Build Log AI research pipeline

How we built an AI-search radar that writes our newsletter

Keeping up with AI search by hand is a losing game. The field moves every week across scattered sources. So we built a radar: it reads five live sources, keeps only what is new since last run, and drafts the write-up. This is the system, end to end.

Published June 2026 by Balázs Turán, Creative Data Engineers.

What we built
  • One radar, five live sources: X, YouTube, vendor docs and changelogs, AI engine answers, and Reddit, pulled fresh every run.
  • Capture raw, then keep only what is new. Every finding is stored without opinion, and a dedup ledger remembers what it saw last time, so each run surfaces only genuine change.
  • An AI drafts the report, a human approves it. Nothing publishes on its own.
  • The output is the AI Visibility Report, our free fortnightly newsletter on where AI search is heading.
Why we built it

A manual scan does not survive contact with the week

AI search changes faster than any one person can track. The signal is spread across a dozen vendor changelogs, the people worth following on X and YouTube, what the engines themselves answer, and the community arguing it out on Reddit. We wanted a repeatable system that does the reading, remembers what it already saw, and hands us a draft, not a manual scan we have to repeat from zero every week.

The rule we set ourselves: the machine collects and remembers, the machine drafts, and a person decides what goes out. That split is what makes it both fast and trustworthy. The collectors stay dumb and honest, the writing happens once, and the judgement stays human.

The system

From signal to newsletter, end to end

Three stages run in order: collect five live sources, sort and write by keeping only what is new and drafting it up, then publish and send. A watchlist on top decides what is in scope, and the whole thing repeats every week.

Sources we pull Our pipeline AI writes it

Hover or focus any box for a plain-language explanation.

WATCHLISTS Watchlists who and what we track the lists decide what we track A Collect Five live sources, pulled fresh every run SOURCE X / Twitter voices and traction SOURCE YouTube channels and reviews SOURCE Docs and changelogs what changed this week SOURCE AI engine answers what the engines say SOURCE Reddit engagement-ranked sample every finding, captured raw B Sort and write Keep only what is new, then turn it into a report keep only new then write it up STORE Raw capture findings, no opinions DEDUP New only what is new since last run SYNTHESIS Write the report an AI drafts it saved and published every week C Publish and send A weekly report that becomes a fortnightly newsletter every two weeks to the inbox every week, on repeat REPORT Weekly report saved and published NEWSLETTER AI Visibility Report fortnightly READERS Subscribers straight to the inbox

This diagram is the engine that produces the report. What happens next, how the report becomes a newsletter that subscribers join and leave, is its own system. See the newsletter system →

What we learned

Three decisions that make it hold up

1

Collectors stay dumb and honest. Every source emits raw, structured findings with no summarizing. The judgement comes later, in one place, so the inputs stay auditable and we can always trace a claim back to where it came from.

2

The dedup ledger is what makes it sustainable. An append-only record of what we have already seen means each run is only the change since last time. Without it, every run would re-surface the same items and the noise would bury the signal.

3

The machine drafts, a person decides. An AI writes the report, but a human reads it before anything reaches a reader. Speed where it helps, judgement where it matters.

FAQ

Questions we get about the radar

Is the radar fully automated?

The collection and the first draft run on their own. A person reads every report before it becomes a newsletter, so nothing reaches a reader unreviewed.

What does the radar run on?

Five source collectors, a raw capture step, an append-only dedup ledger that remembers what it has already seen, and an AI write-up step. The collectors stay opinion-free, so the judgement happens once, at the write-up.

How is this different from a news feed or an alert?

A feed gives you links. The radar captures findings raw, keeps only what is new since the last run, and drafts an analysis of what changed and why it matters.

Can you build something like this for us?

Yes. It runs on the same engineering as our AI Search Visibility work. Book a call and we will scope it.

This is the kind of system we build for clients

The radar runs on the same skills we bring to client work: read the sources that decide AI answers, capture them cleanly, and turn them into something a team can act on. That is the heart of our AI Search Visibility work.

We also teach operators to build systems like this, hands-on. Balázs runs the build sessions at Agent-J+.