Why Your Competitors' Emails Feel So Good (And How to Match Them with AI)

By Dr. Elena Vasquez, AI Researcher

📧 You open your inbox each morning and notice something: the email from your competitor reads like a personal note from a smart friend, while yours reads like a template that a junior intern filled in on a deadline. You click both links. You compare the two. And you feel that quiet, nagging sense that you're playing a different game than they are.

You're not alone. Most marketing and product teams assume the gap is a budget issue — a bigger agency, a more expensive tool, a larger team. But if you've spent time in AI research, you know that's only part of the story. The real difference is a computational one. Your competitors have already solved the personalization problem with models that weren't available five years ago. And you can use the same ones.

Let's break down what's actually happening inside those "good" emails, and how to replicate the effect.

What "feels good" actually means in an email

When we say an email "feels good," we're usually reacting to three things that compound:

1. Relevance — it speaks to a specific moment, not a generic persona.

2. Specificity — it names the actual product, the actual task, the actual pain.

3. Tone consistency — it sounds like one human, not four humans arguing over a copy deck.

In information-theoretic terms, a "good" email reduces your uncertainty about the sender. You can almost predict the next sentence. A "bad" email maximizes uncertainty. You're left guessing who this is for, what they want, and why you should care.

Consider the entropy of a generic subject line:

$$H( \text{subject line}) \approx \log_2(\text{number of possible audiences})$$

A subject like "Unlock Your Potential" applies to everyone, so its entropy is high. A subject like "Your Q3 onboarding is 62% complete — here's the 8 min fix" applies to a narrow slice of people, so its entropy is low. Low entropy feels specific, and specificity is what readers experience as "this was written for me."

That's the first insight: good emails are low-entropy communications. They narrow the audience so the message can be precise.

The hidden pipeline behind those emails

Behind every "good" email is a pipeline that looks surprisingly simple. You can sketch it in four stages:

[Signals] → [Segmentation] → [Drafting] → [Tone Control] → [Send]

Let's unpack each.

1. Signal collection

The best email teams don't write from imagination. They write from data:

The key is that these signals are combined, not used in isolation. A single signal gives you a segment. Five signals give you a person.

2. Segmentation

Segmentation is the art of grouping people who will react the same way to the same message. In classic marketing, this meant SQL-style buckets: "SMB," "Enterprise," "Churn-risk." Modern teams use latent segments — clusters discovered by models rather than defined by hand.

A simple way to think about this: if you have $N$ users and $K$ features, your user space lives in $\mathbb{R}^K$. Clustering finds subspaces where users behave similarly. You're not just splitting by "size of company." You're finding the axis along which "who is this person actually?" becomes answerable.

3. Drafting with AI

This is where the modern gap opens up. Your competitor's team can now ask a model to draft one email per segment — or even one email per individual user — with a consistent voice. The model isn't writing from scratch. It's writing conditioned on:

$$P( \text{email} \mid \text{user}, \text{segment}, \text{brand-voice}, \text{channel-constraints})$$

That conditional is the whole trick. The model knows the user, the segment, the brand voice, and the constraints (length, CTA, tone). It then produces text that fits all four simultaneously.

4. Tone control

This is the stage most teams skip, and it's the one that makes or breaks the "feels good" factor. A well-drafted email with a mismatched tone reads like a costume. A slightly imperfect draft with a consistent tone reads like a person.

Tone control is a constraint, not a style. It's the difference between "enthusiastic" and "warm." It's the difference between "professional" and "collegial." Most teams should write a one-paragraph voice document and feed it to the model as context.

A concrete example

Let's say you sell an analytics tool. A user in your system has:

• Opened the pricing page 3 times in the last 5 days

• Used the dashboard 12 times, but never the API

• Is in a 300-person company

• Filed a support ticket about "how to export to BigQuery" last Tuesday

A generic email would say: "Unlock the full power of your data with our enterprise features!"

A good email, conditioned on those signals, might say:

Hi Sam,

Noticed you've been checking out our pricing page a few times — and that you've been using the dashboard a lot. Quick thought: if you're in a 300-person org and you're already exporting to BigQuery, you probably don't need the dashboard. You need the API.

Here's a 2-line snippet that gets you from our API to BigQuery in about 8 minutes:

client.export(dataset="q3", dest="bigquery://project/dataset")

Want me to set up a 15-min call to make sure it works in your stack? No deck, no slides.

— Maya

Notice what's happening:

• Relevance — it names the exact behavior (pricing page, dashboard, BigQuery).

• Specificity — it gives a code snippet, not a feature list.

• Tone — it reads like a colleague who knows your stack.

• Low entropy — you know exactly who this is for and what it's for.

That's not a 5-person agency. That's a model with a clean signal pipeline and a voice document. And you can build the same thing.

How to build your own pipeline

You don't need a data team or a PhD in ML. You need four things:

Step 1: Write your voice document

One paragraph. Not a style guide. A voice document. It should answer:

• How do we talk to a new user on day 1?

• How do we talk to a user who's about to churn?

• How do we talk to an enterprise buyer vs. a founder?

• What do we never say?

Keep it under 150 words. The model needs a reference, not a novel.

Step 2: List your signals

Open a spreadsheet. Column 1: signal. Column 2: where it lives. Column 3: what it tells you about the user. You want 8–15 signals, not 80. More signals means more noise.

Step 3: Define your segments

Don't start with 200 segments. Start with 5–8. For each, write one sentence: "This segment is people who ___ and they care about ___." That sentence is your segment's intent.

Step 4: Draft per-segment, then per-user

Start with per-segment emails. Get the tone right. Then, if your tooling supports it, condition on the individual user's signals. The model will produce a version of the segment email that names the user's actual behavior.

Step 5: Add a human review step

Don't let the model send directly. Have one person read 10–20 emails per batch and flag:

• Any sentence that sounds like a template

• Any CTA that feels pushy

• Any mention of a feature the user hasn't used (this is the #1 AI email mistake)

Human review is not a bottleneck. It's the quality layer.

Common mistakes that make AI emails feel worse

A few patterns that betray AI-generated copy:

1. The feature-list paragraph. "With our platform, you can do X, Y, and Z." Readers skip this. Replace with one concrete example.

2. The enthusiasm stack. "We're so excited to share this amazing, game-changing update!" Pick one adjective. The others are noise.

3. The generic CTA. "Learn more" or "Explore our platform." Make the CTA specific: "Read the 2-page case study" or "Book the 15-min call."

4. The tone whiplash. One sentence is casual, the next is corporate. Fix this with your voice document.

5. The unconditioned personalization. "Hi {first_name}!" with no other signal. That's a template, not personalization.

A small experiment you can run this week

Pick your top 3 segments. For each, write:

• Your voice document (1 paragraph)

• 5 signals (from your spreadsheet)

• 1 segment-intent sentence

• 3 draft emails (varying CTA and length)

• 1 human review pass

Send to a small list. Measure: open rate, click rate, and — this is the one most teams skip — reply rate. Reply rate is the truest signal that your email felt personal. People don't reply to templates.

The deeper point

The gap between your emails and your competitors' isn't a budget gap. It's a modeling gap. They've treated email as a computation: given these signals, produce this text, in this voice, for this person. You're still treating it as a copywriting task.

That's a fixable difference. You don't need a bigger team. You need a cleaner signal pipeline, a voice document, and a model that can condition on both.

The good news: the tools to do this are now cheap, fast, and good enough that a 2-person team can produce what a 10-person agency used to produce. The bad news: your competitors already figured this out.

So the question isn't can you match them. It's how fast you'll start.

Dr. Elena Vasquez

PhD, Artificial Intelligence