Copy

AI personalization is a merge field with extra steps

By Travis Heurtelou · 5 min read · September 14, 2026

The short version

Scraping a profile and generating a compliment for line one is the same as a first-name token. Readers can tell.
The model belongs before the email: deciding who's on the list, reading what the signal means, sorting what comes back.
A variable is worth having only if deleting it changes what the next sentence says.
Why-now comes from the list, not the writer.

What the tools sell

Most 'AI personalization' takes a LinkedIn profile or a website, generates a compliment about a recent post or the company's growth, and puts it in line one. 'Loved your post on hiring in Q3.' 'Saw [[Company]] is expanding into Denver.' Then the same pitch as everyone else.

The reader has had forty of these this month. The compliment has nothing to do with the ask, so it gets filed where 'I hope this email finds you well' gets filed. It's a merge field that cost more to fill.

Where the model earns its place

Before the email, and after it. In order of how much each matters:

Deciding who's on the list. Given a company's website and a description of the buyer, does this company fit? A model can read two thousand sites in an hour and answer that. A filter on industry code and headcount can't.
Reading the signal. A posting for a second estimator means something different from a posting for a receptionist. A 4.2 rating on 80 reviews means something different from a 4.2 on 6. The model turns the raw signal into the one fact the email opens on.
Sorting what comes back. Out of office, wrong person, not interested, interested, interested but later. Getting the fourth one to a human within the hour is worth more than any first line.

The variable test

A personalization variable is worth having if removing it changes what the next sentence can say. 'You're hiring a second estimator' leads to 'so the bid volume is there and the pipeline to feed it isn't.' Delete the hire and the next sentence has nothing to stand on.

'Congrats on the new office' leads nowhere. Delete it and the email reads the same. That's decoration.

The other tell is grammar. A real variable changes the shape of the sentence around it: 'your two locations' against 'your location', 'the estimator role' against 'the crew lead role'. If every version of the email is the same sentence with a different noun dropped in, it's a template, and it reads like one.

Why-now comes from the list

Every email has to answer two questions the reader is asking: why me, and why now. Why-me is who they are. Why-now is what changed: they posted the role, the rating dropped, they moved buildings, they lost a contract. A fact about them from the last few weeks.

No amount of rewriting produces a why-now. If the list was built on 'HVAC contractors in Texas' there's nothing to say about this week, and the writer ends up inventing urgency ('as we head into Q4'). When the copy feels like it's missing something, the fix is usually a different list.

Give the thing first, then ask

One more pattern from the tools: the first email asks the prospect for something (fill in this form, hop on a call, answer three questions) in exchange for a free thing later. Reverse it. The deliverable needs nothing from them. The homework, if there is any, is step two, after they've replied.

Use the model to build the list and read the signal. Write the email yourself, under 75 words, on the one fact the signal gave you. If there's no fact, there's no email yet.

Questions

Is any first-line personalization worth doing?

The kind that's also the reason you're writing. 'You're hiring an estimator' is both. 'Great post' is neither.

What about letting AI write the whole email?

It writes the statistical middle of every cold email ever sent, which is exactly what the reader has learned to ignore. Have it draft if you want, then cut it down to the one fact and the one question.

Doesn't research at list-build time cost more?

Per row, yes. Per reply, no. You send fewer emails to people who fit instead of more emails to people who don't.

The cold email that gets a reply is under 75 words

6 min read

Build the list from signals, not filters

7 min read