How it works
From URL to alert, noise-filtered and AI-summarized
Every check runs through eight stages. Most end early (hash match, noise-only change) so you are not alerted. Only the ones that clear every gate reach the LLM and become a notification.
- 1
You add a URL
One-click templates, the Chrome extension, or the dashboard
Paste any public webpage. Pick a check frequency (daily on free, up to every 15 minutes on Business). Optional: pick a CSS selector, keywords to filter on, or the alert channels you want.
- 2
Baseline snapshot
Captured within 15 minutes of creation
Zikit fetches the page and stores a structured snapshot: visible text, section hashes, HTTP status, response time. This is the reference point every future check compares against.
- 3
Fetch on your schedule
Direct HTTP first, then residential proxy, then headless browser
Every scheduled check runs server-side on Zikit's infrastructure (no browser needed). We auto-escalate: plain HTTP first for speed. If a site blocks, we retry through a residential proxy. If that is blocked too, we render with a headless browser. About 95% of real-world URLs work through this cascade.
- 4
Hash compare
Skip instantly if nothing changed at all
If the content hash matches the previous snapshot, we stop here - no further processing, no alert, no LLM cost. Most checks end at this stage, which is how Zikit stays efficient.
- 5
Noise filter
40+ patterns that strip timestamps and rotating content
Strips timestamps ('2 hours ago', ISO dates), session tokens, CSRF values, rotating testimonial carousels, and view counters. If nothing meaningful is left, we stop and stay silent.
- 6
Semantic similarity gate
Embeddings-based - skip if semantically the same
Even if text literally changed, an embedding comparison checks whether the meaning changed. Paraphrased copy scores above 95% similarity and gets dropped. This is the stage that catches 'we reworded the hero' vs 'we raised prices'.
- 7
AI summary + importance score
GPT 5.4 mini - plain English, not a raw diff
Only the changes that pass every gate reach this stage. A focused prompt extracts the before-and-after delta and returns a one-paragraph summary plus an importance score 1-10 (and a suggested action when relevant).
- 8
Alert routing
Your threshold, your channels
If the importance clears your threshold, alerts go to whichever channels you configured per monitor: email, Slack, Telegram, Discord, or a custom webhook. Below threshold: logged in the dashboard, no ping.
Technical questions
Does the pipeline run in my browser?
No. All checking happens on Zikit's servers on your schedule. You can close your laptop; monitors keep running. The Chrome extension is a control surface, not the runtime.
How does the semantic gate avoid false positives?
We embed the before-and-after text using OpenAI's embedding model and compute cosine similarity. At above 95% similarity we treat the change as noise and skip it. This catches paraphrasing, reordered testimonials, and cosmetic rewrites that a raw diff would flag.
What does 'importance 1-10' actually mean?
A rubric-based score. 1-2 is pure cosmetics. 3-4 is a minor content edit. 5-6 is notable (new paragraph, tier label changed). 7-8 is material (price change, new feature, Terms revision). 9-10 is critical (service deprecation, security notice). You set your own threshold.
How accurate is the AI summary?
The summary is grounded in the actual diff we extracted (we don't pass the whole page to the LLM, just the changed sections). GPT 5.4 mini handles this size reliably; hallucinations are rare because the prompt is constrained and the input is small.
What happens if a page is temporarily down?
A single failure marks the monitor 'unstable' but doesn't alert. Three consecutive failures mark it 'error' and we email you so you know the page stopped working. We don't treat transient 5xx responses as changes.
Paste a URL. See the pipeline work.
Free plan runs the full pipeline on 3 sites with unlimited pages. Upgrade when you need more sites or faster checks.