Same visitor. Same URL. Two different pages.
Dana lands on our pricing page twice, three days apart. Nothing about the URL changes, nobody logs in, and no rule was hand-written for her. The only thing that moved between these two renders is what she did on the site in between.
Illustration – authored copy showing the two renders this page could serve, not a capture of a live result. The signals listed are ones the tag genuinely reads in-session; the wording is ours. Nothing here is a customer, and no figure on this page is a measurement.
Sense and act at the tag. Steer in the app, or through an agent.
Sensing and acting happen at the tag. Steering happens in the app, or through an agent.
One loop. Every visitor. Every session.
Four things happen to every visitor, automatically, in the time it takes the page to paint. Then it repeats, getting smarter each cycle.
Capture
The tag observes behavioral signals: scroll velocity, dwell, hover, pauses, tab-switches, source, time of day. No form, no login, no PII. Just what the visitor actually does.
Decide
Client-side scoring sorts the visitor into a segment: live, this pageview, recomputed as behavior shifts. No SQL, no batch refresh. Deterministic: same signals in, same segment out.
Serve
The matched variant renders in the zone: sub-200ms, before paint, zero flicker, frequency-capped. The visitor never sees the swap. They just see a page that fits.
Learn · compounds
A multi-armed bandit measures lift per segment per zone, reallocates toward what's winning, retires what isn't. Every cycle sharpens the next. No manual A/B readout, no stats team.
The loop is the idea. This is the surface.
Four screens, one tag. Every stage of the loop has a real place your team (or its agent) can watch, steer, and prove it: nothing mocked.
Four things happen, in this order.
Nobody believes a no-code claim from a sentence. This is the screen the work happens on, and the sequence you move through on it.
- 01
Put the tag on the page
One script tag, deferred. It goes in your tag manager or your head, and from that moment every session is being scored – before a single variant exists.
your tag manager, or the head - 02
Point at what should change
Click the element you want AXO to be able to swap and it becomes a zone, with a self-healing selector and a declared set of fields. No template edit, no release.
Personalize · Content - 03
Write the variant, or have the agent draft it
Pick the segment, write the copy. Draft with AI writes a first pass offline for you to approve – nothing generative runs on the live page. The preview beside the editor renders the result against your real site.
Personalize · Content · Variants - 04
Publish, and the holdout starts with it
From the first impression, a tenth of visitors keep the standard page. You do not opt into measurement later or run a separate test – the control group is already there.
Publish
Capture of the running app against a demo storefront, the same way the four screens above were taken. The content in it is demo data, not a customer's.
The loop runs itself. You give it range.
The four-stage loop is automatic. These are the things you (or your agent) set up once and let it carry across every channel.
Install
One script tag, deferred. Live in an afternoon, not a six-month implementation.
Triggers
Cart, browse, dormancy, custom events. The same segment that personalizes the page fans out to email, paid social, push, and SMS, triggered by behavior, not a scheduled blast.
Campaigns + content
You write a brief. A drafting agent writes segment-specific copy and generates on-brand imagery under your brand guardrails. Bring your own model and DAM. A human approves before anything ships.
Everything above is a claim. This part is the product.
Every other vendor shows you a picture of their tool working. We are the personalization layer and this site runs our production tag, so the panel below is not an illustration of a decision – it is the decision AXO made about you, read back out of your own browser.
The tag loads deferred, after your analytics, so this takes a moment.
Whichever side of that you are on was decided before this section rendered.
Variants earn traffic. Nobody declares a winner.
Testing tools that wait for significance hold a fixed split for weeks, then wait for someone to read the dashboard, call the winner, and ship it. AXO runs a multi-armed bandit instead: every variant's share of traffic is recomputed each hour from the conversions it has actually driven, so traffic flows continuously toward what converts – and away from what doesn't – for as long as the variants are live.
Earning starts at the first impression
A new variant starts with an equal claim on traffic and earns more as evidence arrives that it converts. There's no test window to schedule and no end date to wait for – allocation is simply always running.
Losers get starved
A variant that isn't converting loses share automatically, until it's getting almost none. Nothing has to be switched off in a meeting – by the time you look, the winner is already carrying most of the traffic.
Steady for each visitor
The draw is deterministic per session, so a visitor keeps the same variant across pageviews instead of watching the page flip-flop mid-visit. The odds move between sessions, not under someone's cursor.
AXO's simulator can exercise a site before real visitors arrive – useful for demos and dry runs, dangerous if fake traffic could steer real allocation. So the optimizer enforces one rule: the moment a real visitor is exposed to a variant, simulated traffic is excluded from the evidence entirely. A simulated session can never outvote a real conversion.
And while a site is still running on simulated evidence, the dashboard says so: a SIMULATED DATA badge sits on the optimizer until real exposures arrive, then clears on its own. Numbers that came from a simulator are labeled as such – you are never left to guess.
One more boundary: the bandit only decides traffic shares. Every variant it allocates between was written, or approved, by a human – allocation is the part that runs on its own; authoring never is.
A human can drive it. An agent can drive it to infinity.
Every operation in AXO (read a segment, draft a variant, rank an experience, fan out to a channel, tune the bandit) is one of 130+ MCP tools. The same actions your team takes by hand, your agent takes programmatically, at a speed and breadth no team could click through.
Explore the MCP server →The human stays in the loop, not out of it. You decide where the agent runs, and where it asks first.
No CDP. No identity graph. No warehouse. No hallucinating model in the decision.
The legacy way is four tools (a CDP, an experimentation tool, an activation tool, and an AI bolt-on) stitched across four contracts and a six-month implementation. AXO is one layer that does the whole loop.
AXOFOUR GUARANTEES THAT HOLD THE WHOLE THING UP
Deterministic
Rules and classical ML on the live path. No generative text deciding what a visitor sees. Same signals in, same experience out.
Identity-optional
Segments work on anonymous behavior. Layer identity in from your CRM only when it pays off.
PII-optional
Behavioral segmenting needs no PII. It enters only the channels that require it (email, CRM, paid audiences) and stays there.
Consent-first
Built-in consent gate, CMP integration, Google Consent Mode v2, GDPR + CCPA tiered. A visitor can drop their segment in one click.
The awkward questions, answered first.
Including the two most vendors leave out: what you get when there is not enough traffic to prove anything, and what happens when the measurement says the personalization did not work.
The arguments behind the architecture.
You don't need to know who they are to know what they want
Why the anonymous 90% doesn’t need an identity graph
Read →The three-stack tax on customer journeys
What the CDP + A/B + automation stack really costs
Read →No hallucinating agent in your decision chain
Why the live decision path stays deterministic
Read →