Essay · 006
2026-10-30From Better Inputs to a Closed Loop
When I started Context Catalog earlier this year, the problem I was trying to solve was mostly about inputs.
A prompt is a pretty thin container for everything a brand knows about itself. Claims, voice, audiences, product truths, scenes, creative history and all the other things that help a good marketer make a good decision tend to live across briefs, decks, Slack threads and people’s heads.
So the original idea was simple: give generative creative systems a better input layer. A structured, portable Context Catalog that says what is true about the brand and what the system should know before it starts making things.
I still believe that.
But over the last few months, I’ve become increasingly convinced that better inputs are only half the problem.
The other half is what comes back.
AI has made it dramatically easier to produce more advertising. What it has not necessarily made easier is learning from that advertising.
That is the direction Context Catalog is now moving in.
The new proposal is called the Context Catalog Exchange. The catalog remains the foundation, but the goal is now to connect how an ad was made with how the individual parts of that ad actually performed.
In other words, move from better inputs to a closed loop.
More ads, not necessarily more learning
There are three problems I keep coming back to.
The first is the flood. If one idea can become 50 or 100 variants almost instantly, those variants can end up competing with each other for the same audience and splitting the available test budget. More output does not automatically mean more useful information.
The second is that every new ad effectively starts from zero. You might reuse the exact same product demonstration that has already performed well in three previous ads, but the new creative is still generally treated as an unknown. You pay to establish confidence in it all over again.
And then there is the scoreboard itself.
Advertising results usually come back at the level of the whole asset. That can tell you which ad got more clicks or drove more conversions, but it does not necessarily tell you which part of the creative actually did the selling.
A sensational opening might attract attention while a fairly boring product demonstration is the thing that convinces a valuable customer to buy. If all we know is that the ad “won,” those two effects get bundled together.
That makes the next creative decision surprisingly crude: make more things that look like the winner and hope we guessed correctly about why it worked.
What if the ad carried its own history?
The Exchange is an attempt to make that loop more explicit.
It starts with the catalog. The brand maintains a versioned record of what is true: approved claims, voice, scenes, audiences and the creative ideas currently being tested. Every ad is built from a specific version of that record.
The ad then carries a digitally signed description of its parts. This might say that seconds zero through two are the hook, seconds two through four describe the problem, seconds four through eight show the product, and the final section is the call to action.
Each segment has a fingerprint of its content and points back to the catalog version it came from. If a segment has been reused from an earlier ad, that lineage travels with it too.
The approach follows the ideas behind the C2PA content-provenance standard: content should be able to carry a trustworthy record of where it came from and what happened to it.
Before money is spent, there is a pre-flight check.
The goal here is to stop two kinds of waste before they enter the auction: near-identical creative that teaches us very little, and content that conflicts with the brand’s own rules.
The interactive simulation at contextcatalog.ai/exchange makes this concrete using a fictional B2B security company called Bramblegate Security.
In one simulated run, 10 near-identical variants arrive at the pre-flight check. Eight are stopped because they are effectively copies of ads already running. Two are allowed through. An eleventh variant is rejected because it uses a claim the brand has specifically prohibited: “guaranteed zero breaches.”
That is not a claim about live platform performance. It is an illustration of what the protocol is designed to do.
The click is not always the point
The more interesting part happens once performance comes back.
In the Bramblegate simulation, a sensational opening at 0:00 to 0:02 produces the highest click-through rate at 3.1%.
If CTR were the scoreboard, you would probably make more of it.
But the customers acquired through that section are worth only 0.4x over the following 90 days.
The plain product demonstration at 0:04 to 0:08 looks much less exciting on the surface, but the customers associated with it are worth 3.2x over 90 days, eight times as much.
That is the kind of distinction I think matters.
The Exchange is designed to connect segment-level creative history with longer-term customer value. Advertiser customer data can be matched in a data clean room, with only aggregated results coming back.
The system can then draft a proposed learning: perhaps the dramatic opening is attracting attention but the product demonstration is doing more of the actual selling.
Importantly, the system does not write that learning into the brand record by itself.
A person reviews it.
A person approves or rejects it.
Only then does the catalog move from one version to the next.
I think that human approval matters a lot. If the answer to AI-generated creative is simply another autonomous system producing more AI-generated creative based on noisy correlations, we have mostly built a faster machine for producing slop.
The point is to make the marketer better informed, not remove them from the loop.
Keep what worked. Replace what did not.
Once the learning has been approved, the next ad does not need to reinvent everything.
In the simulation, the next Bramblegate ad replaces the weak opening but keeps the proven product demonstration.
Because the demonstration carries a verified history, the simulation assumes that the new ad can establish confidence with about 800 test impressions rather than roughly 5,000, saving 4,200 impressions of redundant testing.
Again, those are illustrative numbers from the simulation, not measured results from a live platform.
The important idea is the inheritance.
If a piece of creative has already earned a track record, why should that history disappear simply because it has been placed inside a new ad?
And if someone tries to fake that history, the signed record should make tampering detectable. In the demo, changing a single byte in the claimed lineage causes the verification to fail.
Two possible paths
There are two ways I imagine this working.
The first is bring your own pixels. Make the creative wherever you like. The signed record travels with the finished ad, so provenance and learning are not tied to a particular creative tool.
The second is more ambitious: guided native generation.
In that model, the advertiser sends a signed blueprint containing its brand rules and the proven creative segments it wants to preserve. The advertising platform’s own generative systems could then build the remaining parts around them.
No advertising platform offers this today. It is a proposal.
But I think it becomes more interesting as generation moves closer to the point of delivery. If platforms are eventually generating or adapting creative in near real time, the advertiser needs a much richer way to communicate what is known, what is allowed and what has already been proven.
Why this might matter
For marketers, the goal is fairly straightforward: spend less time relearning the same things, understand more about which parts of creative are actually contributing to value, and keep the brand’s own knowledge portable and under human control.
For platforms, there is a different potential benefit. Fewer redundant ads could mean less unnecessary testing and ranking. Verified creative history could provide useful prior signal. And if advertisers become more confident that they understand what is working, it is reasonable to think that could support greater investment over time.
Those are design goals, not claims about current platform behavior.
Today, the Context Catalog Exchange is a patent-pending protocol demonstrated through an interactive simulation. The document formats are open, and the signing and tamper-detection pieces work. There are no live ad-platform integrations yet.
So this is still early.
But I think the direction is increasingly clear.
Essay 001 started with the idea that AI creative systems need better context before they make something.
Essay 006 adds the other half: they also need a way to remember what happened afterwards.
The interesting future is not just generating more ads.
It is making each ad leave the next one a little smarter.
You can try the interactive simulation at contextcatalog.ai/exchange, or reach me at info@contextcatalog.ai.