Guide

Put a watch on a web page: monitor changes with a scheduled AI task

To monitor a web page for changes with AI, set up a weekly scheduled task in ChatGPT or Claude Cowork with the page’s link, a copy of today’s wording and this instruction: “Every week, compare this page with the text below. Report only the differences, quoting old and new side by side.” Most weeks the report is empty. When a clause moves, it shows both versions, so you can check the change against the live page before anyone acts on it. It works on public pages only and can miss a change, so treat it as an early warning and keep the formal review.

AI web page monitoring. AI web page monitoring is a scheduled assistant task that rereads a public web page on a fixed cadence, compares it with a saved copy of its wording and reports only what has changed.

Written by Tayyeb Mahmud, founder and Principal Architect. Claude Certified Architect – Professional · LinkedIn profile of Tayyeb Mahmud

1AYM is an OpenAI Select Partner.

Checked 11 October 2026. OpenAI’s Help Center articles on scheduled tasks and on ChatGPT Work and Codex, OpenAI’s ChatGPT documentation on automations, web search, using ChatGPT and teams, Anthropic’s Help Center articles on scheduled tasks in Cowork, getting started with Cowork, Cowork on Team and Enterprise plans and using Cowork safely, the Claude Cowork product page, the Claude Academy Cowork course and Anthropic’s documentation on reducing hallucinations were read on this date. Plans, limits and menu names change, so check a detail before you rely on it.

Put a watch on a web page · 47 seconds, with captions

Why a public change sits unread for a month

A supplier’s terms page says the fuel surcharge is reviewed quarterly. One month it says monthly. Nothing was hidden: the new wording sat on a public page the whole time, and the first anyone heard of it was the invoice.

The same thing happens with a regulator’s guidance, a framework’s tender notices and a competitor’s price list. The information is published. What is missing is anyone whose job it is to reread a page that almost never changes, because doing it by hand every week is dull, the answer is nearly always that nothing moved, and it is the first task to go when a week gets busy.

That makes it a good job for an assistant on a schedule. OpenAI says ChatGPT can run recurring tasks and monitor for changes [1], and Claude Cowork can run saved tasks on a schedule [3]. A weekly task rereads the page, compares it with wording you saved and stays quiet unless something moved. When something does, you see the old and new text in the week it changed, rather than at renewal or at audit.

The prompt: two sentences, the link and today’s wording

Start a new chat in ChatGPT, or a new scheduled task in Claude Cowork. Type the two sentences below, then paste the page’s link and the current wording of the part you care about underneath. Copy that wording straight from the page rather than summarising it, because the comparison can only be as exact as the text you give it.

Set it to run weekly. For terms, guidance and price pages, a week is soon enough to act, and a modest frequency keeps the load on someone else’s website light.

The weekly watch prompt

Every week, compare this page with the text below. Report only the differences, quoting old and new side by side.

[Paste the link to the page here]

[Paste the wording as it reads today]

Most weeks the report says nothing changed. When a clause moves, from “reviewed quarterly” to “reviewed monthly”, it quotes both versions, and you check the quote against the live page before anyone acts on it.

Once you have accepted a change, paste the new wording into the task as the new baseline. Leave the old one in place and every run after it reports the same change again, which is how a useful alert turns into one people ignore.

Why the two sentences work

Each phrase in the prompt closes off one way a page watch goes wrong.

“Compare this page with the text below”: The comparison runs against wording you pasted, not against what the assistant remembers. That matters most in Claude, where each scheduled task runs as its own Cowork session [3], so there is no memory of last week to compare with. ChatGPT’s monitoring tasks can use information from previous runs [1], and the pasted text still earns its place there: you know exactly what the page is being measured against, and so does anyone who inherits the task.
“Report only the differences”: Silence is the default, so a report that does arrive is worth opening. OpenAI’s advice for scheduled tasks is a prompt that says how to decide whether anything is worth reporting [2]. Here that rule is four words long.
“Quoting old and new side by side”: You get evidence rather than a summary to take on trust. A quoted clause can be found on the live page with a text search; a paraphrase cannot. Anthropic’s guidance on reducing hallucinations makes the same move, asking Claude to work from word-for-word quotes and to withdraw any claim it cannot support with one [7].

Test it before you trust it. OpenAI recommends running a prompt by hand in a normal chat before you schedule it, then reviewing the first few runs [2]. Run the watch once against the live page straight away and expect to hear that nothing changed. If it reports differences on day one, the wording you pasted does not match the page, and every weekly run would have been noise.

Worked variations for other pages

The core instruction stays. What changes is how much of the page you point it at and what counts as a difference.

A long page with one clause that matters: Paste only that clause as the baseline and add: “Compare only the section headed [heading].” Footers, dates and banners stop producing reports, and the clause you care about is the only thing being compared.
A regulator’s guidance page: Add: “Ignore changes to dates, links and formatting. Report changes to what the guidance asks organisations to do.” A new review date with no change in substance should not reach your inbox.
A competitor’s price list: Paste the table as text and type: “Every week, compare this page with the table below. Report only the prices, plan names and limits that differ, quoting old and new side by side.” The quoted rows are what you forward to whoever sets your own prices.
The line worth adding to every version: “If you cannot open or read the page, say so instead of reporting no differences.” A page that failed to load and a quiet week can look identical in a report. OpenAI’s own advice is to ask the assistant to say plainly when it could not reach a source, because an explicit gap is easier to deal with than a confident guess [9].

Where it lives in ChatGPT and Claude

Both products can run the watch. They differ in how you set it up, how often it can run and whether a run knows what the last one saw. The table is from the vendors’ own pages as read on 11 October 2026; plans and menus change, so check a detail before you rely on it.

Setting up a weekly page watch in ChatGPT and in Claude, from the vendors’ own pages as read on 11 October 2026
Dimension	ChatGPT	Claude
Plans	Eligible plans. How many active tasks you can have depends on the plan	Paid plans (Pro, Max, Team and Enterprise) where Cowork is switched on
Setting it up	Ask for the task in a chat and confirm it, then manage it under Scheduled	In Cowork, Scheduled > New task > Create with Claude. Describe the job, choose weekly, check the proposed name, schedule and instructions, then click Schedule
How often	Up to once an hour on paid plans	Hourly, daily, weekly, on weekdays or manually
Between runs	Monitoring tasks can use information from previous runs	Each scheduled task runs as its own Cowork session, so the pasted wording is the baseline
Seeing a change	Settings > Notifications, then Push, Email or both	Open each run in Cowork
Where it runs	A desktop task that uses local files needs the computer on and the app running	In the cloud, so your computer can be asleep, unless the task needs local files or apps
For a team	In Enterprise workspaces, Team Tasks run scheduled work through a team’s service account	Owners can turn Cowork off for the organisation. On Enterprise, running Cowork in the cloud is off until an owner turns it on

The ChatGPT rows are from OpenAI’s Help Center article on scheduled tasks [1] and its documentation on automations [2] and on teams [11]. OpenAI’s article on ChatGPT Work and Codex also lists monitoring for changes among the things Work can do through scheduled tasks [10]. The Claude rows are from Anthropic’s articles on scheduled tasks in Cowork [3], on Cowork for Team and Enterprise [4] and on getting started with Cowork [5]. All were checked on 11 October 2026.

Two of those differences change how you run the watch. In Claude the pasted wording carries the whole comparison, so keep it current, and plan to open the runs yourself, because a watch nobody reads protects nobody. On Pro and Max accounts that have moved to the new combined Claude, you can also create the task by describing it in any conversation [3].

Scheduled runs draw on your plan. Anthropic says Cowork uses a plan’s limits faster than chat [13], and OpenAI says inactive tasks may pause automatically [1]. One weekly watch is light; forty of them is a decision for whoever manages the plan.

What goes wrong, and what it cannot do

The watch is an early warning. It is useful because it takes one prompt to set up and stays quiet, and those same qualities set its limits.

It can miss a change: The assistant decides what counts as a difference, so a one-word edit low on a long page, “may” becoming “will”, can slip through. Read a quiet week as no news, not as proof. Contract reviews, regulatory monitoring and price checks that someone owns stay exactly where they are.
Public pages only: It reads what anyone can read. A page behind a login, a supplier portal or a site that blocks automated readers will not work, and a password never belongs in a task. A page you have to sign in to is a job for a proper integration.
A failed read can look like a quiet week: Add the line from the variations above, so a page that would not load comes back as a gap rather than as no change.
The baseline goes stale: Update the pasted wording each time you accept a change, or the same change is reported every week until people stop reading the reports.
The prompt guides, it does not guarantee: “Report only the differences” is an instruction the assistant tries to follow. Expect the odd run that flags a cosmetic change or adds commentary, and judge the watch on its first few runs before anyone depends on it [2].
Keep it read-only, in the work account: The watch needs a public link and public wording, nothing else. Run it in the account your company has approved, keep confidential notes such as contract values or the reason you are watching out of the task, and connect nothing that can send or change things. Anthropic’s Cowork course says scheduled tasks should draft for your review before they send anything [12].
The page is outside content: Whatever the page says flows into an agent. OpenAI says to treat web results as untrusted input [8], and Anthropic says the risk of prompt injection in Cowork is not zero and that you remain responsible for what Claude does on your behalf, scheduled tasks included [6]. A read-only task with nothing connected has little to misuse. Let a watch act on what it reads and it becomes a fuller agent, which needs the controls to set before any AI agent goes live.

Where 1AYM fits

Setting up and rolling out AI work tools is core work for 1AYM, ChatGPT Work and Claude Cowork among them. One real workflow per team is how a rollout starts, and a watch on the pages a team depends on is a good one: procurement on supplier terms, compliance on guidance, the commercial team on competitor prices.

When a watch outgrows an assistant, because a missed change would be expensive or the list runs to dozens of pages, it becomes our AI integration and automation services work: a scheduled job with monitored runs and an audit log, which stores every version of the page and posts each change to its owner in Slack. If you have a list of pages worth watching, book a call or email enquiries@1aym.com.

For engineers: state between runs, failure modes, cadence and when to move it into code

The mechanics behind the two sentences, from the OpenAI and Anthropic pages listed in the sources, as read on 11 October 2026.

State lives in the prompt: Each Claude scheduled task runs as its own Cowork session, so the pasted baseline is the only state between runs. ChatGPT monitoring tasks can carry information from earlier runs, and a task scheduled inside a chat returns to that chat with its existing context. Pasting the baseline anyway makes both behave the same way and makes a run reproducible: anyone can repeat the comparison by hand with the same text.
Scope the input: Paste the clause or table, not the whole page, and name the section by its heading so the model can still find it after a redesign. A smaller baseline leaves less to misread and fewer cosmetic differences to report.
Make every outcome explicit: Ask for one status line on every run, such as “No differences”, “Differences found” or “Could not read the page”, so an empty week, a changed week and a failed fetch are three different reports rather than one silence. Review the first few runs before relying on it, as OpenAI recommends.
Least privilege: Keep the task read-only. OpenAI’s automations guidance is to start unattended tasks with the narrowest access that lets them succeed, and the watched page is untrusted input to an agent, so a task that can read one public URL and write nothing has the smallest blast radius.
Cadence and quotas: ChatGPT runs tasks up to once an hour on paid plans and caps active tasks by plan: 5 on Plus, 10 on Business and Edu, 15 on Pro and Enterprise. Claude offers hourly, daily, weekly, weekdays or manual runs, and Cowork draws on plan usage faster than chat. Weekly is the sensible default for terms and guidance.
Where it runs: Claude runs scheduled tasks in the cloud unless they need local files or apps, and on Enterprise, cloud running is off until an owner turns it on and grants it to a group. In ChatGPT, a task created inside a project cannot read that project’s files, and a desktop task that uses local files needs the computer on and the app running.
Ownership: A watch in one person’s account stops when that person leaves. In ChatGPT Enterprise workspaces, Team Tasks run in the cloud through a team’s service account and its configured app connections, so the watch belongs to the team rather than to whoever set it up.
When to move it into code: If a missed change has a real cost, put the watch in a scheduled job that fetches the page, stores every version and diffs it deterministically, and use the model only to summarise the diff. A diff of the stored text cannot miss an edit to that text; the model makes the alert readable. The assistant version is still the right first step, because it shows which pages are worth that build.

Sources

Questions people ask about watching a page with AI

Can ChatGPT monitor a website for changes?

Yes. OpenAI says ChatGPT can run recurring tasks and monitor for changes, and that monitoring tasks send a notification when a relevant update occurs. On an eligible plan, ask for the task in a chat, manage it under Scheduled and choose Push, Email or both in Settings > Notifications. Paste the page’s current wording into the prompt so every run compares against the same baseline.

Can Claude monitor a web page for changes?

Yes, through scheduled tasks in Claude Cowork on paid plans. In Cowork, open Scheduled > New task > Create with Claude, describe the job and choose weekly. Each scheduled task runs as its own Cowork session, so paste the page’s current wording into the instructions as the baseline, update it after each change you accept and open each run in Cowork to read the result.

How often can an AI scheduled task check a page?

ChatGPT runs tasks up to once an hour on paid plans. Claude offers hourly, daily, weekly, weekday or manual runs. For supplier terms, guidance and price pages, weekly is enough, and a modest frequency keeps the load on the site you are reading light.

Can it watch a page that needs a login?

No. Keep it to public pages anyone can open. A page behind a login, or one that blocks automated readers, will not work, and a password never belongs in a task. Watching a portal or a system you sign in to is integration work, built with proper access and an audit log.

Is AI page monitoring reliable enough for compliance?

Not on its own. The assistant decides what counts as a difference, so it can miss a small edit, and a page that failed to load can look like a quiet week. Use it as an early warning that puts a change in front of a person sooner, and keep the formal contract or compliance review.

Does my computer need to be on for the watch to run?

Not usually. Claude runs scheduled tasks in the cloud, so your computer can be asleep, unless the task needs local files or apps; on Claude Enterprise, running Cowork in the cloud is off until an owner turns it on. In ChatGPT, a desktop task that uses local files needs the computer on and the app running.

Why not use a website change-detection tool instead?

A tool that compares a page character by character catches every edit, including the ones nobody cares about, such as a new date in the footer. The assistant reads the page more like a person would and quotes the clause that moved, but it can miss one. Where a missed change is expensive, store every version and diff it in code, and use the assistant to make the change readable.