Generate personalized copy at scale:** Paste your baseline messaging into the chat and instantly receive variants for subject lines, email body copy, and landing pages tailored to each Persona.
- **Build synthetic focus groups:** Save time and budget by asking the AI to generate a panel of virtual customers that proportionately mirrors the distributions of your actual Persona set.
- **Enable Sales roleplay:** Improve outreach by asking the agent to "act" as a specific Persona so your sales team can practice overcoming objections and perfecting their pitch.

_(Note: The chat does not currently retain memory or history between sessions, so you will want to provide your brand context in each new conversation.)_

## How to talk to your Personas

This feature is included at no additional charge for clients on our usage-based pricing (UBP) plans.

You can find the new chat feature directly within your existing Persona sets:

1. Go to a Persona set on the Dashboard.
2. Click the **Chat** tab at the top of the page, or simply add `/chat` to the end of the URL.
3. Start talking to the interface as you would with any other LLM.

[![Chat with personas feature in Faraday's dashboard](/images/blog/persona-chat/persona-chat-filled.png)](/images/blog/persona-chat/persona-chat-filled.png)

## A note on privacy and safety

We know that using AI with customer data requires trust. The chat agent **does not have access to PII (Personally Identifiable Information)**. The model operates on the aggregate patterns and statistical definitions of your Persona groups, not the private data of individuals.

## Ready to start?

The days of choosing between personalization and productivity are over. By talking directly to your Personas, you can finally operationalize the deep insights you already have, turning static data into dynamic content in seconds.

_Keep an eye out for our next post, where we will dive deep into step-by-step playbooks and actual prompts for getting the most value out of these new use cases!_

Faraday’s newest feature lets you talk to your Personas using an AI agent—preloaded with your unique Persona definitions—to instantly generate targeted marketing copy, subject lines, and campaign strategies without manual tedium.

What's new: Talk to Personas

We recently announced the biggest update to the Faraday platform ever: an overhauled version of the [Faraday Identity Graph (FIG)](https://faraday.ai/features/faraday-identity-graph). FIG is the foundational dataset on which all other services at Faraday are built and if you want the big picture, check out our [main announcement](https://faraday.ai/blog/now-live-faraday-identity-graph-version2).

But today, I want to zoom in and talk about one specific, highly anticipated aspect of FIG v2: the data catalog.

## Starting from scratch

Over the last decade, we’ve learned a lot about how to structure data, and we’ve been continuously improving that whole time. But for FIG v2, rather than continue to make incremental updates, we took those years of learnings and rebuilt the dataset from scratch—ensuring every attribute is standardized, strictly categorized, and crystal clear.

Nowhere is that rigor more obvious than in our new data catalog.

## A better data dictionary

This new data catalog goes far beyond a standard data dictionary. It features an entire dedicated page for every single attribute in FIG v2.

[![FIGv2 data catalogue](/images/blog/fig-v2-data-catalog/data-catalog.png)](/images/blog/fig-v2-data-catalog/data-catalog.png)

Historically, a lot of this highly detailed data has only existed internally at Faraday and clients were required to go through their account management team to get these details. Now, for the first time, we are publishing it so that all of our users can get direct value out of it.

[![FIGv2 data catalogue example](/images/blog/fig-v2-data-catalog/data-catalog-example.png)](/images/blog/fig-v2-data-catalog/data-catalog-example.png)

When you land on these new attribute pages, you won't just find basic metadata like names and descriptions. We are showing you the true depth of the data by surfacing the really important technical details and historical context you need to make decisions, including:

- **Data distribution:** Exactly how many people are observed in each category, or how a numerical attribute is distributed across a histogram.
- **Technical properties:** The API name, category, unit, data type, statistical type, allowed values, and deprecation status.
- **Coverage breakdowns:** Exactly how many individuals have a known value versus a value imputed from their address or geography.
- **Data derivation:** Whether a value was directly collected/observed, modeled, or survey-based — so you know exactly how each attribute was produced.
- **Time series and contribution history:** A look at how many observations we've collected over time and the authorities that contributed them.
- **Directionality and interpretation:** Guidance on how to use the data (e.g., how missing values should be interpreted—does a missing value mean "we don't know," or should you assume it's "false"?).

## Built for marketers and data scientists alike

How you use this level of transparency really depends on what you're trying to build.

**If you’re a marketer,** let's say you’re looking for a specific piece of context to personalize an email campaign. In the past, finding that exact attribute usually meant digging through a dense, traditional data dictionary or—more likely—having to reach out to your account manager and wait for an answer. Now, you can skip the middleman, search the catalog yourself, and find exactly what you need to make a campaign feel bespoke.

**If you're a data scientist**, say using Faraday's data for internal modeling, things like data type and directionality matter immensely—they have a massive impact on the resulting accuracy of the models you're building. Instead of hunting for those specs, all of those crucial technical details are now front and center.

## Easier to browse, more to discover

When you start searching through the new data catalog, you might notice the data feels much cleaner. By doing a massive amount of consolidation and normalization behind the scenes, we’ve made FIG v2 much easier to navigate. Better yet, that cleanup created headroom for us to bring in additional vendors and add entirely new attributes.

So go ahead and explore—there is a lot more context about American consumers waiting for you.

**Interested in a specific data point?** If you're looking for a particular attribute or curious about our coverage for a specific demographic, we’re happy to pull those details for you. Just send a note to your account management team directly.

## What comes next

As we noted in our [main announcement](/blog/now-live-faraday-identity-graph-version2), FIG v2 is set to officially release in April. However, we are currently looking for beta testers to get in early. If you're interested in getting your hands on the new data catalog and taking it for a spin, reach out to your account management team to learn more.

And keep an eye out—we'll be dropping more deep dives into FIG v2 in the coming weeks\!

Explore the new Faraday Identity Graph (FIG) v2 data catalog to access detailed attribute pages, visual data distributions, and technical metadata designed for marketers and data scientists.

Deep dive: FIG v2 data catalog

_Faraday automatically groups your customers into distinct, actionable subgroups based on shared traits. To ensure your segments are highly useful, we dynamically scale the number of Personas based on your total audience size. This gives you the perfect level of granularity for personalized marketing._

In 2026, marketers know that a generic message is a campaign killer. Your buyers are unique. Blasting your entire list with the same offer won't win the deal anymore. In this context, the question is: how do you effectively segment a massive, messy audience into groups you can actually talk to? And the answer is generally by building customer [Personas](https://faraday.ai/docs/abstractions/persona_sets).

But how many Personas do you need? If you have a customer list of 1,000 people, organizing them into a few distinct personas makes sense. But what if your list is 100,000 people? Or 10 million?

If you are still squeezing a massive, diverse audience into those same basic Personas, you aren't personalizing at all. You are just generalizing at a slightly more granular scale.

## The problem with traditional Persona clustering

Historically, many data scientists have used a technique called the “elbow method” to help decide how many personas (or clusters) to create. The method tests different numbers of groups—one, two, three, four, and so on—and measures how tightly each customer fits within each grouping. As you add more clusters, the fit improves, but the improvement slows over time; each additional cluster helps a little less than the one before it. The “elbow” is the point where adding another group no longer meaningfully improves the separation.

In theory, that elbow represents an optimal number of personas.

In practice, real customer data rarely produces a clean, obvious elbow. Human behavior overlaps. Traits blend together. Instead of a sharp bend in the curve, you often get a gradual flattening. When that happens, the algorithm generally defaults to a standard output: three clusters. And often, three is not what you need

## The myth of the "true" persona

The elbow method breaks down because it is optimized to find a mathematically correct answer. It is hunting for what data scientists call “true clusters”.

But when our product team talks to clients, we often say: “In marketing there is no such thing as correct personas, only useful personas”.

For a mathematically true persona to exist, your audience data needs a massive, undeniable gap. For example, if your brand sells two entirely unrelated things—like skateboards for teenagers and mobility aids for seniors—your buyers will naturally split into two separate demographic groups. The math easily finds that gap. But if your groups are that distinct, you probably didn't need clustering tools to point them out to you at all.

For almost every other brand, human data is messy. Your customers have overlapping traits and nuanced behaviors. Because those massive gaps don't exist, the elbow method gets confused and tends to lump everyone into the default number of buckets: three. But often, three personas is not what you actually need.

## Segmenting for usefulness, not perfection

To build personas that drive value, we realized we needed to stop chasing mathematical perfection and start optimizing for utility.

Instead of relying on an algorithm that forces arbitrary limits, Faraday has rewritten the rules. We now dynamically scale your persona count based directly on the size of your audience. Using a proportional sizing model, Faraday now assigns a persona count ranging from a minimum of 2 up to a maximum of 8 based on the volume of your cohort.

Here is why this changes the game for your campaigns:

- **Granularity that scales with you:** An audience of 1,000 yields 3 personas. 10,000 yields 4\. 100,000 yields 5\. Massive lists finally get the rich segmentation they deserve.
- **Characteristics shine through:** Because large audiences are no longer crammed into three tiny boxes, the unique traits that define your buyers become visible and actionable.
- **No more micro-segments:** Conversely, smaller cohorts will never be artificially chopped into too many tiny, useless personas. Your groupings stay distinct and easy to manage.

Marketing isn't about hitting an arbitrary number of segments. It's about finding the best way to speak directly to your customers' unique characteristics. With proportional personas, you have the exact right number of segments to make your next campaign your highest-converting one yet.

**Ready to see your new segments?** Log in to Faraday to explore your dynamically scaled persona sets today, or read the updated docs for more details on analyzing your audience.

Faraday has replaced traditional persona clustering with a dynamic scaling model that automatically adjusts the number of personas based on audience size, ensuring marketing segments are always granular enough to be actionable without becoming overly complex.

Smarter segmentation: why we rewrote the rules on Persona sizing

_Why do AI tools so often fail to drive ROI in marketing? Frequently, the problem isn't the model—it’s the inputs. To drive real value, AI needs to understand the people it’s engaging with. Without context like financial signals, lifecycle indicators, or real-world constraints, agents miss the meaning behind behavior—like whether someone buying baby clothes is shopping for their own child, signaling future purchases, or buying a one-time gift for a baby shower. Faraday provides essential consumer data so agents have context that allows them to act with nuance and precision._

## What AI knows, and what it doesn’t

You likely recall your first conversations with an LLM chatbot. You probably noticed that the AI knows almost everything about the world, but absolutely nothing about you: the customer standing right in front of it. Indeed, if you set this giant brain to complex tasks like writing a short story or generating a video of a monkey, it jumps into action. But ask an LLM a seemingly simple question, like “How should I talk to the person who just landed on my pricing page?” and suddenly you remember that AI doesn’t actually know everything. Without the specific identity and history of that visitor, the world's most advanced AI can do little more than offer a polite, generic greeting.

This is the paradox of the agentic era. The "raw muscle" of Large Language Models (LLMs) has placed every piece of shared human knowledge at our fingertips, but that means that shared knowledge is no longer a differentiator. In a world where every competitor has access to the same models and internet, the superpower isn’t volume of information—it’s relevance.

The winner today isn’t the brand with the most information; it’s the one with the specific knowledge that illuminates the path to value. Faraday gives your AI agents the consumer context they need to stay grounded in reality, enabling them to confidently orchestrate engagement while your competitors are still teaching their agents how to guess.

## Your AI needs context

The goal of any agentic AI customer engagement workflow is to move beyond generic automation and toward true situational awareness with 1:1 personalization. In the high-stakes environment of modern commerce, generic anything—whether it is messaging, targeting, or personalization—isn’t just inefficient, it’s a campaign killer. If your AI cannot distinguish between a price-sensitive churn risk and a financing-ready buyer, it will default to a “middle-of-the-road” response that resonates with no one.

**Context is the guiding light that turns a cold interaction into a warm conversion.** It is the vital distinction between having a list of names and possessing the clarity to know exactly who a prospect is, and how to best engage them. Grounded with this context, the model sees that a specific visitor isn't just a generic web session, but a high-intent lead with a unique persona and a high predicted lifetime value: giving you the clarity to act.

### The context budget

But there is a catch. Once you solve the data scarcity problem, you’ve got a new one: [too much data increases entropy instead of reducing uncertainty](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents).

When AI underperforms, the natural instinct for many is to "feed the beast" by dumping in every available row of first-party data or loosely stitched third-party attributes; however, this approach hits a hard technical ceiling: the [context budget](/blog/context-engineering-and-demographic-data).

Every AI model has a limited attention window. If you flood that window with uncurated information, the agent wastes its processing power on noise rather than signal and starts to hallucinate. What’s worse, even if the AI is not overwhelmed, this internal or uncurated 3rd party data likely lacks the specific answers the AI actually needs to answer the questions that actually drive value. Things like:

- Who is this person?
- Are they a good fit for my brand?
- What’s the best next action to ensure conversion and maximize value?

Think about it like this: if you’re trying to decide whether to market a home renovation project to someone, is knowing their blog click history or whether they have a dog or a cat really helpful? No. You need to know whether they own their home or rent it—and whether they’re likely to have the means and intent to take on a renovation.

## Faraday’s context makes your AI better

To succeed, your systems need a complete picture of the consumer, but they also must operate within strict processing limits.

Faraday balances this equation by judiciously infusing your agentic workflows with context. We curate and deliver the deep external knowledge your agents need to thrive. We allow you to solve the context gap while simultaneously compressing that complexity into clean, clarity-driving signals that don’t overwhelm your context budget. We understand how agents operate, so we provide specific context they actually need—tailored to your industry and business goals.

At Faraday, we offer two kinds of context for AI agents:

- **Consumer data:** We close the "context gap" with carefully curated external details missing in your internal systems. From financial health to lifestyle details and shopping habits, this data helps the agent understand the human behind the screen.
- **Predictive intelligence:** We distill thousands of signals into clear guidance. Instead of flooding the agent with a tempting smorgasbord of potentially irrelevant details, we use traditional machine learning techniques to boil it all down to predictive indicators like "churn risk," "purchase likelihood.," or “next best offer.”

Finding the right blend of context—one that maximizes lift without overwhelming the context budget—is an art in itself, but Faraday includes analytics and tooling to help you make an informed decision.

## Better context, better outcomes

This isn't just a technical theory; [we are seeing it work in the wild right now](/blog/hazel-agentic-marketing-consumer-context-faraday-api).

Take [Hazel](https://fromhazel.ai/), an agentic marketing platform that provides brands with an "AI coworker" to analyze market trends and customer behavior. To fulfill that promise, Hazel’s agent needs to know more than just what a customer bought yesterday; it needs to understand who that customer is outside of the transaction: their demographics, lifestyle, and purchasing power.

Hazel faced a classic "[cold start](/stories/dtc-retailer)" problem, and traditional data brokers demanded six-figure upfront fees and multi-month engineering projects just to get access to the data they needed. So instead, Hazel integrated with Faraday. By plugging into our platform, they started receiving consumer data from the Faraday Identity Graph immediately, via API and MCP. And this data made a real impact:

- **Speed to value:** They reduced a potential multi-month build-out to a single day of prototyping.
- **Cost efficiency:** They eliminated over **$100k in upfront data licensing fees**, shifting to a usage-based model that scaled with their growth.
- **Privacy-first:** They gained a built-in privacy compliance infrastructure, allowing their team to focus on agent logic rather than regulatory hurdles (such as building a system to handle data access and deletion requests).

The result? Hazel’s agent now has instant, privacy-safe access to rich consumer context, allowing it to make optimal decisions for their clients without the engineering overhead. This is the blueprint for the next generation of software: specialized agents running on specialized context.

## Context and AI: fuel for the engine

If the rest of the AI industry is building the "engine"—the LLMs and autonomous agents—Faraday provides the fuel. An engine without fuel is just a heavy piece of hardware; an AI model without context is just a convenient way to deliver the wrong message.

As we move deeper into the agentic era, the distinction between winners and losers won't be defined by who has the smartest model, but by who has the most relevant context to feed them.

**Ready to power your agents with the context they need? Let’s talk.**

AI models are powerful, but without consumer data to provide real-world context, they can’t make relevant, high-value marketing decisions.

Your AI needs context. Faraday’s got it

See our complete guide: [How predictive lead scoring works — and why it outperforms traditional methods](/blog/predictive-lead-scoring-b2c)

_The optimal target for predictive lead scoring is the final **revenue event** (e.g., "Closed Won" or "Funded Loan") rather than intermediate funnel stages like "Appointment Set."_

_Intermediate stages often capture leads based on immediate need (e.g., a broken appliance) regardless of their budget, leading to high volume but low conversion. By training on revenue data, models filter out these low-fit leads and prioritize leads who have the **financial capacity** to purchase, significantly increasing ROI._

## The priority problem: why score at all?

Before discussing what to predict, we have to acknowledge why we are doing this. Predictive scoring exists because your most expensive resources—time, money, and people—are finite.

If you had an infinite marketing budget and a sales team that never slept, you wouldn't need a model; you’d just treat every lead like a VIP. In reality, lead scoring is the filter you use to manage scarcity:

- **The call center:** Your reps only have 8 hours to make dials. They need to know which calls to answer first to hit their quota.
- **The service tech:** Your technician only has four service windows today. You want them at the house most likely to result in a full system replacement, not just a $50 patch-job.
- **The budget:** You can’t afford to send a high-end direct mailer or a field rep to every door. You need to know who gets the premium treatment.

By predicting the outcome, you aren't just "sorting data"—you are ensuring that your team's limited hours are spent on the opportunities with the highest potential value. But this leads to a critical tactical error: in the rush to find "value," many teams point their AI at the wrong target.

## Avoiding the middle-funnel signal trap

In the world of predictive modeling, we spend a lot of time obsessing over the "how." _How_ do we ingest this data? _How_ do we tune the algorithm? But we rarely spend enough time on the "what." Specifically: What precisely are we trying to predict?

This sounds like a trivial question. If you run a home services company or a financial institution, you might say, "I want to predict who becomes a customer."

But in a multi-stage funnel, "becoming a customer" is a journey, not a moment. You have leads, you have appointments, you have quotes, and finally, you have revenue. The most common mistake I see intelligent teams make is setting their sights on the middle of that funnel—predicting for appointments or quotes—rather than the end.

It feels intuitive. Appointments happen more often than sales, so you have more data. But predicting for appointments is a **signal trap**. To understand why, you have to understand the difference between a broken furnace and a healthy bank account.

## The two types of propensity: situational vs. intrinsic

At Faraday, we’ve found that consumer behavior isn’t monolithic. It’s actually composed of two distinct forces. If you mix them up—or ask your AI to predict the wrong one—your model breaks.

### 1. Situational propensity (the "broken furnace")

Imagine it’s January. It’s freezing. Your furnace starts making rattling noises. You are not proactively looking for an upgrade. You aren’t browsing HVAC websites because you love the tech. You are cold, and you have a **situational need**.

You will almost certainly call a provider. You will book an appointment. You might even get a quote. But here is the kicker: **This event is random.** No amount of third-party data can tell us your furnace is about to die on a Tuesday. When you ask a model to predict a coincidence, you’re asking it to be a psychic, not a strategist.

### 2. Intrinsic propensity (the ideal customer fit)

Now, let’s look at what happens _after_ the technician arrives. You have the need (situational), but do you have the means? Do you have the credit score to finance a $12,000 system? Do you have the home equity? Are you the type of homeowner who invests in premium solutions, or are you looking for the cheapest patch-job possible?

This is **intrinsic propensity**. Unlike the random breakdown of a machine, these are stable, knowable attributes about you and your household.

## Why intermediate stages undermine model effectiveness

Here is where the "appointment prediction" strategy falls apart.

If you train your AI to predict who will book an **appointment**, you are asking it to predict the broken furnace. You are asking it to find people in crisis. Since data providers can't see inside houses, the model grasps at straws, finding noisy correlations that don't hold up.

Worse, it optimizes for low-fit leads. We’ve confirmed this through experimental testing: when we try to predict intermediate stages, we consistently see low signal. It’s only when we target the final revenue stage that the model can bring all of its "tricks"—our data and tech—to bear. You fill your sales team’s calendar with activity, but you starve their commission checks.

However, if you train your AI to predict **revenue** (Closed Won), the model changes its behavior. It accepts that it can’t predict the furnace breaking (the noise), so it focuses entirely on finding the optimal fit for the solution. It stops chasing every 'hand-raiser' and instead highlights the leads where the need and the budget actually align.

### Comparative analysis: choosing your prediction target

| Target Stage             | Signal Type              | Predictability | Business Impact | Recommended?           |
| :----------------------- | :----------------------- | :------------- | :-------------- | :--------------------- |
| **Inquiry / Lead**       | Purely situational       | **Low**        | High noise      | **No**                 |
| **Appointment / Quote**  | Need and interest        | **Medium**     | Low-fit risk    | Only if data is scarce |
| **Revenue / Closed Won** | **Intrinsic: ideal fit** | **High**       | **Maximum ROI** | **Yes: best practice** |

## Strategic guidance: defining your attainment class

If you run marketing or product, your job isn’t to deploy AI; it’s to drive value. AI tools can certainly help you do that, but you have to give them the right context. If you point it at "appointments," you’re asking it to be a psychic. If you point it at "revenue," you’re asking it to be a strategist.

To operationalize this, follow this checklist:

- **Don’t settle for volume.** It is scary to see your lead volume drop when you switch targets. But if you’re filtering out the people who were never going to buy anyway, that drop is actually efficiency.
- **Predict the fit, not the worry.** Use third-party data to find long-term customer alignment, not temporary situational crises.
- **The Golden Rule:** Always predict the **Ultimate Revenue Producing Signal**.

## The bottom line: ask the right question

Your predictive model is only as smart as the target you give it. If you ask it to find "activity" (appointments), it will fill your calendar with noise. If you ask it to find "value" (revenue), it will build your business.

The shift from volume to value can be daunting—your lead counts will drop, and your sales team might panic at the empty space on their calendars. But that empty space is an opportunity. It is time saved from chasing tire kickers, which can now be spent closing the customers who actually have the capacity to buy. Don't optimize for the middle. Optimize for the end.

---

### FAQ: improving lead scoring performance

**Q: Why is my lead scoring model prioritizing low-value leads?**

**A:** You are likely predicting for situational propensity (e.g., who needs the service _now_) rather than intrinsic propensity (who can afford the service). If you train on "appointments," the model will find people in crisis regardless of their ability to pay. Shift your target variable down-funnel to "revenue."

**Q: What if I don't have enough data at the "revenue" stage?**

**A:** If you lack sufficient volume at the bottom of the funnel (typically \<500 events), move up one stage (e.g., to "quote generated"), but proceed with caution. Acknowledge that you are introducing "situational noise" and rely on your sales team to qualify for budget.

**Q: My sales cycle is extremely long (6+ months). Can I really wait that long for a feedback loop?**

**A:** In long sales cycles, you may need a proxy variable. Look for a "committed" stage—like "underwriting approved" or "contract sent"—that happens earlier but has a high correlation to final revenue. The key isn't that "Closed Won" is the _only_ acceptable variable; the key is that you must identify the stage in _your_ specific cycle that concretely indicates revenue potential (capacity) rather than just interest (activity).

To maximize the value of your lead scoring, you must target the funnel stage that directly produces revenue (e.g., "Closed Won"), not intermediate steps like appointments. Targeting revenue ensures you predict actual buying power, whereas appointments often just capture curiosity or immediate need.

Predictive lead scoring best practices: Optimize for revenue, not appointments

**See our complete guide:** [What is the Faraday Identity Graph?](/blog/whats-in-faraday-identity-graph).

---

When brands partner with Faraday, they aren’t just looking for a score; they need to understand the mechanics behind that score. A common question we receive is: "How exactly is the system acting upon our data?"

While the underlying math is complex, the process is straightforward: identity resolution and enrichment, model training and validation, and deployment. Here's how each step works.

[![diagram showing how customer data flows through Faraday](/images/blog/faraday-under-the-hood.png)](/images/blog/faraday-under-the-hood.png)

## Identity resolution and enrichment

The process begins when your first-party data, whether it’s historical orders, lead lists, or churn records,is securely ingested into the Faraday system. But we don’t just store these rows; we resolve them.

We match your PII identifiers (emails, phone numbers, addresses) against the Faraday Identity Graph (FIG), a proprietary consumer dataset that contains over 1,400 attributes on more than 240 million U.S. adults and their households. This step is critical because your raw data is often fragmented or incomplete. By resolving identity first, we link your scattered records to rich, privacy-safe consumer profiles. This ensures our models are analyzing a complete person rather than a disconnected data point such as their address.

Once identity is resolved, the system enriches your data with relevant attributes from FIG, ranging from demographics to lifestyle interests. This enriched dataset becomes the foundation for modeling.

## Model training and validation

Rather than relying on manual guesswork or generative AI such as ChatGPT, our models identify complex patterns in behavior that a human analyst might miss.

We first create a training set that includes your examples (customers who converted) and counter-examples (those who didn't). We then test various algorithms, such as Random Forests and Logistic Regression, on this enriched dataset. Each algorithm evaluates thousands of potential attribute combinations to find which approach best fits your data and business goals. We often combine multiple techniques to improve performance.

For instance, the model might discover that home ownership combined with recent purchase frequency and regional trends strongly predicts conversion for your business. This automated testing finds hidden signals that may take a human analyst months to find.

Before deployment, we validate the model on fresh data (holdout) to confirm the patterns hold up using methods, such as the 5-fold-cross-validation, to ensure model performance.

## Deployment and transparency

Finally, the validated model generates actionable outputs that answer your business questions:, “Who is most likely to convert,” and “Which products will be most appealing to them.” The system then assigns a Propensity Score (percentiles from 1–100) to each subject for which one has been requested.

This process is transparent by design. Propensity scores are pushed directly back alongside explainability reports that details the factors driving each prediction, and how valid the predictions are. We don’t just deliver the prediction; we detail exactly _why_ the prediction was made, giving your team full visibility into the model’s decision-making

## Conclusion

In short, Faraday works by turning incomplete customer records into enriched profiles, then trains validated models that predict behavior. By combining secure identity resolution with the latest in machine learning, we ensure that the insights you get are not only accurate but fully explainable and ready to drive revenue.

Faraday turns your first-party customer records into privacy-safe, identity-resolved profiles, trains and validates machine-learning models on the enriched data, and then deploys transparent propensity scores (with explainability) that predict who’s most likely to do what next.