DEV Community

Cover image for An AI outreach planner that is only allowed to name domains it was given
Petteri Pucilowski
Petteri Pucilowski

Posted on Edited on

An AI outreach planner that is only allowed to name domains it was given

LinkCensus answers one question from a dated Common Crawl edition: which sites share linkers with mine? Two sites are "neighbours" when the same domains link to both. The more linkers they share, the closer the neighbours.

A neighbour list is useful but raw. People asked the obvious next question: "Fine, who do I actually pitch, and with what?" So LinkCensus now has one AI feature, the free outreach planner.

It's the only place LinkCensus calls a language model. Search results, reports and exports stay model-free, and the method page says so.

The constraint that matters

The model gets the result rows you already saw and nothing else. For each row that's the domain name, the number of shared linkers, and up to five of those linkers. It never sees page content and never browses.

That makes the obvious failure easy to rule out. An LLM asked for outreach targets will happily invent plausible ones. So the server doesn't trust the output:

allowed = {r.domain.strip().lower() for r in rows}
segments = [
    Segment(..., domains=[d for d in (x.strip().lower() for x in s["domains"])
                          if d in allowed])
    for s in data["segments"]
]
segments = [s for s in segments if s.domains]   # drop segments left empty
Enter fullscreen mode Exit fullscreen mode

Any domain not in your input is dropped, and a segment left empty is removed. The plan can be wrong about why two sites are related, but it can't send you to a site the data never showed.

Two other rules do most of the remaining work:

  • Structured output. The response must match a JSON schema (segments with name, why, angle and domains, plus first steps), so there's no prose to parse.
  • "Unclear - check first" is an allowed answer. The model may only infer what a site is when the name makes that obvious. Otherwise the site goes into this segment instead of getting a made-up description.

A real run

I took the rows for simpleanalytics.com (privacy-friendly analytics) from the 2026-01 edition. The free preview shows five neighbours: trustpilot.com (20 shared linkers), mailerlite.com (18), brevo.com (18), easings.net (17) and google.com (10).

The plan it returned:

  • Email and marketing platforms: mailerlite.com and brevo.com. Angle: a joint guide on privacy-friendly analytics alongside email campaign tracking.
  • Review and trust platforms: trustpilot.com. Angle: claim or improve a listing in the software category.
  • Unclear - check first: easings.net. "Does not clearly indicate what the site is from its domain alone." Correct: it's a CSS easing reference, and you'd want to look before pitching it.
  • Large platforms and search: google.com. "Do not pitch directly." Also correct.

The first steps rank Trustpilot first because it has the most shared linkers. They also point out that the shared linkers themselves are worth checking for list or directory pages.

Where it's honestly weak

Shared linkers are noisy at the top. Domains like google.com, wordpress.org and github.com link to almost everything, so they show up in many rows. The planner is told that shared linkers are the strongest signal and that list or directory linkers are worth pitching themselves. It still can't tell you which page on wordpress.org links to whom; that needs the full report.

The free preview is also capped at a few neighbours, so this plan covers five sites rather than your whole neighbourhood.

Cost and limits

Claude Haiku with low effort, well under a tenth of a cent per plan. Each IP address gets 4 plans a minute and 30 a day, with a server-wide daily cap on top. Nothing is stored.

Try it with your own domain: https://linkcensus.com/tools/outreach-plan

Top comments (0)