Set up overnight AI competitor monitoring

What to watch, before you watch anything

Most competitor monitoring fails before the first check ever runs, because nobody decided in advance what’s worth watching or what a real change looks like. “Watch our competitors” can’t run on a schedule; “tell me if Acme changes its pricing page” can. Decide which competitors and which of their pages: pricing, product and feature pages, release notes, help pages, app store listings. That’s where a real change shows up first, not a generic homepage.

The harder decision is what counts as worth an interruption. Write it down once, so it isn’t re-judged every night: a change to pricing, supported channels, model options, availability, approval or privacy controls, or a major launch is material. A copy edit or a reordered page is not. That distinction is what keeps a nightly brief short enough to read instead of noise you start skimming past.

Three scheduled checks, not one

A single check a night is cheap, but it misses timing. If a competitor’s page changes and then gets corrected within a few hours, one check might catch only the correction and never see the original change. Three checks spread through the night, for example at 1:00, 4:00 and 6:30, catch more of what happened and let the operator note roughly when a change first appeared, rather than only that it existed by morning.

Each check repeats the same routine: visit the named pages, compare against the last check, and log the difference, if any. That’s mechanical work, and it’s meant to be. The judgment about whether a difference is material happens once, later, when the checks get pulled together into a brief.

A fast model for scanning, a stronger model for the brief

Watching pages is a narrow, repetitive job: fetch, compare, flag a difference. A fast, inexpensive model handles that well, run three times a night without costing much to keep the watch running. Writing the brief is different work: reading three checks’ worth of raw differences, deciding which ones cross the bar you set, and explaining why a given change is worth your attention. That’s worth a stronger, slower model, run once, after the last check of the night.

This split is where “any model” stops being an abstract claim: you can see which model did the scanning and which one wrote the brief, and swap either one without losing the watch’s history.

One run, worked through

That’s an example run: it shows the shape of the flow, not a screenshot of a specific night. The competitor facts inside it are real and sourced, drawn from Meta Connect’s September announcements about Muse, so the illustrative brief holds up if you follow the links. A real captured run, once we have one, replaces it without changing anything else on this page.

Notice what the brief doesn’t do: it doesn’t say “Meta is winning” or editorialize about what the change means for your roadmap. It states what changed, links to where, and stops. Reading judgment into it is still your job, the same reason the note to a co-founder is a draft, not a sent message.

What waits for your approval

Checking pages, comparing them and writing the brief happen without asking you anything, since none of it leaves your workspace or affects anyone else. The moment that needs a decision is narrower: if the watch turns up something material, it can draft a short note and hold that note for your approval, rather than sending it on your behalf. You see the exact text before it goes anywhere, and you can approve it as written, edit a line, or hold it entirely.

That single approval point is deliberate. A competitor watch that only reads public pages doesn’t need much gating; one that also messages a colleague about what it found does. Drawing the line at “send” rather than “notice” keeps the watch unattended without making it unaccountable.

What’s in the receipt

Every run leaves a record: how many checks ran and when, how many pages or sources got checked, which model did the scanning and which one wrote the brief, and exactly when anything was sent, if it was. If a drafted note went out, the receipt shows who approved it and at what time, down to the minute. If you need to check why a message to a co-founder went out, or confirm a scheduled check ran at 4 a.m. rather than being silently skipped, the receipt is where that answer lives, not a memory of what you think happened.

When it doesn’t go cleanly

Not every night is a clean three-for-three. A source can go down mid-check, a pricing page can 404, a help center can move behind a login it wasn’t behind the week before. Say so rather than skip the source silently: “Instinct’s pricing page was unreachable at 4:00, retried at 4:15, succeeded” beats no mention of Instinct at all, because it tells you the watch is still working even when one source isn’t cooperating.

The more common outcome, most nights, is quieter: nothing material changed. That’s not a failure, and the brief should say so plainly, something like “no material changes; one minor wording update on a help page.” A one-line confirmation is worth more than silence, because silence leaves you unable to tell “nothing happened” from “the watch didn’t run.”

The honest limit

This only monitors what’s public. It reads the same pricing pages, product pages and release notes anyone could open themselves; it runs on a schedule, at hours you’d rather not spend refreshing tabs. If a competitor makes a change through a channel your operator can’t reach, a login-gated portal, an invite-only rollout, a conversation at a conference, it won’t show up in the brief. The watch is careful, repeated reading of pages already open to everyone, nothing more.

Setting it up

1. Name the competitors and the specific pages: pricing, product, release notes, help pages, app store listings.
2. Write down, once, what “material” means for you, so it doesn’t get re-judged differently each night.
3. Pick a schedule, three checks overnight is a reasonable default, and a time for the brief.
4. Decide where the brief should land, and whether a material change should also draft a note for someone else, held for your approval.

From there, the watch runs itself. See always-on for how the scheduling and background pieces fit together, and receipts for what gets recorded on every run.