Watch — a standing question, asked on a schedule

A watch is a standing job, and its output is a timeline rather than a document. Each check holds three things:

• The snapshot — the pages or search results that answered your subject on that run, trimmed to the sentences that matter.
• The change diff — line by line, what this snapshot says that the previous one did not. It is computed mechanically, by comparing canonicalised sentences, so the same page fetched twice with no edits produces no change at all.
• The digest — a few sentences of plain English about what the diff means. This is the one part a model writes.

The timeline is the artefact — there is no separate report — and the Compare tab reads any two checks side by side.

• Frequency — every 3 hours, 6 hours, 12 hours, once a day, every 3 days, weekly, monthly, yearly, or a custom interval in minutes.
• Source — whole internet (web search), or specific websites, one per line.
• Advanced query — optional boolean term groups layered on top of the subject.
• Notifications — in-app bell and email, independently.
• Highlight theme — classic, ocean, violet or mono, for the diff colours on the timeline.
• Ends — an optional end date, after which the watch stops checking.

Checks do not fire on the exact minute. The scheduler adds a jitter of up to 15% of the interval, capped at 20 minutes, so a hundred daily watches do not all hit the same sites in the same second.

• 20 watches per account.
• 30 minutes is the shortest interval, and one year the longest.
• 2,000 characters for the subject.
• 5 websites when you name them yourself.
• 12 sources per snapshot and 6,000 characters kept per source — enough to represent what a handful of pages say, small enough that a year of checks stays a small table.
• Fetching respects each site's robots rules and a per-host delay, so a watch is a polite reader rather than a crawler loop.

Watch has no public share link and no export endpoint. If you need the text of a check somewhere else, copy it out of the timeline.

A check costs at most one text model call, billed through the ordinary credit path and recorded as text usage on your usage dashboard. The arithmetic is deliberately in your favour:

• The baseline check writes a digest, so it costs one call.
• A check that finds a change writes a new digest: one call.
• A check that finds no change makes no model call at all. The previous digest is carried forward and the check is free.

So a watch on a page that rarely moves is close to free to leave open, which is the whole point. If the model is unavailable the check still completes, with a keyword digest written offline.

Can I watch a page by pasting its URL?

You can name up to five sites under Specific websites, but you still have to say what you are watching for. A bare URL tells Watch where to read, not what the answer is.

Does an AI decide whether something changed?

No. Change detection is mechanical: the snapshots are reduced to canonical sentences and compared line by line, and a model only writes the digest afterwards. That is why an unchanged page cannot produce a false alert, and why a no-change check is free.

What is the shortest interval I can set?

Thirty minutes, enforced on the server as well as in the form. Anything faster turns one watch into a crawl loop.

Will it email me on every check?

Only when something changed. A check that finds no difference writes a timeline entry and stays quiet.

Which plan do I need?

Watch is a Pro app, because an active watch is a standing recurring cost rather than a one-off generation.

Can I share a watch with someone?

Not today — there is no public link and no export. A watch belongs to the account that created it.

• Open Watch — the app this guide teaches.
• Watch the web and hear only when an answer changes — the article about it.
• Reader — an AI reading room for the documents you already have
• Operator — autonomous runs with a budget ceiling
• All tutorials