Rawuser · Privacy-zero personalization

The page that readswhy they came.

Rawuser serves each visitor the version of your page written for their intent. Never their identity. One snippet, thirty versions, zero cookies.

This visit · liveclassifying…
Signals in
reading…
Vector
no strong signal · you get the general version
Variant outserved default × direct·<50ms at the edge·kv read·0 llm calls in your path
View this page as
Preview the page as a different visitor type
1.3KBsnippet, gzipped
<50msedge selection, p95
$0.4530 versions, one page
0cookies · fingerprints · profiles
taskpilot-landing-page.lovable.app
TaskPilot landing page as built: headline 'Get your week back, every week' with a Start free trial button

Direct visitor · sees the page as built

…?utm_campaign=pricing-plans
The same TaskPilot page personalized for pricing intent: headline 'Transparent pricing. No surprises.' with a View plans button

Paid pricing-intent visitor · same page, rewritten

Same URL. Two visitors.

One snippet did this. No code changes, no A/B setup, no cookies. A real site built with Lovable, running Rawuser in production right now.

Crawlers and AI bots always receive the original page. Personalization is for people, so search and AI answers stay canonical.

The mechanism, not a black box

How it works.

01 · Signals

Signals, not surveillance.

UTM, referrer host, entry path, language, country, device class, AI-source. Ambient data the visit already discloses. Never cookies, never fingerprints, never a stored ID. The classifier on the right is the same code the edge runs, executing in your browser right now. Type into it.

LIVE · runs in your browser, no network call
context.offhours_context
100%
source.community
85%
intent.competitor
75%
context.technical_depth
65%
intent.problem
35%
intent.evaluation
0%
intent.feature
0%
intent.brand
0%
dominant: intent.competitor + source.community

02 · Vector

Every visit becomes arithmetic.

A 20-dimension intent vector across intent, source and context. Several dimensions can be high at once; the source never overrides the intent. No segment guessing, no black-box scores. You can read every number in the dashboard.

Animated visit ledger: signal chips appear, vector bars grow, and the served variant line resolves

03 · Variant

The version they needed, instantly.

The edge snaps to the dominant pair and reads a pre-written version from cache in under 50ms. The snippet swaps only the marked elements. Generation happened offline, before they arrived; there is no LLM in the visitor path, no flicker, no layout shift.

snappricing × paid
kv_readv1:variant:…:pricing
latency42ms
swap3 elements · CLS 0.000
llm_calls (visitor path)0

The operator side

Run it from one screen.

Paste a URL. AI finds the headline, buttons and sections worth personalizing; sites built with Lovable or v0 get a one-click rendered-page capture. You review every selector before anything ships.

dashboard / connect
Rawuser dashboard connect screen showing a detected element manifest for reviewRawuser dashboard versions screen with one page version per audienceRawuser dashboard analytics screen with cohort comparison and conversion lift

Compared honestly, not trashed

How Rawuser compares.

vs Mutiny

Identity-first personalization

Built on cookie + IP reverse lookup to identify who's visiting. Hit by GDPR + ePrivacy enforcement; Mutiny pivoted away from website personalization in 2025. Identity-based personalization is a regulatory liability now.

How Rawuser differs →
vs ClearBit / 6sense

Identity-first enrichment

Anonymous-visitor reverse lookup is expensive per visit, frequently wrong, and depends on third-party data brokers. Same regulatory ceiling, same identity-first architecture, same path.

How Rawuser differs →
vs DIY A/B tools

Manual variant authoring

Hand-instrumenting variants in Optimizely / VWO / homegrown tools doesn't scale past two or three. The vector approach generates 30 versions per page at about $0.45, and you author the brief once.

How Rawuser differs →

Built on two moats

Two things that aren’t copyable.

I

Cookies are not disabled here. They are never set.

Nothing outlives the tab: two labels and a random visit marker, gone when it closes. No cookie, no profile, nothing that follows anyone. That is an architecture, not a policy, and it is why GDPR and ePrivacy stop being a project.

What the visit tells us

utmreferrer hostlanguagecountrydevice classentry path

What we never touch

cookiescanvasuserAgent hashIP lookupemailcross-session id

II

Design-grade output

The strongest model where it matters.

Brand, problem, and competitor versions get Claude Sonnet 4.6. The other three intents get Haiku 4.5. That split is a documented decision, not a coin flip.

A design system per surface, never a template.

Default AI-tool aesthetics never ship. This page, the dashboard, and every generated version run deliberate, documented systems. If it looks like a template, it doesn't merge.

Structure changes, not just sentences.

Different intents get different section compositions. A competitor visitor gets a comparison table first; a problem visitor gets validation before any product mention. The structure is the personalization.

Agshin Rajabov

Who’s behind this

Pilots are founder-run. When you book one, you’re talking to me.

Agshin Rajabov: sixteen years in UX, formerly at Electronic Arts and Deutsche Bank, teacher to 400+ designers at Emotix School. Design-grade output isn’t a marketing line here; it’s the personal bar every generated version has to clear.

16 yrs in UXex-Electronic Artsex-Deutsche Bank400+ designers taughtLinkedIn →

Start your pilot

See it on your site.

We’ll generate the versions we’d ship for your exact ICP, on your real page. About $0.45 in API cost. Two-day turnaround.