Personalisation Line Generator

Give it a list of company domains. Get back one factual opening line per company, grounded in that company's own website, with the quote that proves it.

When we cannot ground a line, you get an honest HOLD instead of a generic sentence. You are not charged for those.

---

What you get

Feed it five domains:

{
  "domains": [
    "monzo.com",
    "gocardless.com",
    "starlingbank.com",
    "wise.com",
    "some-dead-domain.example"
  ],
  "style": "observation",
  "maxWords": 18
}

You get one row per company, ready to paste into your sending tool:

domain	status	line	evidence
monzo.com	OK	found that you build current accounts and lending entirely inside a mobile app	quote from /about
gocardless.com	OK	found that you handle recurring bank payments and Direct Debit for subscription businesses	quote from /features
starlingbank.com	OK	found that you run business current accounts with built in bookkeeping tools	quote from /business
wise.com	OK	found that you move money across borders at the mid market exchange rate	quote from /pricing
some-dead-domain.example	HOLD (site_unreachable)

Four usable lines. One honest blank. You pay for four.

Every OK row carries evidence_url and evidence_quote, so you can click through and check our work in about five seconds. If a line is wrong, you can see exactly which sentence on their site produced it.

Who this is for

• Cold email teams who want a real first line per prospect instead of a merge field
• Agencies running outbound for several clients at once
• Founders doing their own outreach who do not have time to read 300 websites
• Anyone with a lead list who already knows that a generic opener gets ignored

If you are sending to companies and you want the first sentence to prove you looked, this is for you.

How it works

YOUR DOMAINS
    |
    v
1. FETCH      We read the company's homepage and up to two internal pages.
              If that is not enough to say what the company does, we run one
              web search limited to the company's own site and read up to two
              more of its pages. Search results are only used to find pages;
              the quote always comes from a page we read ourselves. robots.txt
              is respected. Sites that do not respond, or that still render
              nothing without JavaScript, are held rather than guessed at.
    |
    v
2. EXTRACT    We pull out what the company actually does: its service phrases,
              the city it operates in, the language the site is written in, and
              whether this is a company at all rather than a directory listing.
    |
    v
3. GENERATE   A language model writes one line from those extracted phrases.
              It never sees the raw page, only the extracted text, and it has
              to tell us which snippet it used. That snippet becomes your
              evidence quote.
    |
    v
4. CHECK      Enforced in code, not asked for in a prompt: the word cap, no
              em or en dashes, nothing negative about the company, and a check
              that the evidence actually supports the line. A line that fails
              is regenerated once, then held.
    |
    v
5. DEDUPE     Across your whole batch, no two lines may read alike. A collision
              is regenerated once, then held as duplicate_risk, naming the
              company it collided with.

What makes it different

1. Grounded or held. Every line comes with the sentence from the company's own website that supports it, and the URL it came from. If we cannot find that sentence, we do not write the line. Generating from a company name and a domain is what produces the openers that read like mail merge, and it is the failure mode we built this to avoid.

2. Batch unique. No two lines in your run may read alike. This sounds obvious until you have sent a campaign where most of the list received the same sentence with the company name swapped in. Recipients compare notes, and the whole batch reads as automated the moment two of them do. The deduplication runs across the entire run, not just neighbouring rows.

3. You pay only for usable lines. OK lines are charged. HOLD rows are free. Not discounted, free. That is the whole pricing model, and it is deliberate: if we charged for HOLDs we would have a reason to ship a weak line rather than hold it, and you would have to trust that we did not. This way our incentive and yours point the same direction.

Pricing

$5 per 1,000 OK lines, plus a start fee of $0.00005 per run. HOLD rows cost nothing.

• A line you can use is charged. A blank with a stated reason is not
• Duplicate domains in your input are collapsed before the run starts, so the same company is never charged twice
• Try it first: turn on preview to process only your first 10 domains. A first run costs at most about $0.05. It uses the identical engine, the same rules and the same evidence, so what you see is what you get. It is a spend cap rather than a free tier

Most tools in this category sell you a subscription plus a bundle of credits that expire whether you use them or not, and charge you per line whether the line is usable or not. Here there is nothing to sign up for, nothing expires, and unusable lines are free. That is the comparison that matters, and it is why this exists as an actor instead of a SaaS product.

Chains with

This actor takes domains and returns lines, so it drops into the middle of almost any list building workflow:

• Any company or lead scraper on the store that outputs a website column
• Lead List Cleaner (ours) to deduplicate and normalise a multi source list before you spend anything generating lines for the same company twice
• Company Lookalike Finder (ours) to build the list in the first place, then personalise it here
• Your sending tool, by exporting to CSV or JSON

The natural order is find companies, clean the list, then write the lines.

Limitations, in plain language

Read this before you buy. It is here so you do not need a refund.

HOLD is a feature, not a failure. A generic line burns your reply rate and costs you the prospect. An honest blank costs you nothing and tells you to skip that row or write it yourself. We would rather return 70 good lines and 30 blanks than 100 lines you have to read one by one to find the bad ones. If you want a tool that always produces something, this is not it, and that is on purpose.

JavaScript only sites cannot be read. We do not run a browser. A site that renders nothing without executing JavaScript comes back as js_only. This is a real gap and it is more common on small business sites built with page builders.

Some sites tell us not to read them, and we listen. If a site's robots.txt asks automated readers to stay out, we do, and the row comes back as robots_blocked. Re-running will not change it. We could ignore that file and most tools do, but an actor that crawls sites which asked it not to gets reported and delisted, and then it is not there for you either.

Non English sites are held, not translated. By default we accept English only, and a site in another language comes back as wrong_language with the language we detected. You can widen this with the languages input, but the line will be written in that language, and our own testing is strongest in English.

Thin sites cannot support a specific line. A one page brochure site that says "quality service since 1998" genuinely does not contain anything specific enough to write about. Those come back as too_thin. We are not going to invent a detail to fill the gap.

Your list quality decides your HOLD rate. A list of real companies with real websites holds very little. A scraped list full of directory pages, parked domains and dead sites holds a lot, and every one of those is free but also means one fewer usable line. Cleaning the list first is worth doing.

This describes companies, not people. No names, no job titles, no emails, no personal details of any kind. The line is about what the company does. If you need contacts, chain this with a contact finder.

We do not scrape LinkedIn. Not as a source, not as a fallback. This keeps the actor stable and keeps you clear of the takedown risk that follows tools which do.

One style. Every line is an observation about what the company does, grounded in a sentence from its own site. A question style was tried and removed before launch: asked as questions, the lines came out as asking a company how it does the very thing it sells.

customContext steers the angle, it never adds a claim. Telling us what you sell makes us more likely to lead with the relevant true thing about a company. It cannot make us write something their site does not say. If you describe an offer that has nothing to do with a company, you get a line about what that company actually does, or a HOLD.

Input

Field	Required	Description
domains	yes	1 to 60 company domains per run. Split a longer list across runs
style	no	observation, the only style
maxWords	no	5 to 40, default 18. Hard cap, enforced after generation
customContext	no	Up to 300 characters on what you are reaching out about. Tilts the angle, never adds a claim
languages	no	ISO 639-1 codes, default ["en"]. Other languages are held
preview	no	Process only the first 10 domains, capping a first run at about $0.05

Output

One row per input domain, whether it produced a line or not: domain, status, line, hold_reason, evidence_url, evidence_quote, matched_phrases, words, style, language, hold_detail, final_domain, pages_fetched, generatedAt.

status is OK or HOLD. On a HOLD, line is empty and hold_reason is one of seven documented values: site_unreachable, robots_blocked, js_only, too_thin, non_company, duplicate_risk, wrong_language.

hold_detail is one sentence saying why a row was held. A duplicate line names the company it read too much like, so you can see the pair for yourself.

final_domain is where the site ended up after redirects. It usually matches domain. When it does not, the company we described is not the one on your list row, and that is worth a look before you send.

Rows are pushed as they are generated, so if you stop a run early you keep everything produced so far. Export to CSV, JSON or Excel, or pipe it straight into another actor.

FAQ

What HOLD rate should I expect? It depends almost entirely on your list. A clean list of real companies with working sites holds very little. A raw scraped list holds a lot. Every HOLD is free, so a bad list costs you nothing except the lines you did not get.

Do I need an API key for anything? No. Not for us, and not for the language model either. Some tools in this category ask you to bring your own paid model key and pay that bill separately. Everything here is included in the per line price.

Can I use the lines as they are? Yes, that is the point. Every OK line is capped at your word limit, contains no em or en dashes, and says something the company says about itself. Read the evidence quote if you want to check one before sending.

Why did one company get held when its site looks fine to me? Open the row and read hold_reason. The usual answers are that the site needs JavaScript to show its text, or that it says very little in words a machine can read, or that its line was too close to another line in the same run.

One row came back describing a different company. Why? Check final_domain. Some domains redirect to a parent or group site, so the page we read genuinely belonged to that company rather than the one on your row. We record both instead of guessing which one you meant.

Can I run the same list twice and get different lines? Yes, the wording can change a little. A language model writes each line, so a second run over the same domain can phrase it differently, and now and then a domain that got a line once is held the next time, or the other way round. What does not change is the rule: every line still has to be supported by a quote from that company's own site, and you are still charged only for the lines you get.

Can I schedule it? You can, but this is a batch tool rather than a monitor. Run it when you have a new list.

What if a line is wrong? Send us the domain and the line. The evidence quote makes these fast to diagnose, because we can see exactly which sentence produced it. A line that its own evidence does not support is a bug we want to know about.

Support

Issues and questions through the Apify listing. We aim to acknowledge within 24 hours and to fix genuine breakage within 72 hours.

The actor runs an automated health check every day, so breakage is usually found and fixed before you would notice it.