UXXRAY's deterministic extraction models. Versioned rulesets, stamped on every result.
UXXRAY runs two deterministic extraction models: versioned rulesets, not an LLM. Each reads a funnel a different way and hands back a different artifact. No model assigns a score or a fact by judgment. Every detection is a fixed rule over the measured DOM and computed CSS, and every number is frozen and stamped with the model version that produced it.
Surface funnel x-ray
Deep funnel x-ray
A sample breakdown is published, rendered from a captured walk. The in-product deep scan is in beta; access opens gradually.
See a sample deep breakdown →Roentgen 1.26.1 gives every walked page its own step label and id instead of letting them collapse when a funnel nests under one directory. Nothing about how pages are judged has changed, so any score that moves does so only because the page could be measured again.
Full Roentgen changelog ↓Sievert 1.1 closes two ways the deep walk could report a false success: an automated form fill is now confirmed by reading it back from the page, and a click is checked against what's actually on top at that screen position so a pop-up covering a button can't be recorded as a successful click. Before this, a walk that silently failed partway through could be published as if it had hit a dead end in the site's own funnel.
Full Sievert changelog ↓Roentgen reads the funnel a visitor walks by URL: lander to product to checkout. Sievert reads the funnel that hides behind interaction, a quiz or onboarding flow that a URL crawler sees only the first screen of. Sievert answers each screen with a deterministic first-valid answer policy, walking as one disclosed crawler identity, so the paywall it reaches is the one personalized to that exact answer path. Its output includes a machine-readable funnel spec plus an agent brief, assembled only from measured facts.
| Version | Date | Change |
|---|---|---|
| Roentgen 1.27.0 | 2026-08-29 | Roentgen 1.27.0 lets a content or editorial funnel be credited for credibility it visibly has. Logo walls and adoption counts were only ever looked for on SaaS, lead-gen and affiliate funnels, so a content site showing "trusted by 500 companies" got no mention of it at all, rather than a clear "we checked and found none". Shopping-cart tactics like discounts and free shipping stay ecommerce-only, and nothing already published moves: every result keeps the ruleset that produced it, and only a fresh crawl uses the new one. |
| Roentgen 1.26.1 | 2026-08-24 | Roentgen 1.26.1 fixes how a walked funnel labels its own steps. Several pages sharing a directory, like three competitor pricing pages under the same section, used to display the identical label, so a report or a listing could not tell them apart. Each step now gets a label built from what actually makes it different from its siblings. |
| Roentgen 1.26.0 | 2026-08-21 | Roentgen 1.26.0 records which section each thing it finds is actually in. It has always noted where a rating, a guarantee or a struck-through price sits on the page, and separately where each section sits, but it never connected the two - so a report could say a funnel uses social proof without being able to say whether the hero has any. Where a page has nothing detected on it the sections stay silent rather than reporting an empty result, because "we did not measure this" and "we measured it and there is nothing" are different answers and only one of them is useful. |
| Roentgen 1.25.1 | 2026-08-15 | Roentgen 1.25.1 restores scanning after a fault that stopped every scan from completing for about a day. The previous version introduced a coding error in the step that records a screenshot's real pixel size, and because the fault only appeared while a scan was actually running, none of the usual checks caught it. Nothing about how pages are judged has changed, so any score that moves does so only because the page could be measured again. |
| Roentgen 1.25.0 | 2026-08-14 | Roentgen 1.25.0 fixes how the crawler measures a page that scrolls an inner panel instead of scrolling the whole page, a layout common to app-like marketing sites. A section near the top of such a page no longer gets reported as sitting a third of the way down, and a section further down that the screenshot genuinely could not capture now reports no position at all instead of an invented one. Ordinary pages that scroll normally are entirely unaffected. |
| Roentgen 1.25.0 | 2026-08-14 | Roentgen 1.25.0 stops calling a section "hero" just because it happens to be first in a detected group - it now also has to sit near the top of the page. Ten published sections were labelled hero while starting 57-97% of the way down their page, which is not a hero by any reading. A section that no longer qualifies is described as whatever it actually is, most often a CTA band, so nothing disappears from the report. |
| Roentgen 1.24.0 | 2026-08-12 | Roentgen 1.24.0 describes every section it crops: a hero now carries its heading, its word and character count, the actual text on its buttons, and whatever the publisher declares about it in schema.org. It is strictly descriptive - counts and quoted text, never a score or a quality judgement. Existing rows keep their sections undescribed until re-crawled, because absent is honest and invented is not. |
| Roentgen 1.23.1 | 2026-08-11 | Roentgen 1.23.1 fixes a report that could not tell two opposite things apart. When a site's menu only jumps around within a single page, the walk correctly finds no other page to visit, but the report said zero links were seen, which reads as though the page failed to load. The report now says how many links were actually on the page and where they went, so a healthy one-page site no longer looks like a broken capture. |
| Roentgen 1.23.0 | 2026-08-11 | Roentgen 1.23.0 fixes a rule that could mistake an entire site for a translation of itself. Sites that live under a language folder like /en/ link to their own homepage from the logo, and that one link was read as proof the folder was a translated copy, so every link was discarded and the funnel was published as a single page; a site's own root no longer counts as that proof, while genuinely translated copies of a page are still skipped. It also fixes the test tool meant to catch this, which was browsing as anonymous Chrome rather than as our clearly-identified crawler and so was measuring pages we never actually visit. |
| Roentgen 1.22.0 | 2026-08-10 | Roentgen 1.22.0 stops publishing corporate pages like investor relations or careers as if they were the funnel, and lets the walk take a second look one click past the homepage when the homepage itself doesn't fill out the step budget. If you're reading a report for a large or enterprise site, the score is now more likely to reflect the real purchase path instead of a nav link that happened to score higher. This only applies to pages re-crawled after 2026-08-10; existing reports keep the score from when they were captured. |
| Roentgen 1.21.0 | 2026-08-10 | Roentgen 1.21.0 stops guessing a business model from a few dollar figures on the page, and it can now say a page has no funnel at all instead of scoring one anyway. A page needs real pricing structure, a pricing table, a dedicated pricing URL, or billing language, before it's classified as SaaS, so a page full of financial figures with no actual pricing can no longer score against free-trial and tier tactics it never had. If the walk finds no cart, checkout, pricing, signup, trial, demo, or lead form anywhere on the site, the report now reads 'no funnel found' instead of a number: a claim about what the walk saw, not a verdict on the business. |
| Roentgen 1.20.1 | 2026-08-05 | Roentgen 1.20.1 changes no scores. It fixes a gap where only the first page of a walk had its layout data saved, so a funnel's pricing or signup page contributed nothing to the design library even though it was fully measured. Existing crawls are backfilled from data already captured; nothing is re-crawled. |
| Roentgen 1.20 | 2026-08-04 | Roentgen 1.20 can now confirm an item actually reached the cart, not just that a button was clicked. Most stores open the cart in a pop-up drawer with no URL to follow, so before this only 14% of ecommerce funnels reached a cart or checkout step at all. The walk now reads the store's own cart data before and after the click and only counts the step when the quantity increases and the added item matches what was clicked; a cart icon that only animates on click doesn't count. |
| Roentgen 1.19 | 2026-08-04 | Roentgen 1.19 fixes two bugs that could strand a walk at a single page. A site served entirely under a language prefix like /en/ was treated as one giant translated variant and lost every link, including cart and signup; the walk now only treats a prefix as a translation if the same path also exists without it. /register also moves from a skipped page to a conversion step, since some sites use it as the sign-up flow itself, and affected sites went from a 1-page capture to 2 or 3 pages deep. |
| Roentgen 1.18.1 | 2026-08-04 | Roentgen 1.18.1 changes no scores. When the walk can't find a funnel, it now records why: a rejected locale variant, an excluded page type, or simply no matching link, instead of an unexplained miss. If a report reads 'no funnel found,' this is what lets a future fix target the actual cause on that site. |
| Roentgen 1.18 | 2026-08-03 | Roentgen 1.18 stops the walk from treating a blog, docs, or resource page as a funnel step. A multi-step walk now ranks above a single-page one, so a 3-step trip through a resources section could otherwise outrank a real homepage funnel and get published as the conversion path. Checked against the full corpus at the time: zero published funnels were affected, so this closes a risk rather than a live defect. |
| Roentgen 1.17 | 2026-08-03 | Roentgen 1.17 fixes one bug that inflated funnel depth and one that made whole categories of site unwalkable. A translated page, like a Korean copy of a reviews page, was being counted as a real next step, making a 2-page funnel look like 3; that's now excluded unless the segment isn't a real language code. Separately, the walk could never follow a link deeper into its own section, which had capped affiliate and category-review sites at a single page; some of those sites now reach 5 steps instead of 1. |
| Roentgen 1.16 | 2026-08-03 | Roentgen 1.16 adds four accessibility checks: a missing page title, a missing skip-to-content link, a button label that doesn't match its visible text, and a heading level that jumps, like an H2 straight to an H4. Each one only reports the failing side, since confirming a pass would require a judgment call the crawler can't make deterministically. These are display-only and don't affect the conversion score. |
| Roentgen 1.15 | 2026-08-02 | Roentgen 1.15 tightens two section labels that were producing confident false positives. A product carousel with toggle swatches was being labeled an FAQ, and a photo with a few overlay icons was being labeled a logo wall; both now require the actual shape of that section, question-formatted rows for an FAQ, a row of small, similarly sized, distinct images for a logo wall. Existing reports need a re-crawl to pick up the correction. |
| Roentgen 1.14 | 2026-08-02 | Roentgen 1.14 fixes mobile screenshots so they render sharp. The live capture path wasn't setting a pixel density for mobile, so a phone screenshot was stored at 390px wide and looked soft on a real phone display; desktop captures were already sharp and are unchanged. No score or extraction logic changed; existing mobile images stay soft until their page is re-crawled. |
| Roentgen 1.13 | 2026-08-01 | Roentgen 1.13 adds three UX check groups: page weight and render-blocking resources, landmark regions and keyboard reachability, and form input correctness. Page weight is reported as a floor, at least X MB, since some resources can't be measured exactly, and it's never presented as a full total. These are display-only checks and don't move the conversion score. |
| Roentgen 1.12 | 2026-08-01 | Roentgen 1.12 fixes three ways the walk could miss the actual funnel. Hyphenated conversion paths like /instant-quote weren't matching at all; B2B paths like /talk-to-sales weren't recognized as conversion steps, so consultative sales funnels scored as if they were one page; and /contact-sales was being excluded as a generic contact page instead of read as the conversion step it is. The walk also now records why a page's discovery came up empty, instead of leaving that as an unexplained single-page result. |
| Roentgen 1.11 | 2026-08-01 | Roentgen 1.11 fixes a two-week gap where the live crawler and the local tool could disagree on a score. A rule requiring at least 3 distinct logos for a trusted-by section, added to stop a grid of repeated icons from counting as one, had only been applied to one of the two engines. Both now apply it the same way; a report with a logo wall counted from repeated icons will lose it on re-crawl. |
| Roentgen 1.10 | 2026-07-29 | Roentgen 1.10 now records the exact on-page location of individual conversion elements, like a strikethrough price, a rating, or a countdown timer, alongside the full-page screenshot. This is what lets a report crop and show you the precise pixels for each tactic instead of just naming it. A candidate sitting over a photo or background image is dropped rather than guessed, so a crop only ever appears when it can be shown cleanly. |
| Roentgen 1.9 | 2026-07-28 | Roentgen 1.9 clears cookie, consent, and newsletter pop-ups, including inside iframes, before taking a screenshot. Before this, a report captured mid-crawl could show a frozen pop-up instead of the actual page. It only dismisses overlays on page load, never after a click inside the funnel. |
| Roentgen 1.8 | 2026-07-24 | Roentgen 1.8 recognizes a price written with a trailing currency code such as USD, not only a leading dollar sign. A page that only wrote prices that way was previously read as having no price at all. |
| Roentgen 1.7 | 2026-07-23 | Roentgen 1.7 adds two tactic detectors: stock-level scarcity messaging, like 'Only 3 left', and press or media logos. Both are display-only and don't change the conversion score. |
| Roentgen 1.6 | 2026-07-21 | Roentgen 1.6 adds multi-funnel discovery: a site with more than one product line can have each one walked as its own separate funnel instead of only the first one found. This is flag-gated and not on for every crawl. |
| Roentgen 1.5 | 2026-07-20 | Roentgen 1.5 lets the walk follow a review or affiliate funnel that starts one click past an editorial homepage, instead of stopping at the article itself. That means an editorial or review site's actual funnel can show up in a report, instead of the walk stopping at the homepage article. |
| Roentgen 1.4 | 2026-07-17 | Roentgen 1.4 closes detection gaps in three tactic types: comparison tables written in prose rather than a grid, proof that a stated engagement metric is real, and links to docs or a live demo. A report can now credit these where earlier versions missed them entirely. |
| Roentgen 1.3 | 2026-07-17 | Roentgen 1.3 recognizes that a site can have more than one distinct funnel. Different answer paths through the same domain now coexist as separate funnels instead of overwriting each other, and the deepest one is published as the primary. |
| Roentgen 1.2 | 2026-07-17 | Roentgen 1.2 tightens scoring precision across guarantee wording, store-URL weighting, free-trial detection, and per-unit pricing. These are refinements to existing checks, not new tactics. |
| Roentgen 1.1 | 2026-07-16 | Roentgen 1.1 adds a DOM snapshot and layout skeleton for every step of the walk, not just the first, plus video and embed detection and broader SaaS proof-language matching. A report can now show the layout of every step in a funnel, not only the entry page. |
| Roentgen 1.0 | 2026-07-15 | Roentgen 1.0 is the baseline model. It walks a funnel, detects conversion tactics, scores against a rubric that adapts to the funnel type, checks UX signals, and crops the sections it finds. |
| Version | Date | Change |
|---|---|---|
| Sievert 1.1 | 2026-08-01 | Sievert 1.1 closes two ways the deep walk could report a false success. A filled form field is now read back from the page and only counted as entered if it actually matches what was typed, since some inputs silently revert an automated fill; and a click is now checked against what's actually on top at that screen position, so a cookie banner or pop-up sitting over a button can no longer be recorded as a successful click. Before this, a walk that silently failed partway through could be published as if it had hit a dead end in the site's own funnel, when the failure was ours. |
| Sievert 1.0 | 2026-07-24 | Sievert 1.0 is the first deep model. It drives a multi-step quiz or onboarding funnel all the way to its paywall, picking the first valid answer at each screen, and extracts the pricing structure it finds: plan cards, any pre-checked default, discounts, ratings, and bonuses, then renders a screen-by-screen breakdown plus a machine-readable funnel spec. |
Scores and facts are a measurement of a page at crawl time by a specific model version. They do not silently change: they are frozen until the funnel is re-scanned. When a model's detection improves, we bump its version, add a changelog line here, and note it in the engineering log. So any number is always traceable to the exact ruleset that produced it. For the full measurement method, see Methodology.
Two more model names are reserved for future work: Gray and Curie.