Building an AI Agent That Watches Websites for Changes

Yes, you can build a working AI agent that watches websites for changes in an afternoon, usually for free. The pattern: a change-detection tool snapshots a page on a schedule, compares it with the last version, and hands the difference to an AI model that says what changed in plain language, and whether it matters. You have three realistic routes in 2026: a hosted watcher with AI built in (Visualping or changedetection.io), a no-code pipeline wiring a watcher to an LLM (Zapier, Make, or n8n), or an AI-native agent (ChatGPT scheduled tasks and OpenAI’s dots).

Prices below are a snapshot from November 7, 2026.

What does an AI agent that watches websites actually do?

It’s two jobs stapled together: a robot with a stopwatch, and a brain that reads its output. Website change detection is the mechanical half: software fetches a page on a schedule, strips noise like ads and timestamps, and compares the result against the previous snapshot. That comparison is the diff, and on its own it’s dumb: it screams about a new cookie banner as loudly as about a price drop.

The AI half fixes that. The diff goes to an LLM (a large language model like GPT or Claude) with a prompt such as “summarize what changed and tell me if it affects the price.” The model tags the change as important or ignorable, and only then does the system notify you by email, Slack, or Telegram: usually through a webhook, a URL one tool calls so another wakes up and runs. Every option below varies by who runs each part.

What are the four ways to build an AI website watcher?

Four, ranked from least to most effort, for most people the first row is the right answer.

Approach	Tools	Typical cost	Maintenance	Reliability
Hosted watcher with built-in AI	Visualping; changedetection.io hosted	Free to $35/mo	Nearly none	High: vendor-run servers
Self-hosted watcher plus your own LLM key	changedetection.io (open source) + OpenAI, Gemini, or Ollama	~$5/mo server + pennies of API	Low after setup	High: you own it
No-code pipeline	RSS.app or a watcher’s webhook → Zapier, Make, or n8n → LLM → Slack/email	Free to $20/mo	Medium (chains break quietly	Medium) more failure points
AI-native agents	ChatGPT scheduled tasks; OpenAI dots	Included in a ChatGPT plan	None	Medium: coarse schedules, no true diffing

The first two rows are the classic experience; the third suits builders who want digests and deduplication; the fourth, people who refuse to open a settings page.

Which hosted watcher gives you the most for your money?

changedetection.io wins on volume per dollar, Visualping on polish, Distill.io on the best free browser-based tier.

changedetection.io’s hosted plan costs $8.99/month for up to 5,000 page watches with re-checks as fast as every 5 minutes, plus 85-plus notification services, EU/US/Tor proxies, and one real Chrome browser for JavaScript-heavy sites, per the official site (November 2026). The AI is bring-your-own-key: in Settings → AI/LLM you plug in a key from OpenAI, Anthropic, Google Gemini, OpenRouter, or Ollama, and the tool writes summaries and filters alerts. Open source too, so self-hosting the identical machinery on a cheap VPS costs nothing but setup time.

Visualping starts free (150 checks monthly across 5 pages at hourly frequency) and Personal 1K is $14/month ($10/month billed annually) for 1,000 checks on 10 pages at 15-minute intervals. The extra money buys polish: AI change summaries (unlimited on every plan), an important-or-not flag on each change, sentence-long AI alert criteria, and an assistant that builds monitors for you. REST API and MCP server access come on all plans, free included.

Distill.io sits between them: the free plan covers 25 monitors (5 cloud) with 1,000 cloud checks at 6-hour intervals; Starter is $15/month for 50 monitors and 10-minute checks; Professional $35/month at 5 minutes; Flexi from $80/month at 2 minutes. Its party trick: browser-extension local monitoring, checking as fast as every 5 seconds on your own machine, off the cloud quota.

changedetection.io’s hosted plan is $8.99/month for 5,000 pages at 5-minute checks. Visualping’s $35/month tier buys 5,000 checks spread across just 20 pages. Same money, wildly different volume: that’s the decision in one comparison.

Visualping’s summaries are the nicest and its free tier is genuinely usable, but at dozens of pages the math isn’t close: changedetection.io is the value pick.

Zapier, Make, or n8n: which glue should you pick?

n8n for cheap volume and built-in AI nodes, Make for mid-priced visual building, Zapier for the widest app coverage: your watcher spots the change, one of these turns it into a workflow.

Zapier prices by tasks: free 100/month, Pro $19.99/month billed annually ($29.99 monthly) for 750 tasks, per its pricing page (November 2026). Two details matter: triggers and polling don’t count as tasks, and Zapier MCP is on every plan: one MCP tool call costs 2 tasks. So a webhook-triggered Zap with an AI step costs a couple of tasks per real change, not per check.

Make switched its billing unit from operations to credits on August 27, 2025, at a 1:1 conversion, per its help center. Free includes 1,000 credits/month; Core is $9/month billed annually with 10,000 credits, extra bundles at $0.90 per 1,000. Most actions cost 1 credit; Router and error handlers are free. Caution: polling burns credits on every check: point your watcher’s webhook at Make and only pay when something changed.

n8n charges per workflow execution, not per step: its pricing page says it plainly: “If it doesn’t run, you don’t pay.” Cloud Starter is 20€/month billed annually for 2,500 executions; Pro is 50€/month for 10,000; the Community Edition is free to self-host with unlimited executions. Paid plans include the AI Agent node and an MCP Server Trigger; templates like RSS monitoring with AI filtering and Slack alerts adapt easily to webhook input. Comfortable with a VPS? Self-hosted n8n plus changedetection.io is the cheapest serious setup.

Can ChatGPT or OpenAI’s dots watch websites without any setup?

Yes, but coarsely: these are agents on a clock, not precision instruments: fine for “check this page each morning,” useless for 5-minute price alerts.

ChatGPT’s scheduled tasks are capped by plan, per OpenAI’s help center (November 2026): 3 active tasks on Free and Go, 5 on Plus, 10 on Business and Edu, 15 on Pro and Enterprise. Free accounts run one-time tasks or recurring ones once per day in flexible windows; paid plans get hourly schedules and exact times. Event-triggered tasks (reacting to Gmail, Slack, or GitHub) run up to 30 times per hour and 720 per day, paid plans only, with no generic “web page changed” trigger.

Dots are OpenAI’s always-on agents, announced September 29, 2026. Each runs on its own cloud computer with a browser, keeps persistent memory, and connects to 4,000-plus apps; your first dot comes included with Pro or Business Premium, and dot conversations don’t count against ChatGPT usage limits. The launch catch is regional: per OpenAI’s help docs, dots roll out to Pro users in markets excluding the EEA, Switzerland, and the UK (Enterprise and Edu get admin-enabled access). Nothing in the docs describes a dots feature for scheduled page-diffing. The always-on framing suggests monitoring could fit, but that is not a stated capability.

Build one yourself: the changedetection.io stack, step by step

The fastest credible build is changedetection.io with its built-in AI tab: one tool, no glue code, about 20 minutes.

1. Choose your hosting. Self-host the open-source version with Docker on a budget VPS, or pay $8.99/month for the official hosted service with 5,000 pages and 5-minute checks.
2. Add your first watch. Paste the target URL and set a check interval, 5 minutes on hosted, your choice self-hosted.
3. Narrow the target. Add a CSS/xPath filter (the price element, the “In stock” button) and ignore-text rules so nav bars and cookie banners never trigger you: this decides whether alerts are useful or noise.
4. Turn on the AI. Settings → AI/LLM: pick OpenAI, Anthropic, Google Gemini, OpenRouter, or Ollama, paste your API key, hit Test connection. It’s bring-your-own-key: the AI bill is separate, usually pennies per month.
5. Write your alert criteria. One sentence: “Summarize what changed and only alert me if price or availability changed.”
6. Pick notifications. There are 85-plus channels: email, Slack, Discord, Telegram, or a raw webhook via Apprise’s custom JSON URLs for chaining into Zapier, Make, or n8n.
7. Test deliberately. Watch a page you control, change it, confirm the AI summary arrives and reads sensibly: misconfigured filters are the likeliest failure.
8. Optionally chain it. Point the webhook at an n8n workflow that logs changes, dedupes repeat alerts, or rolls up a daily digest: the RSS-plus-AI templates adapt with minimal edits.
9. Set a failure path. Make sure you’d notice if checks start erroring: a monitor that silently dies is worse than none, because you’ll trust a signal that no longer exists.

What about anti-bot walls, terms of service, and ethics?

Monitoring public pages is normal and, in the US, not a federal crime, but Cloudflare has made agent traffic harder since mid-2026, and terms of service still bite.

Cloudflare’s July 1, 2026 announcement gave every customer, including free sites, granular switches for Search, Agent, and Training bot categories. From September 15, 2026, new domains block Agent and Training bots by default on pages that display ads, and a “Verified” bot badge no longer guarantees access. Translation: some sites will refuse your watcher outright. Respect it: use the site’s RSS feed, newsletter, or official API instead. Disguising your bot or smashing through challenges stops being monitoring.

The legal reference case is hiQ Labs v. LinkedIn (2017-2022). The Ninth Circuit ruled twice that scraping publicly accessible data doesn’t violate the Computer Fraud and Abuse Act, but in November 2022 a district court found hiQ had breached LinkedIn’s user agreement, and the case settled. Contract law survives where the CFAA doesn’t reach, and logged-in pages are riskier still. This is general information, not legal advice: commercial monitoring deserves a lawyer’s review.

The ethics that cost you nothing: honor robots.txt, keep intervals humane, prefer RSS or APIs where they exist, and don’t monitor behind logins you were given as a human. That is not a statute, but it’s where ToS claims and bans come from.

What are the honest limitations?

The pipeline works, but five failure modes will find you eventually.

• Anti-bot walls. Cloudflare’s default-blocking shift means a slice of the web is off-limits to any watcher.
• Free-tier slowness. Visualping free checks hourly, Distill’s free cloud tier every 6 hours, ChatGPT free tasks once per day: five-minute alerts cost money.
• JavaScript-heavy pages. Client-side-rendered content needs a real browser; changedetection.io hosted includes one Chrome runner, extras are add-ons.
• False positives. A/B tests, rotating promos, and cookie banners all look like changes; filters and ignore-text rules are your defense.
• LLM summaries can be wrong. Compression can misstate what changed: verify anything that costs money against the page itself. No vendor’s docs promise perfect accuracy here.

Last, Visualping states checks don’t roll over month to month: quiet months waste budget.

Which setup should you pick?

Match the stack to your volume and patience.

• One to five competitor pages, zero setup: Visualping free, then Personal 1K at $10/month billed annually.
• Dozens to hundreds of pages on a budget: changedetection.io hosted at $8.99/month, or self-hosted if Docker doesn’t scare you.
• A custom agent, not just alerts: n8n (self-hosted for best price, Cloud Starter at 20€/month otherwise) with your own LLM key; Zapier if your team lives there.
• Already paying for ChatGPT Plus: a scheduled task for morning checks is a fine start; don’t expect 5-minute alerts.
• In the EEA, Switzerland, or the UK: dots aren’t available on personal Pro at launch; hosted watchers work everywhere.

FAQ

Can ChatGPT monitor a website for changes?

Partially. Scheduled tasks check pages on a recurring schedule, but limits are tight: 3 active tasks on Free and Go, 5 on Plus, 15 on Pro or Enterprise (OpenAI’s help docs, November 2026). Free accounts can only recur tasks once per day; paid plans get hourly schedules. There’s no generic “page changed” trigger: event tasks react only to Gmail, Slack, and GitHub.

What is the best AI tool for website change monitoring?

For most people, changedetection.io: $8.99/month hosted for 5,000 pages with 5-minute checks and bring-your-own-key AI, or free self-hosted. Visualping is the most polished. AI change summaries and an important-or-not flag on every plan, starting free with 150 checks monthly. Distill.io has the strongest free tier for browser-based monitoring. Pick by volume, polish, or price.

How does AI website change detection work?

A watcher fetches the page on a schedule, filters to isolate what you care about, and diffs it against the previous snapshot. The diff goes to an LLM with your criteria; the model summarizes the change in plain language and flags it as important or noise, then notifications go out via email, Slack, or webhook. changedetection.io does this in one tool via its AI/LLM tab.

How much does it cost to monitor websites with AI?

You can start free: changedetection.io self-hosted is open source (server costs only), Visualping’s free plan gives 150 checks monthly on 5 pages, and Make’s free tier includes 1,000 credits. Paid entry: changedetection.io hosted at $8.99/month, Visualping Personal 1K at $10/month billed annually, Make Core at $9/month for 10,000 credits, Zapier Pro at $19.99/month for 750 tasks. LLM API usage is typically pennies.

What are the main alternatives to Visualping?

changedetection.io (open source, cheapest at volume, BYOK AI), Distill.io (browser extension with local checks from 5-second intervals, cloud plans from $15/month), PageCrawl.io (plain-language AI summaries), and the RSS route: RSS.app turns any page into a feed (free plan: 2 feeds, 24-hour refresh). ChatGPT tasks and OpenAI’s dots are the zero-infrastructure options.

Which website monitoring option is best for beginners?

Visualping. Paste a URL, select the part of the page you care about, and it emails AI summaries of what changed: no setup beyond an account, and the free plan covers 5 pages at hourly checks. ChatGPT scheduled tasks are simpler still if you already pay for Plus. Avoid self-hosting until you’ve outgrown hosted tools.

What are the main limitations of AI website monitoring?

Five recurring ones: anti-bot systems (Cloudflare has blocked “Agent” bots by default on ad-displaying pages for new domains since September 15, 2026), slow free-tier checks (Visualping hourly, Distill every 6 hours), JavaScript-heavy pages needing a real browser, false positives from A/B tests and dynamic content, and LLM summaries that can misread a diff: verify anything important against the page itself.

Is it legal to monitor or scrape websites?

In the US, hiQ v. LinkedIn (2017-2022) established that scraping publicly accessible data is not a federal crime under the CFAA, but in November 2022 a district court found hiQ had breached LinkedIn’s user agreement, and the case settled. Terms of service remain a contract risk, especially behind logins or at commercial scale. This is general information, not legal advice.

Can AI agents check pages behind a login?

Some can technically. OpenAI’s dots run on their own cloud computer with a browser, and changedetection.io supports logged-in browsing. But logging in almost always means agreeing to terms that prohibit automated access: a contract violation even when scraping public pages is not, and password handling adds risk. Get permission from the site, or use an official API.

What should I set up first?

Start free: a Visualping account for up to 5 pages, or changedetection.io self-hosted if Docker doesn’t scare you. Add one page that genuinely matters, set a sane interval, write one line of AI criteria (“tell me if the price changes”), and let it run a week. If alerts prove useful, upgrade to the $8.99 hosted plan or a Visualping Personal plan.

Sources

• Visualping pricing: plans, checks, AI features, API/MCP access (November 2026)
• changedetection.io: hosted plan price, features, BYOK AI providers
• changedetection.io AI summary tutorial. AI/LLM setup steps
• Distill.io pricing: monitor counts, check quotas, intervals
• Zapier pricing: plans, task tiers, free triggers, MCP task cost
• Zapier MCP. MCP overview
• Make pricing: credit model, Core plan, bundle pricing
• Make Help Center: Credits: operations-to-credits transition
• n8n pricing. Cloud plans, Community Edition, per-execution model
• n8n workflow: smart RSS monitoring with AI filtering and Slack alerts: template for the watcher-LLM pattern
• OpenAI Help Center: Scheduled tasks in ChatGPT: task limits, frequency caps, event triggers
• OpenAI: Introducing dots: dots capabilities and plan inclusion
• OpenAI Help Center: Getting started with your dot: eligible markets and exclusions
• Cloudflare blog: Your site, your rules (new AI traffic options) Search/Agent/Training controls and September 2026 defaults
• RSS.app pricing: feed generation plans and refresh rates
• hiQ Labs v. LinkedIn (Wikipedia) case history and outcomes
• Apprise wiki: custom JSON notifications: webhook payloads from changedetection.io