What is collected
Exactly what the browser part and the server part send, and what they never read.
On this page
Everything below is what the code does today. The browser part and the server part are open source, so each line can be checked against github.com/parlox-io/parlox-js.
The browser part
For each page, it sends:
- A page view the moment the page loads, with the page's HTTP status (where the browser reports it) and the signals used to tell an automated browser from a person: automation flags, the browser's brand list, the graphics renderer, the number of touch points, and the notification permission's state.
- An engagement summary at 8 seconds, sent again when the tab is hidden or closed and on a route change. Numbers only: time on the page; counts of mouse moves, scrolls, clicks, keys and touches; the time to the first input; the pointer path's length, straightness and speed; and a count of script errors on the page (never their messages).
- The page address, as origin plus path. Path segments shaped like tokens (in checkout, order, account-reset and activation links) are replaced with
:token. From the query string, only campaign parameters are kept:utm_source,utm_medium,utm_campaign,utm_content,utm_term,ref,gclid,fbclid,srsltid. - The referrer: another site's as its origin only; your own site's with the same token rule.
- Form friction, per field: its name, type and
autocompleteattribute, how often it was entered and changed, the number of keystrokes, validation errors and the time it was focused. Never its value. - Clicks: the control's label and its accessibility role and name (an
aria-labelortitlewins when present), the tag, id, up to three classes, link path and a short structural path; itsnameandtypeattributes, its form's action path and anydata-actionattribute; and whether the click had no effect or was a rage click. Text is read only from real buttons, submit inputs and simple links, with emails, phone numbers and digit runs masked. Up to version 1.0.3, a click on a link or button that wraps a larger block (an address card, say) also sent that block's text as the control's name, at most 80 characters with emails, phone numbers and digit runs masked; since October 16, 2026 Parlox drops it on arrival, and version 1.0.4 no longer sends it. - Which CAPTCHA or payment form appeared: reCAPTCHA, hCaptcha, Turnstile or Arkose; Stripe, PayPal, Adyen, Braintree, Shopify, Klarna or a plain card form.
- Form submissions: the form's id or name and its action path.
- Error messages shown after a click or submit: that one appeared, and where. Never its text.
- Commerce events, with only
item_id,value,currency,items,order_idandstep(Cart and checkout events).
For a browser it judges automated, it also sends a masked recording. People are not recorded, with one edge case: a person on a virtual machine or a remote desktop can be judged automated, and their visit is then recorded, masked in the same way. A browser with graphics acceleration turned off may be judged automated too (inferred from how the judgement works, not tested).
In the browser it stores, in the tab's sessionStorage only: the visit's random id, its last activity time and its page count (_plx, 30 minutes idle expiry), and a short outbox of beacons not yet answered (_plx_q, at most 30, for 10 minutes). It sets no cookies, reads none (your analytics tools' included), and uses no other storage.
The server part
For each page request that looks automated, one report: the path, method, status, user agent, the Referer, Accept, Accept-Language and Sec-Fetch-Mode headers, the Host and any Web Bot Auth signature headers, and a one-way hash of the client address and the day.
The client address itself is sent only when the user agent names a bot, so it can be checked against the vendor's published ranges. Requests that look like a person's browser are never sent at all: they are filtered on your server.
What is kept
- People's visits: the pages in order, the referrer's host and the campaign parameters they arrived with, interaction counts, commerce events and the coarse device signals above. No user agent, no address, no cookie, no identifier beyond the tab's random visit id. They are kept so agents can be compared with people, not to identify anyone.
- Automated visits: each request or page, with its time, path, user agent, referrer host, category, the basis of its label and the observations behind it.
- Daily page counts, per page and kind of visitor, with nothing about any single visit.
Never collected
- No IP addresses stored. Parlox never writes an IP address to its database. The beacon's address, and the address the server part forwards for a declared bot, are compared with published ranges inside the request and dropped; only the verdict is kept. The rate limiter holds addresses in memory only.
- No cookies, and no identifier that lasts beyond one tab or is shared across sites.
- No values typed into fields and no keystroke content. Emails, phone numbers and runs of four or more digits are masked before anything leaves the page (Masking).
- No page text, beyond the labels and accessible names of the controls clicked (see Clicks above).
- No screen images. The visit's player is rebuilt from events, and an agent's recording is a masked copy of the page's structure: every input a visitor can type into is masked, always. Hidden form fields and link, image and form addresses are rewritten by the address rules, and the page address in a recording keeps only its path and campaign parameters (on arrival since October 16, 2026; in the recorder from 1.0.4).
Your privacy notice
The browser part reads device characteristics (the graphics renderer, the browser's brand list, touch support) to tell agents from people. Depending on where your customers are, that may need a mention in your privacy notice. Whether it does is for whoever handles privacy for your store to decide; Parlox supports consent mode either way.