How do you write personalized cold emails at scale?
Enrich each prospect row with grounded company website ($0.006/URL) and LinkedIn ($7.50/1k company, $10/1k personal) facts, gate ICP fit with Decision Mode ($1.00/1k), and run one self-healing AI prompt across qualified rows (1.65x tokens plus $0.0035/row, or 0% BYOK on $39/mo Pro) in Google Sheets, Claude, ChatGPT, or Cursor.
What does personalized cold email at scale actually mean?
It means every recipient gets a message that could only have been written for them, produced by a repeatable process rather than by hand. The distinction that matters is between two kinds of personalization. The shallow kind inserts a company name and an industry into a template. The deep kind references a specific trigger: a funding round, a new hire, a product launch, a sentence from the prospect's own site.
That difference is not cosmetic. Industry benchmarks across 2026 put generic name-and-industry personalization at roughly a 1% reply rate, while messages that reference a specific signal land closer to 9%, and signal-based campaigns as a whole report reply rates of 15 to 25% against a baseline near 3.4%. The gap between an average campaign and a top one is mostly the depth of the data the message is built on, not the cleverness of the copy.
Why most personalization fails
Most personalization fails because it runs on data that is not specific enough to be true of one company. A prompt handed only "Acme Corp" writes something true of every company: "I noticed Acme Corp is growing fast." The recipient has read that exact sentence a hundred times and deletes it. The problem is upstream of the writing. You cannot personalize on facts you do not have.
The second failure is doing it by hand. Genuine per-prospect research does work, but a person can research maybe thirty prospects a day, and outbound needs hundreds. So teams either send shallow merge-tag mail to everyone or send deep mail to almost no one. Scale and specificity look like they trade off against each other.
They do not have to. If the specific fact for each row lives in a column, the model can write a specific message for every row in one pass. That is the whole technique: move the research into the data layer, then let one prompt read it.
How to personalize cold email at scale in a spreadsheet
A spreadsheet is the natural place to do this because the data is already there in columns and each step writes its result back as a new column you can inspect and filter. The workflow has three stages: enrich, then write, then run. Sorzio runs all three inside Google Sheets, and it prepares the list rather than sending it, so the output is a set of columns you hand to your own email sender. The full end-to-end version is in the guide to cold email from Google Sheets.
Stage one: enrich every row with a real fact
The opener is only as specific as the column it reads. Before any AI runs, each row needs a fact worth referencing. You run Enrich on the company website or LinkedIn page to pull the content into a new column, so a row that started as just a name and a domain now also carries an "About" column with real text in it. Enriching a 6-page company website bundle costs $0.006 per URL (or $0.005 on Jina), Company LinkedIn costs $7.50 per 1,000 rows, Personal LinkedIn costs $10 per 1,000 rows, and you are only charged for rows that return content. The mechanics are covered in lead enrichment in Google Sheets. This stage is the one that decides whether your personalization is the 1% kind or the 9% kind.
Stage two: write one prompt that reads the enriched columns
A good prompt references the enriched columns by header, not just the name. For example:
Write a one-line cold-email opener for
{{First name}}at{{Company}}. Reference one specific detail from this about text:{{About}}. One sentence, no greeting, under 25 words. If the about text is empty, write a warm generic opener that mentions{{Company}}and nothing else.
Each row substitutes its own values for the merge fields, so the prompt runs once but produces a different sentence per row.
Stage three: sample, then run the full list
Run the prompt on twenty rows first. Read all twenty. Most prompt problems (a tone you dislike, a fact the model invents, an opener that is too long) surface in the first twenty. Fix the prompt, then run the full list. Because the AI runs server-side rather than as a spreadsheet formula, the six-minute Apps Script limit never applies, so one prompt can fill an opener for fifty rows or fifty thousand.
What makes a good personalization prompt?
A good prompt is specific about its inputs, its format, and its failure behavior. Four habits carry most of the quality:
- Reference columns by header. Use merge fields like
{{Company}}and{{About}}so each row gets its own value rather than one hard-coded answer baked into the prompt. - Give a fallback rule. Tell the model exactly what to do when an input cell is empty, so it writes a safe generic line instead of inventing a fact. Invented facts are worse than no personalization, because they are confidently wrong.
- Constrain the format tightly. One sentence, no greeting, a hard word limit. Tight output drops cleanly into a template and is easy to scan when you review the column.
- Sample before you commit. Twenty rows, read every one. This is the cheapest quality control you will ever run.
Personalize after you target, not before
Personalization is the last preparation step, not the first. There is no point writing a brilliant opener for a company that was never going to buy. The order that works is: fix deliverability, clean the list, target against your ideal customer profile, then personalize the survivors. Scoring each row against your ideal customer profile before you personalize means you never spend an AI opener on a company outside your ICP, and it keeps the final list small enough to respect the low per-inbox volume limits that modern deliverability demands.
This ordering also saves money. Every personalized opener costs a fraction of a cent, but on a list of fifty thousand rows the fractions add up, and there is no reason to spend them on rows you will filter out anyway.
What does it cost to personalize at scale?
| Outbound approach | Throughput | Grounded specificity | Effective cost per 1,000 qualified leads |
|---|---|---|---|
| Shallow name/industry merge tags | Unlimited | Low (~1% reply benchmark) | ESP subscription only, high domain burn risk |
| Manual SDR research | ~30 leads per rep per day | High (first-party facts) | Hundreds of dollars in rep labor |
| Sorzio Enrich + Decision Mode + AI | 50,000+ rows in Sheets or MCP | High (self-healing check on 6-page crawl) | ~$21 all-in (verify, enrich, score, opener; $0 on misses) |
| Sorzio Hosted Links microsites | 1 personalized page per row | Interactive page + $0.15 first human open | $0 for unclicked links, bots, and repeat opens |
In Sorzio, managed AI bills 1.65x raw provider tokens plus $0.0035 per succeeded row (or 0% markup with BYOK on $39/mo Pro), Decision Mode ICP gates cost $1.00 per 1,000 rows, SMTP verification costs $0.01 per email, Hosted Links cost $0.15 per first human open, and new workspaces unlock $20 free welcome credit. Run the full workflow in Google Sheets or conversationally on Hosted Lists in Claude, ChatGPT, and Cursor via MCP, automate sequences with Flows, preview real outputs in the showcase, compare Sorzio vs Clay, and calculate spend in the cold email cost calculator or AI cost calculator.
Common questions
Is AI personalization just mail merge with extra steps?
No. Mail merge drops a stored value into a slot in a template, so every recipient gets the same sentence with one word swapped. AI personalization reads a column of real, enriched facts and writes a new sentence per row. The output is different for every row because the input is different for every row, which is exactly what mail merge cannot do.
Can Sorzio send the personalized emails?
No. Sorzio prepares the list across Google Sheets and Hosted Lists (including via Claude, ChatGPT, and Cursor MCP): verify, enrich, score, and personalize. You export the finished columns and send through your own email service provider or sequencer.
How many rows can one prompt handle?
There is no spreadsheet-formula ceiling, because the AI runs server-side rather than as an Apps Script function. One prompt can fill an opener for fifty rows or fifty thousand. The practical limit is your budget and your row count, both shown in the cost preview before the run.
Will personalized email at scale read as automated?
Not if the personalization is real. Recipients spot shallow merge tags instantly, but a line built from a genuine fact about their own company reads as researched, because it was. The work is in the enrichment column, which is why that stage comes first.
Is Sorzio affiliated with reply.io?
No. Sorzio is an independent product built by Empra Consultancy LTD and is not affiliated with reply.io. You can read more on the Sorzio home page.