AICon USA · June 11, 2026 · AI Real-World in Practice
AICon USA · June 11, 2026 · Hello, personalization track 👋
AICon USA · June 11, 2026 · For the people who will build it
Personalization…
is the next frontier of AI.
is AI's next frontier — and its next moat.
is AI's next frontier — the layer above the model.
From intelligence to relevance. From models to memory. From everyone to you.
Why the next product moat isn't bigger models — it's the surface that adapts to each user.
Why the differentiation moves up the stack — into the context layer that wraps every AI client.
Jason Arbon
Founder & CEO, Testers.AI
Previously:
Search personalization @ Google
· Relevance measurement @ Bing
· Founder, Test.ai (Google-funded)
· Author, How Google Tests Software
Previously:
Built product on Google Search personalization
· Relevance & quality @ Bing
· Founder, Test.ai (Google-funded)
· Author, How Google Tests Software
Previously:
Engineered Google Search personalization (Twitter & Google+ social-signal ranking)
· Bing relevance evaluation at scale
· Founded Test.ai for AI test infrastructure (Google-funded)
· Author, How Google Tests Software
Google had the researchers to hand-tune equations. Microsoft used AI to leapfrog.
00 · THE TRAINING LOOP
How the neural ranker learns.
01
Human Labelers
"best pizza NYC"→pizzahut.com3
"best pizza NYC"→joespizza.com9
"python tutorial"→docs.python.org8
Rate (query, page) tuples 1–10
02
AI Trains
Model learns what humans consider relevant
03
Ranked Results
#1joespizza.com0.94
#2difara.com0.87
#3pizzahut.com0.31
Ordered by model's predicted relevance
Human judgment becomes machine intelligence. Millions of labels, one ranking model.
The setup · Why this talk exists
Every engine flattens out at 80.
Climb the relevance curve with better models, more compute, smarter ranking — and you still
asymptote at the same ceiling: the best possible answer for everyone.
You cannot engineer your way past 80 with a better engine. The last 20 points belong to the person.
The problem · one word, two people
Search "bush". Now — which bush?
A plant, a president, and a rock band walk into a results page.
Whether this ranking is right depends entirely on who's looking at it.
Shrubs, right on top!
THE GARDENER · RESULT #1 = PERFECT
I wanted the BAND…
THE FAN · THE BAND IS BURIED AT #3
⌕bush✕
Aalmanac.com › gardening › shrubs
How to plant and prune bushes — complete shrub care guide
Boxwood, hydrangea, and juniper: planting depth, watering schedules, and seasonal pruning…
Wwhitehouse.gov › about › presidents
George W. Bush — 43rd President of the United States
Biography, term highlights, and the presidential library and museum…
Bbushofficial.com › tour
Bush — official site of the rock band
Tour dates, "Sixteen Stone" anniversary edition, and the new album…
About 2,140,000,000 resultsranked for the average user
One ranking cannot be right for both of them. Relevance is personal — that's the whole talk.
A story · Bing, c. 2010
I once proposed personalizing search by demographic. Then we did the math on measuring it.
The algorithm was easy. The hard part was
proving each cohort's ranker was actually better
— with paid human raters, at statistical confidence.
The product was simple. The
per-cohort eval
— raters × queries × cohorts × statistical floor — wasn't.
The ranker change was trivial. Per-cohort offline eval needed
N_raters × M_queries × hours × $rate × C_cohorts,
each cohort hitting its own significance threshold.
$12M
to measure it. For the basic cohorts.
Even Microsoft couldn't afford it.
At least, not back then.
Personalization didn't stall on compute. It stalled on measurement economics.
That's the story of the next 45 minutes — and why it's about to flip.
Personalization didn't stall on compute. It stalled on measurement. AI just collapsed the cost of evaluation.
Personalization stalled on rater throughput, not ranker quality. LLM-as-judge + synthetic users collapsed the eval cost curve.
The problem · complexity
Personalization is harder than it sounds.
The cohort math gets ugly fast.
The naive approach blows up combinatorially.
Forget AI for a second. Just take the basic
Madison Avenue demographics
— the categories the ad industry has used to slice audiences for 60 years.
Watch how fast it blows up.
Age bracket×6
Gender×3
Income tier×5
Education×4
Geography×3
Household×4
Ethnicity×6
Industry×12
6 × 3 × 5 × 4 × 3 × 4 × 6 × 12 = 1
311K+
distinct audience cohorts.
And that's before
psychographics, behavioral signals, intent, life stage, or anything personalized AI
would actually want to know about you.
Every one of them deserves a relevant AI.
Personalizing for the "average" means under-serving every one of them.
311K cohorts is 311K subsets to score — and the eval cost compounds at every one.
War story · social signals, circa 2012
So we borrowed your social graph instead.
If per-cohort relevance was unaffordable, maybe your tweets, likes, and circles could tell us
what you cared about. It worked.
🐦 Twitter
Who you follow, what you tweet — live interest signal.
👍 Facebook
What you like and share — your taste graph.
🪦 Google+
Who you circled. (2011–2019, rest in peace.)
ACCESS REVOKEDAPIs SHUT DOWN · DATA DEALS CANCELED · 403
+8%
relevance lift 🎉
UNTIL EVERY NETWORK PULLED THE RUG
The signals worked — and that's exactly why they got locked up. Borrowed context is a rug that gets pulled.
The thesis
Intelligence is converging.
Personalization is the new moat.
The next decade of AI competition won't be won by smarter models.
It'll be won by systems that know the user — their context, their history, their intent.
Yesterday
Bigger models
Today
Better reasoning
Next
Knowing you
The thesis · For this room
The advantage is moving up the stack.
Raw model quality is converging. The durable product advantage is the surface around the model —
how it adapts to each user, what it remembers, how its output reshapes over time.
Commoditizing
Model quality & price
Where value now lives
Context & memory
What product wins
Surfaces that know you
The thesis · For builders
The model is becoming a dependency.
Frontier models converge in quality, latency, and price. The layer above the model —
user context, memory, behavioral signals, identity — is where the durable advantage will live.
Substrate
Foundation model
Context layer
User-context APIs
Surface
Adaptive client
The flip side · When personalization misfires
And sometimes the right answer is not to personalize.
And sometimes the right product call is not to filter.
And sometimes the right ranker behavior is not to filter.
Another major search company was anti-personalization for years.
The argument was sharper than the Silicon Valley caricature.
It went like this:
Suppose you're a foodie. You told the system you hate McDonald's.
So the search engine filters it out.
Then there's an incident at McDonald's
— a food-safety scandal, a labor moment, a major news story.
You'd want to know.
You wouldn't.
The paradox
Stated preferences are not the same as information needs.
Filter what someone said they don't care about, and you eventually filter out
the moments they would have absolutely wanted to see.
Personalization without judgment becomes suppression.
Stated preferences need a layer of judgment — what's important sometimes overrides what someone said they want.
Stated preferences alone over-filter. The ranker needs an importance signal that can override negative preferences in rare cases.
The deeper problem · Validation
And you can't really validate personalization on one person.
And per-person validation breaks the eval pipeline.
And N=1 wrecks any offline-eval setup.
~50K
graded results per new ranker version.
~few K
searches a typical person issues in a lifetime.
The gap
You can prove a ranker works at population scale.
You can't prove it works for you.
Even with infinite money and perfect human raters, an individual user
will never issue enough queries — in a testing window, or in a lifetime —
to give you statistical confidence the personalized ranking is right
for them.
Population proxies break down at N = 1.
For fifteen years this was the wall — not the algorithms, the
measurement.
That's the wall AI is finally taking down.
For fifteen years the wall was measurement, not ranking.
AI just made per-user evaluation possible.
Eval at N=1 was the wall, not retrieval or scoring. LLM-judges and synthetic users finally make per-user eval tractable.
Meanwhile · The intelligence race
Frontier models are converging on the same answer.
Model differentiation is collapsing across the frontier.
Frontier model outputs are converging within noise.
GPT
CLAUDE
GEMINI
LLAMA
MISTRAL
→
Nearly identical output.
Generic. Polished. Forgettable.
Raw intelligence no longer separates them — and the labs know it. So they're racing to
bolt on the one thing that does: memory and personalization.
Shipping right now
They're adding memory.
›ChatGPT Memory — recall across every chat
›Claude projects & persistent context
›Gemini personalization from your Google data
›Custom instructions, profiles, saved preferences
But watch the motive
Built for stickiness — not relevance.
The goal is to raise switching costs and keep you on one platform.
Your memory becomes their moat.
Useful to you? Sometimes. Designed around you, owned by you, portable for you? Not even close.
Memory that locks you in is not the same as memory that works for you. That gap is the opportunity.
Memory built for retention isn't the same as memory built for relevance. The product opportunity is the gap between the two.
Vendor-resident memory is a moat for the vendor. User-resident memory is a moat for the product. The architectural fork is the opportunity.
Definition
Personalization, defined.
What personalization actually means.
Defining the context layer.
Adapting experience, output, and decision-making
to the individual user
— over time, across contexts, with memory.
A system that adapts output and recommendations
to each user's context, history, and intent
— continuously, across the products they use.
A context layer that exposes
a user's preferences, history, behavior, and intent
to any AI client via stable, queryable APIs.
Inputs the system learns
Preferences
History
Goals
Expertise
Writing style
Relationships
Constraints
Behaviors
Workflow
Location
Devices
Long-term memory
Precedent
Every major wave of consumer tech won the same way.
The same product playbook keeps winning, era after era.
Every era's breakout product was a per-user ranker.
Each era's breakout product won on one thing — showing each person what was relevant to them.
1998
Search
Google's ranking
2006
Social
Facebook News Feed
2010
Streaming
Netflix recommendations
2015
Music
Spotify Discover Weekly
2018
Feed
TikTok For You
NEXT
Every experience
will be personalized.
The product that knows you wins. The product that doesn't, leaves.
The evidence · Happening now
You can already see it everywhere.
We just don't call it personalization.
The product market is already shipping this — under a hundred different names.
Every platform is building context plumbing — under its own brand and API.
Every frontier AI platform is racing to ship features that customize the AI for the user.
They each invented their own brand for it. They're all the same idea —
personalization, in pieces.
Frontier Tuning
Microsoft · just announced
Tune frontier models against your own data and context — adapted at the weights layer.
Personalization, baked into the model.
Claude Skills
Anthropic
Bundles of instructions, tools, and context telling Claude how to behave for a specific job.
Personalization for a task.
Custom GPTs
OpenAI
Pre-loaded prompts, knowledge, and actions wrapped around a use case.
Personalization as a product.
ChatGPT Memory
OpenAI
The model remembers facts about you across sessions.
Literally memory.
Gemini Gems
Google
Reusable AI personas with custom instructions and context.
Personalization as personas.
MCP servers
Industry protocol
Plug your data, tools, and services into any LLM.
Personalization as context plumbing.
Cursor rules & IDE memories
Cursor, Windsurf, Cline, …
Codebase-specific instructions the AI follows on every change.
Personalization for code.
Different names. Same shape. The market is already moving — but every piece is trapped inside one platform.
Skills, GPTs, Memory, Gems, MCP — same shape, different vendors. The market is shipping personalization in pieces, locked to each platform.
Skills, GPTs, Memory, Gems, MCP are all user-context plumbing under different brands. Each is platform-local; none are portable.
The evidence · WWDC, this week
Apple just bet the operating system on knowing you.
Siri AI: the OS itself is now the personalization layer.
iOS 27: personal context becomes a system-level capability.
This week at WWDC, Apple unveiled Siri AI in iOS 27 — Siri rebuilt
around your personal context: photos, calendar, email, messages, contacts,
files, and what's on your screen right now.
Personal context, system-wide
Siri reads across Mail, Messages, Photos, Calendar, Contacts, and Files —
and answers through what it knows about you.
On-screen awareness
Acts on whatever you're looking at — building multi-step calendar events from on-screen text,
surfacing recommendations from past messages.
Context without the prompt.
Persistent memory
Conversation history that persists and syncs across devices via iCloud.
Memory as a default, not a feature.
But: locked to the platform
The deep context is siloed in Apple's native apps — third-party access is an open question.
Your context, their walled garden.
The world's biggest company just validated the thesis: the next moat isn't the model — it's knowing the user. And it's platform-locked. The land grab is on.
Apple didn't ship a smarter model — it shipped deeper user context. And that context stays inside the OS. Remember this slide when we get to who owns your AI identity.
Apple's differentiation is the context store, not the model — system-level, on-device-first, and non-portable. The strongest proof yet that the context layer is the product.
Architecture
The personalization stack.
The product stack above the model.
The system stack around the model.
Adaptive reasoning
Identity & preferences
Workflow integrations
Cross-app context
Behavioral signals
User memory
Foundation model (LLM)THE SUBSTRATE · EVERYTHING STACKS ON TOP
Key idea
The personalization layer may become more valuable than the model itself.
The personalization layer is where durable product advantage lives — above the model.
Treat the user-context layer as its own service with versioned APIs. The model becomes a swappable dependency.
Models are increasingly interchangeable. The stack that surrounds them —
memory, signals, identity — is what compounds.
Live example · Search
Search, re-ranked for you.
Per-user re-ranking, in one model call.
An LLM re-ranker over a generic SERP.
Same query. Same web. A single LLM pass re-orders the results for who you actually are.
Containerize and deploy a Python API on GKE Autopilot
Production-grade: build the image, push to Artifact Registry, autoscale with zero node management…
↑ FOR YOU
Ffastapi.tiangolo.com › deployment › cloud-run
FastAPI on Cloud Run — a staff engineer's deployment guide
The production checklist: httpOnly auth, CI/CD, zero-downtime revisions, cold-start tuning…
↑ FOR YOU
Rreplit.com › templates › python-api
Deploy a Python API instantly — Replit starter template
Click Run. Your API is live. Great for prototypes, hackathons, and demos…
iibm.com › think › topics › api
What is an API (application programming interface)?
An API is a set of rules that lets software applications communicate with each other…
10 results⚡ applying your profile…re-ranked in 740ms
What's happening
One LLM call rewrites the ranking using who the user actually is.
Same query. Same web. The order is now tuned to a senior engineer who already knows what Heroku is.
Cost: fraction of a cent. Latency: sub-second. Infra: none — just a prompt.
Live example · Chat
Even AI can be personalized.
Personalize the response — with guardrails — in one click.
A follow-on prompt that re-asks with context + policy.
One click fires a follow-on prompt that re-asks the model with your context — and your guardrails — attached.
AI assistant · thread
What just happened
The button doesn't change the model. It silently re-asks with context.
A follow-on prompt is auto-generated — pulling who the user is, what they're building, and what they can't share — then re-runs against the same model.
Guardrails travel with the user.
Compliance, safety, and tone all live in the personalization layer — not baked into the foundation model.
The horizon
Really everything.
Same facts. Same sources. Rewritten in real time to match your politics, your reading level, and your attention.
AI assistant declines meeting on user's behalf, citing "no clear agenda."Finally, an AI with boundaries: assistant RSVPs "no" to agenda-less meeting.AI now declines your meetings — what could possibly go wrong?
The model flagged Thursday's invite as agenda-less and low-priority; its user reports recovering two hours of deep work.Productivity researchers celebrate — calendar invites containing actual agendas are up 300% since the feature shipped.Critics note the model also declined its user's performance review and a dentist appointment. The vendor calls this "directionally correct."
🗞️
by Sarah Chen · Tech Desk · framed for you in 0.3s
✓ Same facts.✓ Same sources.Personalization adapts framing, depth, and tone — not truth.
Each loop deepens the moat — and shallows competitors'. Watch the meters climb with every revolution.
The gap
Personalization today is broken — and the labs can't fix it.
Today's personalization is fragmented — and platforms won't unify it.
Current personalization is siloed by architecture — vendor incentives won't change that.
Every app silos its own thin version of you. And the companies with the most data
have the least incentive to set it free.
What's broken today
›Fragmented — every app rebuilds you from scratch
›Not portable — context can't travel across models or tools
›Locked in — your memory becomes the vendor's switching cost
›Surveilled — personalization gets conflated with data extraction
Why frontier labs won't solve it
›
Users run many models — no single lab sees your whole context
›
Enterprises demand portability — lock-in is a procurement red flag
›
Centralized memory is the wrong shape — and trust keeps eroding
›
Their whole incentive is lock-in — portable personalization is the opposite
Centralized control is exactly the wrong shape for something this personal.
Centralized personalization can't be neutral about portability. The product shape has to change.
The user-context store has to live with the user. Every other architecture leaks data or locks-in by design.
The answer
User-owned personalization.
Personalization that belongs to the user.
Portable, encrypted, user-owned context.
Your AI identity should belong to
you
— not the platform you happen to be using this quarter.
User context belongs to the user — and travels with them across every AI surface they touch. Portability isn't a feature; it's the shape.
User context becomes a portable, encrypted store the user owns
— read by any AI client over standard APIs, never trapped inside a single vendor.
Properties of a user-owned personalization layer
Portable
Inspectable
Editable
Cross-model
Cross-platform
Privacy-controlled
The multiplier
Personalization isn't just for people.
Companies are users too.
Org-context is just another personalization slot.
A generic agent can do the task. A personalized agent does it the way it actually
needs to be done. And here's the part most people miss —
"you" is just as often a company as a person.
Adapting an AI to an organization is personalization too.
A company's context is its profile — regulations, policies, restrictions, brand rules
Compliance as personalization
The a16z rule.
Firms like a16z must append "this is not investment advice" to every public
and social post. A personalized agent carries that compliance profile and enforces
it automatically — on every draft, every time. Nobody has to remember.
Regulation as personalization
The HIPAA rule.
A healthcare company's AI must never expose — or train on — real patient data.
That regulation is part of the organization's profile. The agent adapts to it
exactly the way it adapts to a person's preferences.
An agent that ignores your company's rules isn't just impersonal — it's undeployable.
An agent that doesn't respect an organization's policies can't be shipped into a real workflow.
Without an org-policy bundle wired into every call, the agent fails review long before it reaches production.
In practice · Everyday AI
Your feed, built from your world.
A custom feed assembled from real-time context.
Multi-signal, LLM-ranked, per-user feed.
News, recommendations, what's nearby — gathered from several signals and
ranked in real time. The logic behind it is just an instruction:
The prompt behind the feed
"Given Jason's interests, nearby places, and today's news — rank them in one list with a confidence score for how relevant each is to Jason."
Four signals in. One ranked feed out.
Four signals — location, interests, nearby, news — merged into one ranked feed.
Four context sources → one LLM rank call → an ordered feed with per-item relevance scores.
Step 1Your location
Step 2Your interests
Step 3Fresh news
Step 4Your ranked feed
1
2
3
Fintech
Climate
AI agents
Design
Startups
Cycling
2h ago
4h ago
today
today
NearbyFintech
AI agentsFresh
NearbyStartups
TodayDesign
The horizon · Generated interfaces
Results stop being links. They become apps.
Why hand you ten blue links when the AI can spawn the exact tool you need —
a planner, a lesson, a workflow — personalized, interactive, and disposable.
⌕plan emma's 8th birthday party
APPLET · GENERATED FOR YOU
Emma's 8th — Party Planner 🎈
✓Pick a dateSat, June 20
✓Send invites12 sent · 9 yes
Order the cakechocolate · pickup 10am
Goodie bagsyour craft-store run
Next: order the cake →
⌕what are the names of the planets?
LESSON · BUILT FOR A CURIOUS 10-YEAR-OLD
The planets, in order 🪐
⚠ not officially a planet anymore! …sort of
Mercury
Venus
Earth
Mars
Jupiter
Saturn
Uranus
Neptune
Pluto
In 2006 Pluto was demoted to a dwarf planet — big enough to be round, but it never cleared its orbital neighborhood. So… is Pluto a planet?
YesNoSort of… ✓
The interface itself becomes a personalized, generated artifact — built for one person, one moment, one task.
The horizon · The end of one-size-fits-all
Every app is the same for everyone. Soon it's built for you.
Today's apps are the lowest common denominator: buttons you'll never tap, and the Starbucks
home screen pushing triple-shot promos at someone who has never once ordered caffeine.
☕
☕ STARBUCKS
Menu
Scan
Cards
Stores
Rewards
Gifts
Order ahead
Music
Tipping
Settings
☕ NEW! Triple-Shot Venti — extra caffeine!⚡ Nitro Cold Brew: our MOST caffeinated ever
TODAY · TAP, HUNT THROUGH MENUS · ADS FOR DRINKS YOU NEVER ORDER
💡 Did you know? Frappuccinos don't have caffeine.
Good morning, Jason ☀️
📍 drive-thru detected
🕢 7:42 AM · on your way to work
☕ your usual: decaf oat latte
Pay $5.40 · your usual
SOON · AI SPAWNS THE APP AT THE DRIVE-THRU · ONE GLORIOUS CODE
No menu. No hunting. The interface knows where you are, what time it is, and what you always order.
War story · 2013
"Realtime" personalization — on Google Glass.
I built a Glass app that watched where you walked. Approach a Target, and by the time you
reached the door, the HUD was already showing what's on sale that you'd actually want.
▸ TARGET · 140 FT▸ TARGET · 80 FT▸ TARGET · 20 FT▸ AT THE DOOR ✨
yes, they called us "glassholes." worth it.
GLASS · FOR YOU @ TARGET
LEGO Space sets −30%
AISLE 7
Trail mix you buy monthly BOGO
AT THE DOOR
USB-C cables −20% · grab & go
2013 hardware, 2026 idea: context + identity + location = relevance before you even ask.
The costs
Hyper-personalization has real downsides.
Personalization done wrong becomes a product liability.
Personalization without guardrails becomes a surveillance system.
⚠Filter bubbles
Your AI only shows you the version of the world you already believe.
⚠Manipulation
Knowing the user is the same skill as exploiting the user.
⚠Surveillance
Personal context becomes the asset, not the user.
⚠Overfitting
Models that mirror you can stop challenging you.
⚠Behavioral dependency
Convenience that quietly erodes your ability to choose.
The question
How do we preserve user agency in a world where AI knows us better than we know ourselves?
How do products build agency in — letting users see, edit, and override what the AI knows about them?
Surface the user-context store to the user: inspect, edit, delete. Without those primitives, personalization becomes surveillance.
User-owned personalization isn't just a product choice — it's the structural answer.
The flip side · personalization in reverse
AI can personalize as someone else, too.
Not just content for you — content as seen by someone who isn't you.
The same machinery that builds filter bubbles can break them.
VIEWING AS:
YOUA SKEPTICAL BOARD MEMBERYOUR TOUGHEST PROSPECTSOMEONE WHO DISAGREES WITH YOU
🫧
Break your own bubble
Read today's news as someone who disagrees with you. The bubble pops the moment you can borrow another lens.
🎭
One deck, every audience
Personalize the same slides for execs, builders — or an AICon track session. You're watching this deck do it live.
🪑
Sit in meetings you missed
Simulate the room: replay the pitch as each attendee and learn what they heard — not what you said.
🤝
See it as the prospect. Or the board.
Show the same numbers through their eyes — and surface the objections before the meeting, not during it.
Impersonation is empathy at scale — the antidote to the filter bubble, not just its cause.
THE PRIZEYour AI identitymemory · preferences · behavior
YouPortable, sovereign, encrypted.
Operating systemsApple, Google, Microsoft — Siri AI just claimed this ground.
BrowsersWhere the AI surface actually lives.
Frontier labsCentralized memory inside the model vendor.
EnterprisesYour employer owns your professional AI self.
Open ecosystemsStandards-based, interoperable personalization.
This will define the next decade of AI competition.
This is the next decade's product question — and it's yours to answer first.
This is the next decade's architectural fork. Pick early, and pick portably.
What to do Monday
If I were running AI strategy this quarter…
Three bets that pay off whether you're the CEO, the head of product, or the engineer
picking the next tool. None of them require building a frontier model.
01
Treat personalization as a layer, not a feature.
Stop scattering preferences across products. Build (or adopt) one user-context layer
that every AI surface reads from. The layer compounds. The features don't.
Owner: CTO + Head of Product
02
Measure personalization with AI — not just humans.
The $12M problem is solvable now. Synthetic users and AI personas let you validate
per-cohort and per-user relevance at a fraction of the cost. Build that capability
before you scale personalization.
Owner: Head of Quality + Data
03
Make user context portable on principle.
Export, import, and inspect should be first-class. Lock-in feels like a moat today;
it'll feel like a liability the moment users (and regulators) catch up.
Build the moat from relevance, not friction.
Owner: CEO + Legal + Product
The companies that win won't have the smartest AI. They'll be the ones whose AI knows their customer best.
What to ship after this talk
If I were in your seat at AICon this week…
Three product moves you can ship without convincing the board to fund a research lab.
Each is a roadmap item, not a strategy memo.
01
Ship one personalization layer — not feature-by-feature personalization.
Pick the user-context object you want every surface to read from. Define it once. Wire
search, chat, notifications, recs, settings all through the same object. No more
per-feature preference toggles.
Lead it: PM + Eng lead
02
Stand up an eval pipeline before you scale personalization.
Define personas, write LLM-judge prompts, run them in CI. Per-cohort relevance scores
become a dashboard your team checks daily. You'll find bugs in production rankers
you didn't know you had.
Lead it: PM + Quality lead
03
Make export, inspect, and delete first-class product surfaces.
A user-facing page that shows exactly what the AI knows about them. Edit any of it.
Export it. Delete it. This is the kind of surface that builds trust — and that
regulators are about to require.
Lead it: PM + Privacy/Legal
The product that knows the user wins. The product that lets the user know what it knows wins twice.
The prediction
The winning AI systems
won't be the smartest.
Or the biggest. Or the fastest. Or the cheapest.
They will be the ones that are
Most relevant
Most trusted
Most adaptive
Most personal
Personalization is the moat. The race is on for who owns it.
The prediction · For this room
The winning products
won't ship the smartest AI.
Not the most capable. Not the most novel. Not the most differentiated on the model card.
They'll be the products that
Know the user
Earn trust
Improve with use
Feel made for me
The product layer above the model is the durable advantage. Build there.
The prediction · For builders
The winning stacks
won't own the biggest model.
Not the most parameters. Not the lowest cost-per-token. Not the proprietary weights.
They'll be the stacks that
Own the context layer
Treat models as swappable
Eval at N=1, in CI
Ship context portably
The moat is the layer above the model. Build the API. Make it portable. Make it inspectable.
Intelligence determines what AI can do.
Personalization determines what AI should do — for you.
Q & A
What parts of your digital identity should AI remember — and what should it forget?
What part of your product's personalization should the user own — and what should the platform?
Where does your user-context store live today — and what would it take to make it portable?
Jason Arbon
Founder & CEO, Testers.AI
testers.ai · jason@testers.ai · @jarbon
Play with it yourself
TOOND.AI
https://toond.ai/
Interested in playing with full web personalization?
Sign up — and I'll share the Chrome extension soon.
