Sphere Quarterly · Issue 02 · Adaptation
One URL. Every visitor sees a different page.
The Self-Rewriting Site — Sphere Quarterly · Issue 02
One URL. Every visitor sees a different page. The memory belongs to them. A field guide to browser-resident engrams, declared visitor intent, and pages composed at view-time by the team that ships the reference implementation.
- Anton MaciusField CTO
In this article
The static landing page is the last surface on the internet not yet personalized by AI — and it’s the one that decides whether a buyer talks to you. The fix isn’t more templating. It’s pages composed at view-time, grounded in what the visitor declared, signed at the wire, with the memory living in the visitor’s browser.
If you only read one box on this page.
If you wouldn’t send the same handwritten letter to every prospect, why does every visitor see the same landing page? The version a CIO needs is not the version an engineer needs is not the version a curious reader needs — and yet most B2B sites in 2026 still render one version for all three.
The reason is technical. Templating engines — the way Mutiny (opens in new tab), Intellimize (opens in new tab), and Adobe Target (opens in new tab) personalize — swap pre-written blocks. They cannot compose new content because LLMs at view-time were too slow and too expensive until recently. Both have flipped.
Sphere takes a different approach. The visitor declares who they are and what they’re trying to do (two simple choices, no form). The page reshapes around that declaration. Where templating ends, the LLM picks up — composing a one-off briefing from the engram store and your knowledge base, in well under two seconds, at well under ten cents.
And the memory of all of it lives in the visitor’s browser — signed with a keypair only they hold, viewable at /me, wipeable instantly. The site can’t read it from a server. Article 17 erasure is structural, not a promise.
A CIO and a software engineer land on the same product page.
Both see the same hero, the same screenshots, the same CTA. One leaves.
CIO sees compliance, TCO, audit. Engineer sees the API, the embed snippet, the latency. Both convert.
A returning visitor opens your site for the second time.
Either they re-introduce themselves or you stalk them with a cookie.
Their browser remembers, signed by a keypair only they hold. You can’t read it. You don’t need to.
A regulator asks “why did this person see that page?”
Forensic reconstruction across three vendors and a CDP.
The ledger contains the answer: declared intent, engrams used, content selected. Signed at the wire.
Declared intent beats inferred behavior.
Personalization by inferred behavior is brittle. A visitor reads three blog posts about compliance and the system decides they’re a DPO — when really they’re an engineer doing research before a meeting. The inferred persona is wrong, and the personalized hero now wastes both sides’ time.
Declared intent skips the inference. The visitor says “I’m a CIO, I’m evaluating, I care about audit and TCO” — two simple choices, no form. The site reshapes within the same paint: a different hero headline, a different deck, a different bullet list, a different CTA. Sections that don’t apply collapse via a single CSS variable. The whole change is one DOM attribute and a stylesheet doing what stylesheets were always meant to do.
Inference and declaration combine, when the visitor allows it. Sphere’s engram store consolidates declared intent with anonymized on-site behavior, with the visitor in full control of both sources. The next time they visit, the page knows what they wanted last time — from their browser, not from yours.
From declared intent to composed page, in under two seconds.
The bottleneck for view-time generation used to be cost and latency. In 2024 a personalized landing page from a leading model cost $1–3 and took 8–15 seconds — far too expensive and far too slow. In 2026 both numbers are down by an order of magnitude. A composed brief grounded in 5–10 engrams plus a small knowledge base now finishes in well under two seconds and lands under ten cents.
Sphere ships the runtime as /api/public/brief. The site posts a declared-intent + visitor key, the brief composer reads the engrams the visitor has consented to share, the LLM composes against your knowledge base with cost ceilings, and the rendered HTML lands inline. Every step is signed in the audit ledger. The cost ceiling is set per campaign — runaway bills are not possible.
The memory belongs to the visitor.
The portable engram is the load-bearing piece. A visitor’s browser generates an ED25519 keypair on first use via WebCrypto (opens in new tab). The private key never leaves the browser. Memory entries — preferences, declared intent, page visits — live in IndexedDB and are signed with the local key. Even an exfiltrated copy is verifiable as “this browser said this”.
This inverts the cookie model. The site is no longer the party that knows things; the visitor is. The site asks, the visitor’s browser answers under the grants the visitor has set. Erasure is structural — the visitor wipes IndexedDB and there is nothing left to erase elsewhere because the site never had its own copy. Article 17 satisfied without a forget-cascade ticket.
The widget on the right is not a mock-up. It runs in your browser right now. Generate a keypair. Add an entry. Watch it get signed. Open DevTools → Application → IndexedDB → sphere-portable-engram to verify. Wipe it whenever you want.
Same URL. Four visitors. Four pages.
The right comparison is editorial, not technical. A long-form newspaper article gets edited once. The same words appear on the homepage, in the email digest, in the mobile app, in the print edition. Each surface presents the same content shaped for its reader.
A self-rewriting page does the same shaping job, but for one piece of content addressed to many readers. The CIO sees the compliance/TCO frame. The engineer sees the API surface and the embed snippet. The DPO sees the lawful-basis section and the data-flow diagram. The curious reader sees the manifesto first. Same URL, same source content, four projections.
Why this page? On demand, with a signed receipt.
EU AI Act Article 50 (opens in new tab) requires AI systems to disclose to a person when they’re interacting with AI-generated content. For a self-rewriting site that means: every composed page carries a queryable receipt. The visitor — or the regulator, on the visitor’s behalf — can ask “why did this person see this version of the page?” and get a structured answer.
The receipt isn’t a paragraph of legalese. It’s a structured record: declared intent, engrams used (by kind, with consent timestamps), knowledge-base sources, model, cost, ledger root. Disputes about misuse become a query, not an investigation. Compliance is the path of least friction.
Why this, why now: three curves crossed.
The LLM economics flipped. A composed personalized brief that cost $1–3 and took 8–15 seconds in 2024 now costs single-digit cents and finishes in well under two seconds. The bandwidth budget for personalization moved from “templated swaps” to “a paragraph rewritten on the fly” without breaking the unit economics of B2B marketing.
The transparency obligation arrived. EU AI Act (opens in new tab) general-purpose-AI obligations applicable from 2 August 2025; high-risk obligations from 2 August 2026. Article 50 makes AI-composed user-facing experiences a disclosable event. Either you build the receipt by construction or you build it retroactively under deadline pressure.
Third-party cookies died. Chrome Privacy Sandbox (opens in new tab) shipped, Safari and Firefox already gone. The identity layer of the open web is gone. The replacements being marketed — topic cohorts, on-device aggregation — are weaker than what a declared-intent button does in two seconds, with no fingerprint and full visitor consent.
For B2B marketing teams who already pay Mutiny (opens in new tab) or Intellimize (opens in new tab) for templated personalization, the upgrade path is clear: LLM-composed at view-time, audit-signed by default, with the memory in the visitor’s browser. That’s the offer.
Cited
- EU AI Act Reg. (EU) 2024/1689 — EUR-Lex (opens in new tab); Arts. 13, 26, 50
- GDPR Reg. (EU) 2016/679 — Arts. 7 (consent), 16 (rectification), 17 (erasure), 25 (privacy by design)
- WebCrypto API — MDN (opens in new tab)
- Adjacent personalization tools — Mutiny (opens in new tab), Intellimize (opens in new tab), Adobe Target (opens in new tab)
- Cookie sunset — Chrome Privacy Sandbox (opens in new tab).
This piece is honest about the line.
The marketing site you’re reading already runs declared-intent reshaping (intent.ts), generative briefings (/brief/[token]), and the portable engram (/me). The list below separates what runs today from what ships in the next two quarters.
Nine capabilities running now. Five extensions in the next two quarters.
Where we are, where we’re going
- Declared-intent reshape — one declaration, CSS variables on
<html>, sections gate themselves. - Engram store — nine kinds, four maturity stages, contradiction detection.
- Generative briefings —
/brief/[token]ephemeral pages with per-brief cost ceiling. - Portable engram at
/me— ED25519 keypair, IndexedDB, browser-only. - Hash-chained audit ledger — every compose call signed.
- Consent ledger — per-visitor, per-purpose, signed at policy version.
- Variant + conversion ledger — A/B outcome attribution running.
- Cost guardrails — per-campaign daily ceilings and per-brief caps.
- 3-line embed SDK — reshape any external site under the same runtime.
- Visitor-key sync — carry the portable engram across devices the visitor controls.
- A “why this page?” link in every composed brief.
- Engram-aware A/B — variants composed instead of swapped.
- Per-campaign LLM model routing — cheaper model for first-touch, premium for high-intent.
- Auto-generated Article 50 disclosure block in the page footer when AI was used.
More to read
Generative AI integration explained: the patterns, architecture, security controls and rollout steps for adding GenAI to the systems your business already runs.
Custom AI development explained: when it beats off-the-shelf AI, what a build involves, and how in-house and outsourced teams compare on speed and risk.
A shortlist of top AI development companies for 2026, grouped by firm type, with a transparent selection method and advice on choosing the right partner.
Generative AI use cases in financial services by function: what each does, the data it needs, the controls regulators expect, and how to measure value.