Example prompt
A new RSS item
"Notify me when a new item appears in the feed at example.com/feed."
- Bitmovin
- HBO Nordic
- Obsidian
- Laravel
- Fathom Analytics
- PHP 8
- Stanford University
- Takeaway.com
- IGN
- VRT NWS
- spatie
Up is not the same as working
Your site can be up while the form is broken.
An uptime check asks the server for a page and grades the reply. An AI check opens the page the way your visitor would, works through what you wrote, and tells you what it found.
Same page, two verdicts
Illustrative comparison
Uptime check
Status
200 OK
Response
272 ms
Sees
Status, headers and body. Nothing clicked.
Verdict
Up
AI check
Session
Opened the page in a real browser
Found
Login form missing after a plugin update
Console
1 JavaScript error logged
Verdict
Failed, with an explanation
Uptime checks can also look for a string on the page. That is still one request, not a browser session. Explore uptime monitoring.
How the check runs
Write the expectation. The AI runs the session.
Add an AI monitor and enter a URL. Oh Dear looks at the page and suggests a prompt. Edit it or write your own. There are no selectors to maintain: the prompt says what should happen, not which element to click.
An example prompt
Specific beats clever.
On example.com/contact, fill in the contact form with dummy values. If the submission seems unsuccessful, mark this check as failed and send me a notification. Also take a screenshot and show it in the check results.
A good prompt names the URL, what to look for, when to notify you, and what belongs in the result. The more specific it is, the more consistent the results are from one run to the next.
What the AI can use
A real browser, plus a few tools.
- Click and type. Fill in forms, select options, submit, and move across pages. JavaScript runs.
- Console and network. Read console messages and the requests the page made.
- Screenshot. Added to the result when your prompt asks for one.
- Fetch. Request a URL with GET, POST, PATCH or PUT, with custom headers and a payload.
- DNS lookup. Check records like MX, A, CNAME and TXT.
- Date and memory. Know the current date, and compare against the previous result.
What it checks
- Whether a flow works the way your prompt describes: a form submits, a link is there, a price shows.
- Console errors and failed requests during the browser session.
- API responses, DNS records, robots.txt and sitemap content.
- What changed since the previous run, like a new item in an RSS feed.
What it does not
- Whether an email arrived or an order was stored. The AI reads the screen, not your inbox or database.
- Speed. Use performance monitoring for response times.
- A whole site. A run fetches up to five pages and takes up to 50 browser actions.
- Your release test suite. This is a smoke test on a live site.
Each run stops after five minutes at most. An Error result means the run itself couldn't finish, for example a timeout at your AI provider. That is not the same as Failed, and it doesn't send a notification.
It remembers the last run
Get told when something changes.
Every run gets the AI's previous result. So a prompt can ask what changed, not just whether something works right now.
Example prompt
A price drop
"Alert me when the price on this product page drops more than 25%."
Example prompt
A new job listing
"Tell me when a new job listing appears on example.com/careers."
Built for a client roster
One prompt per client. No test code per client.
A Playwright suite for a brochure site is a second job nobody bills for. An AI monitor is one paragraph of plain English per client flow, running on a schedule.
After every deploy
Stop clicking through client sites by hand.
Describe each flow once. The check runs between deploys and tells you which client flow broke, and why, before the client calls.
- The contact form on client-a.example submits with dummy values.
- The homepage of shop.example shows a login link.
- No errors on the developer console of docs.example.
Oh Dear really helps me to focus on developing PHPMap! I don't need to check my site manually. Oh Dear does it for me.
Florian Wartner, Creator of PHPMap.co
Florian's quote is about Oh Dear in general. An AI check takes the same idea into the forms and flows you used to click through yourself.
You control the spend
Critical checks run often. Costly checks run rarely.
Every run uses tokens, so the schedule is a budget decision. Run each check anywhere from every minute to once every 60 days, and match how often it runs to what's at stake.
Every result shows the tokens it used.
- First five runs free. Every account starts with five runs on Oh Dear's own AI account. Enough to set up a check and read the reports.
- Your own provider key. After that, add a key from OpenAI, Anthropic, Gemini or xAI. Your provider bills you for the usage.
- Latest or pinned. Follow the newest model automatically (the default), or pin one version if your prompts are tuned to it.
- Run history. See the result, duration and token count of every run.
Example schedule
Check
Runs
Checkout form
Every 15 min
Homepage content
Every day
MX records
Every 7 days
A new AI monitor runs once a day until you change it.
Operational ownership
Your prompt decides when to wake you.
Write "alert me if the form doesn't submit" and that's the rule. A run that fails without a notify instruction still shows as Failed in Oh Dear, but stays quiet.
Wait for up to 20 failed runs in a row before alerting, if you'd rather ride out a blip. Repeated OK notifications are held back by default, so you hear about the recovery, not every pass.
Slack
Email
Webhook
No feature tiers
Every check.
Every plan.
AI-powered monitoring is on every plan, and each AI monitor counts as one monitor. Model usage after your five free runs is billed by your own AI provider, not by Oh Dear.
Start with everything on.
Try AI-powered monitoring alongside uptime, DNS, performance and the rest of the website-health toolkit.
- AI-powered monitoring on every plan
- Every monitoring feature included
- Plans scale by monitored site count
- No credit card required
10-day free trial. No credit card.
Before the first check
Questions about AI-powered monitoring.
What the AI can check, what it costs, and where a smoke test ends.
What is AI-powered monitoring?
It's a separate monitor type. You describe what should work in plain English, like "the contact form submits" or "the homepage lists at least five products and a login link". An AI then opens your site in a real browser, clicks, types and reads the page, and reports OK or Failed with an explanation.
How is this different from uptime monitoring?
An uptime check requests a URL and checks the response, optionally looking for a string on the page. It never clicks anything. An AI check starts a browser session and works through what you wrote, step by step. It catches the broken form on a page that still returns a 200.
What can the AI do during a check?
Click, type, fill in forms, select options, and move across pages, with JavaScript fully executed. It can read console messages and network requests, take a screenshot, fetch a URL with GET, POST, PATCH or PUT and custom headers, look up DNS records, use the current date and time, and compare against its previous result.
Will it prove my checkout or payment flow works?
No. Submitting a form with dummy values is a smoke test. The AI can read the confirmation on screen, but it can't confirm that money moved, an order was stored, or an email arrived. Use it to catch broken flows early, not to certify transactions.
What are the limits of an AI check?
Each run can fetch up to five pages with the simple fetch tool and perform up to 50 browser actions, and it stops after five minutes at most. Prompts are 20 to 1,000 characters. AI checks verify function and content, not speed: use performance monitoring for that.
How often can an AI check run?
Anywhere from every minute to once every 60 days. A new AI monitor runs once a day until you change it. Every run uses tokens, so match the schedule to the stakes: a business-critical form often, quiet content rarely.
Which AI models can I use, and what does it cost?
Your first five runs use Oh Dear's own AI account, for free. After that, add your own API key from OpenAI, Anthropic, Gemini or xAI in the AI configuration screen, and your provider bills you for the usage. Pick "Latest" to move to new models automatically (the default), or pin one specific version. Every result shows the tokens it used.
When do I get a notification?
When your prompt asks for one. Write "alert me if the form doesn't submit" and the AI sends a notification when that happens, with its explanation. You can wait for up to 20 failed runs in a row before alerting, and repeated OK notifications are held back by default.
What do OK, Failed, and Error mean?
OK means the page matched your prompt. Failed means it didn't, and the report explains what went wrong. Error means the run itself couldn't finish, for example a timeout or an invalid key at your AI provider, so the AI never got to check your site. An Error doesn't send a notification.
Can it notice changes over time?
Yes. Every run sees the AI's previous result, so a prompt can ask about change: a new item in an RSS feed, a price that dropped, a new job listing on a careers page.
Can it replace my Playwright or Selenium suite?
No. A test suite asserts exact behavior in a controlled pipeline. An AI check is a plain-English smoke test on a live site. Use AI checks to catch breaks between deploys, and keep your suite for release gates.
Can I read the results through the API?
Yes. The AI-powered monitoring API returns the result, the prompt, the full response text, the tools the AI used and the token counts for each run.
Who should use AI-powered monitoring?
Anyone who clicks through a site after every deploy: agencies watching client forms and logins, SaaS teams checking signup flows, shops watching product and search pages, and content sites watching articles and newsletter signups.
Get started
Hear about the broken form before the client does.
Describe what should work, and let the AI check it on a schedule. Every feature on. No credit card.
