Which features use an AI model
The features that send data to an AI model, what they send, and on which sites they run.
On this page
Most of Parlox uses no AI model at all: detection, visit outcomes and the sentences describing a visit are rules over recorded facts. Three features do use one. None of them is meant to receive anything about a visitor; the one exception, in control names sent by the browser part up to version 1.0.3, is described below.
At a glance
| Feature | Model provider | What it is sent | Runs on |
|---|---|---|---|
| Detection: naming a visitor | None | Nothing | Every site |
| Visit outcomes and the sentences describing a visit | None | Nothing | Every site |
| The install wizard | None | Nothing | Your machine |
| The control map's suggestions | Anthropic | What an agent sees of a button: its role, label, tag, attributes and the pages it is on | Every verified site with recent visits, whatever its plan, unless AI features are turned off for the site |
| The weekly checks' fact reader | Anthropic | The visible text of your public product pages | Verified sites on plans that include it, unless AI features are turned off for the site |
| The Intent Monitor | OpenAI, Anthropic, Google | Shopping questions, and your public pages if you ask for suggestions | Off unless it is enabled for your site |
The control map's suggestions
Part of the control map. For buttons and links that agents clicked at least three times and that neither confirmed events nor platform markup could name, a Claude model, through Anthropic's API, proposes a role from a fixed list.
- What it is sent: your store's domain and, for at most 40 controls per run, what an agent sees of each one: its accessibility role and label (already masked in the browser), tag, id, classes, link path, a few attributes (
name,type, the form's action), its structural path, and up to three page paths it appears on. Up to version 1.0.3 of the browser part, a control that wraps a larger block (an address card, say) could carry that block's text as its name, at most 80 characters with emails, phone numbers and digit runs masked, and such a name could reach the model. Since October 16, 2026 Parlox drops those names on arrival, and version 1.0.4 no longer sends them. - When it runs: offline, in the background, never on live traffic, on every verified site whose browser part reported a visit in the last day, whatever its plan, unless an owner or admin has turned off AI features in the site's settings. A site that is not verified is never sent.
- What it changes: nothing until you decide. A suggestion below a fixed confidence is discarded; the rest are shown to you as suggestions, and count only once you confirm them.
The weekly checks' fact reader
Parlox checks the pages agents reach once a week, on verified sites. Most checks read your pages' markup and need no model. For facts that markup often lacks (the price when markup has none, availability, delivery, returns, warranty and reviews), a Claude model reads the product page's visible text and answers, for each fact, whether it is stated and the exact sentence that states it.
- What it is sent: the visible text of your public product pages, as Parlox fetched them (no JavaScript run, no cookies, no login). Nothing from visits.
- When it runs: on verified sites, on plans that include it. Without it, those facts show as "could not check", never as "not stated".
- How its answers are treated: a sentence is accepted only if it appears in the page's text, and a fact found this way is shown as inferred, with its sentence beside it.
The Intent Monitor
The Intent Monitor asks AI assistants the shopping questions your customers would ask, and records which stores and products the answers recommend.
- It is off unless it is enabled for your site.
- What it sends: the questions you write or approve, to OpenAI's, Anthropic's and Google's models, with web search. Anthropic's models also write two rewordings of each question and read the answers to list what they recommend.
- Suggestions, if you ask for them: a Claude model reads your site's home page and up to 20 of its product or collection pages, and proposes questions for you to approve. Only for a site whose ownership is verified.
- What it stores is public: assistants' answers about public stores, and the pages they cite.
Turning AI features off
An owner or admin of a site can turn off AI features in the site's settings. From the moment it is saved, Parlox sends nothing from or about that site to any AI model: no control-map suggestions, no fact reader (those facts show as "could not check" with the reason), and no Intent Monitor questions, which wait and are sent only if the switch is turned back on within the cycle. Work already under way when it is switched off finishes: a page the fact reader is already reading, and questions already sent. What models returned before is kept. Each change is recorded on the account, with who made it. Detection, recordings and every report work as before, since none of them uses a model.
What none of them do
- No model receives an address, a recording, a value typed into a field or an order, or anything that tells one visit from another. The one exception is the control names noted above.
- No model decides a label, an outcome or a number you see as a fact: a model's output is always shown as a suggestion or as inferred, beside its evidence.
- The install wizard has no telemetry and no AI: its edits are fixed templates, shown as a diff first.