AI is doing the shortlisting
Your buyers are asking AI assistants instead of searching. Those systems read, compare, and come back with one answer, and a brand they cannot read properly never makes the list. Everything you publish was built for a person with a browser. ZCET is the layer between that material and the systems now reading it. It rebuilds what you have so an AI agent can find it, read it, check it, compare it and act on it. After that, it is where the new material gets made.
- Increase visibility In the sources
agents read - Drive selection Facts an agent
can compare - Enable action An interface an agent
can call - Build trust Claims with provenance
behind them - Deliver impact Measured by selection,
not clicks
The shortlist is now drawn up somewhere
you cannot see
For twenty years someone searched, compared a shortlist, and chose. That is being handed to AI. An assistant reads, filters, and comes back with a recommendation, or with the purchase already made. Value used to be created when a person arrived on your site. It is now created at the moment an agent picks something, somewhere you are not present and will not be shown.
And there is no such thing as the agent. There are several, and they do not read alike. OpenAI, Google and Anthropic have each built on different foundations, and a company can build its own on top of any of them. A buyer's assistant and a procurement system working through a shortlist are not the same reader. The one reading you this quarter may not be the one reading you next.
So no single format satisfies all of them, and preparing for one is not preparing for the rest. What holds up across them is the same shortlist every time: facts published as data rather than sentences, claims a stranger can check, and something an agent can actually call.
The reflex is to treat this as a visibility problem and buy the same service under a newer name. That gets it backwards. Visibility assumes a person is still choosing and needs to see you.
Nothing in your reporting will tell you
Everything your brand publishes was built for a person reading a page. An agent needs it in a form it can work through, and most brands have never had a reason to produce one.
When that fails, it fails quietly. Facts a system cannot verify get routed around. A catalog it cannot compare on the right attributes loses to one it can. Nothing is rejected. It is simply not carried forward.
ZCET calls this navigational debt: the gap between material built to be found by a person searching and material an agent can select.
And none of it leaves a trace. No impression, no click, no bounce, no drop-off to investigate. The comparison happened somewhere you were never present, and a brand that was not in it does not register as a loss.
Every number your marketing team watches measures what people do on your site. These buyers were never going to be on your site.
PICSA²
Six pieces of work on what your brand publishes.
Presence. Interpretability. Credibility. Selectability. Actionability. Accountability.
The first five decide whether an agent can find you, read you, check you, compare you, and act on what it finds. Each is a separate rebuild, each is individually necessary, and none is a step toward the others. Doing four of them well does not carry the fifth.
Accountability was added to those five. The layer is the t3ai-Ledger, and it has nothing to do with being found or chosen. It decides whether you can show, afterward, what you published and what an agent did with it.
Presence
Your information has to be where AI systems already look. A website waits to be crawled; a product feed, an industry registry and a distributor's data are already being read. A brand in none of them is not ranked low in the answer. It is not in the answer.
- Published in the sources agents read, from Common Crawl to industry registries
- In the formats those sources accept
- Actively distributed, not waiting to be crawled
Interpretability
The same fact can be published two ways. We ship within a week is a sentence a person reads easily. Seven days, in a lead time field, is the same fact a machine can compare. Given only the sentence, it cannot line you up against anyone, so it does not.
- Facts as Schema.org data, not paragraphs
- Explicit relationships between them
- One meaning and one permanent ID per claim
Credibility
A claim either can be checked or has to be taken on trust. Independently tested with nothing behind it is a phrase; a test result with a date, a method and whoever signed it is a fact. A system that cannot verify the first one does not argue with it. It leaves it out.
- Built for Content Credentials (C2PA) on published assets
- Signed and timestamped
- Verifiable by a stranger
Selectability
Your information has to line up with what an agent actually compares on. A brochure leads with heritage and craft; a shortlist is decided on capacity, tolerances and what it is compatible with. A brand that publishes only the first set can be read and cannot be ranked, so it is not ranked.
- Organized around the compared attributes
- Complete on the fields that get weighed
- Current, not last year’s
Actionability
An agent needs something to act against. Contact us for pricing assumes a person will pick up the phone, and an agent supplies none of that effort. It does not wait for a callback. It buys from whoever published a price.
- Current pricing and availability, readable live
- A documented interface to call, in OpenAPI and the Model Context Protocol (MCP)
- Terms that can be read without a phone call
Accountability
What you published, and what an agent did with it, has to be provable after the fact. This one is not about being found or chosen. The sixth is the t3ai-Ledger: an append-only record of what was produced, who acted on it, and under which policy, designed so that it cannot be quietly edited and so that a party outside your organization can check it against an exported archive.
- What was published, and from what sources
- Who approved it, and in what role
- Unaltered since, and checkable by an outsider
Three planes, kept apart
ZCET is designed in three planes, and what separates them is who can reach each one.
The first is internal. It is where your source material is taken in, turned into facts and signed before anything is published. It holds any personal information that material contains, and it is designed so that no agent ever reaches it.
The second faces outward. It holds what you have chosen to publish, in the forms agents read, and the interfaces they call. It is designed to be the only part an outside system can reach, and nothing personal is in it.
The third observes. The accountability ledger records what was published and what an agent did with it, from outside the path the data travels, so it can show both without standing between you and your buyers.
Actionability comes in five depths
Every other layer is something you have or you do not. This one you can be partway through, and placing yourself takes about a minute.
Human-only. Everything a buyer needs is behind a phone call or a contact form. An agent can read that you exist and can do nothing with it.
Read-only page. An agent can fetch your pages and read what is written on them. What you publish is prose meant for a person, so it can quote you while working your prices out of a sentence.
Structured feed. It can pull one fact without reading the page around it, because your facts are published as data rather than sentences. This is where you start being compared on the attribute that matters instead of on how well the page was written.
Live query API. It can ask your systems a question and get an answer that is true at the moment it asks, rather than true on the day you published. Stock, availability and anything else that moves stops being a guess.
Completable. An agent can go from a question to a placed order, under terms it can read, without reaching a person.
Everything below live query is work on a website. The last two mean opening a system, which is a different department and a different budget. That is the line that decides how far up you get, and it is not a technical one.
The Selection Audit
The work starts by finding out how agents see you now. The Selection Audit sends a live agent at what you publish and reads what comes back on four measures: how complete it is, how well it is structured, whether an agent can act on it, and whether it can be trusted.
The result is a report of what is missing and where. Two gaps are named in it: a credibility gap, where claims have no provenance behind them, and an actionability gap, where there is no interface an agent can call.
The output:
- Variable Deficiency Report (VDR)
- Credibility gap(missing provenance)
- Actionability gap(no callable interface)
Selection, not clicks
Click-through rate counts a person arriving on your site. When an agent does the choosing there is no click to count, which is why the loss described earlier leaves no trace.
ZCET replaces it with the Selection Probability Score: how likely an agent is to choose you, measured across the whole path from not being seen at all to being the obvious choice. It is designed to be reproduced from the ledger and the interface records, so the result can be checked rather than simply reported.
Brands whose buyers have started delegating
This matters where the purchase has a specification behind it. Products with attributes that get compared. Services with terms that get checked. Catalogs, inventory, pricing, availability, credentials.
It arrives sooner in regulated sectors, where a claim has to be substantiable and an auditable record of what you published will be asked for regardless.
And it arrives wherever buyers have started handing the search to something else, which is already further along than most reporting can show.