Only new results: why an AI employee should never send the same one twice

An AI employee that runs every day should report each finding once. If Tuesday's email repeats Monday's list with two new lines buried in it, people stop reading by Friday, and the two lines that mattered are missed. The Only new block fixes this by remembering every item the flow has found in its real runs and letting through only the ones it has not seen before, each stamped with the day it was first found.

This article explains why repeats do so much damage, how the block recognises "the same item" from one run to the next, what its memory windows are for, and a few details that decide whether it behaves the way you expect.

Why repeats are worse than they look

• Attention runs out. A daily message is read closely for a week. Once it looks familiar, it is skimmed, then filtered.
• Work gets done twice. Two reps call the same new shop; a lead is created in the CRM again; a road closure is announced to drivers a second time.
• The real signal gets lost. A list of 40 places of which 38 were there yesterday hides the two that are new.
• Trust drops. People start to wonder what else the AI employee is getting wrong.

A person doing the job by hand would never resend yesterday's list. An AI employee has to be told to remember, and that is what Only new is for.

How Only new recognises an item

Recognising "the same one" sounds trivial, but names change, spellings vary and the same shop can arrive from two searches. The block offers two ways:

1. Same place, job or incident (the recommended setting): by the item's own ID. For a place that is its place ID, for a job posting its address, for a road incident its ID, for a ship its MMSI. Items without such an ID fall back to name and location.
2. Same name and location: by the name and the position, to about 100 metres. Use this for rows from your own tables or a webhook, which carry no shared ID.

The first is sturdier. A shop that changes its sign from "Rosa's" to "Rosa's Kitchen" keeps its place ID, so it is not mistaken for a new opening.

The memory window

By default the memory is permanent: once found, an item is never new again. Two shorter windows exist for cases where coming back matters:

The three choices of Only new's window setting.

There is a detail that makes the windows work. Items the block stops are still noted as seen again on that run. A cafe found every week never drifts past the 30-day line, because each week's run refreshes its "last seen". Only an item that genuinely disappears and returns comes back as new.

Details that decide how it behaves

Test runs do not count

Only real runs are remembered: a scheduled run, a run started by a webhook or new rows, or Run once. A test run in the editor treats everything as new and remembers nothing, so you can test as often as you like without using up tomorrow's findings.

Nothing is marked found until it is delivered

If an item reaches Only new but the email or app message after it fails, that item is not remembered. It comes through again on the next run. The memory follows what actually went out, not what was found.

Two runs at once do not double up

When two runs of the same flow overlap, each item can be claimed by only one of them, so the same new place is not sent twice in the same minute.

First runs are long

On its first real run, everything is new. A flow over a whole city may report hundreds of places on day one and a handful a day after. Warn the people who receive it, or run it once before switching on the email.

A made-up week, to see it work

Say a flow searches pharmacies in Lisbon every morning. Monday, the first run, finds 220 and sends all of them. Tuesday it finds 221; the 220 known ones stop at Only new, and one goes on with Tuesday's date as its first-seen day. Wednesday it finds the same 221, and nothing new gets past the block. Thursday a pharmacy that had briefly vanished from the search reappears; with the window set to Ever it stays quiet, with The last 7 days it would also stay quiet because it was missing for only one day. These numbers are illustrative.

Where it sits in a flow

Put Only new right after the block that finds things and before anything that costs effort or reaches people:

1. Schedule and Area of interest.
2. A finding block: Nearby places, Job postings, News or another.
3. Only new.
4. xMap LLM to score or summarise, so only new items are read.
5. Email, Slack, a sheet or your CRM.

Placed before the AI block, it also means the AI only reads what is new each day, which keeps runs short.

Related blocks that are easy to confuse

• Remove duplicates works inside one run: the same shop from two overlapping searches becomes one. It has no memory between runs.
• What changed is the opposite of Only new: it lets through places the flow already knew whose details changed, such as a closure or a rating that moved.
• When something changes is a trigger: it checks around your places and starts the flow only when something new appears, so it never repeats itself either.

Many common setups use them together: Remove duplicates, then Only new for openings, and a second branch with What changed for closures. Ready-made examples are New places every morning, only new, New companies hiring, daily and New outlets to CRM. For the openings side of this in practice, see how to know when a new business opens near you.