Which part of the journey is failing, and what to fix first

This is not a marketing funnel. It is the order an assistant works in when somebody asks it what to buy, and each thing depends on the one before it. Together they are what AI visibility is made of.

• Found. Can the systems behind the assistants reach and read you at all.
• Seen. Do you turn up when a buyer asks a question about your category.
• Trusted. Is what the answer says about you accurate, and does it read favourably.
• Chosen. Are you the recommendation, or one of several names in it.
• Bought. Does the answer end somewhere a buyer can act on.

Two things follow from the order. A brand can be excellent at four of the five and sell nothing. And moving one step lets more buyers through to the ones after it, without anybody touching those.

A buyer lost at the weakest step never reaches the steps after it, however well those steps score on their own. So a gain on your strongest step reaches only the buyers who already get that far, and the same gain on the weakest one also lets more of them through to the steps behind it.

That is also why an ordered plan is not the same as a ranked list of problems. Ranking by how large each weakness looks puts the biggest gap first, and the biggest gap is very often on a step that nothing downstream is waiting for. The measurement knows which one is holding the rest, and a template cannot.

Delphi names the weakest step the binding constraint, and the plan leads with it, unless your robots.txt refuses OAI-SearchBot, Claude-SearchBot, PerplexityBot, Google-Extended, Googlebot or bingbot: opening that comes first.

Four things decide whether a brand can be read. Whether the named crawlers behind the assistants are allowed. Whether pages carry the machine readable labels that say which text is the price, the rating, the ingredient list. Whether the facts a buyer needs are in the HTML as delivered, rather than appearing after a tab is clicked. And whether the assistants already hold the brand as a distinct thing rather than an unfamiliar string of letters.

It is the cheapest of the five to fix and the easiest to miss, because a refused crawler is invisible from the inside. Your analytics show the site working. The readers it is not working for never arrive, so they are not in the report.

Delphi identifies itself by name to every site it reads, and reports a block as a finding rather than working around it. A measurement that routes around a refusal would be measuring a site the assistants cannot see.

The distance between those two is the answers that name you and recommend a rival. Being known is not the problem there: the assistant reached you, and the recommendation went to somebody else.

Two readings sit underneath. How often you are the first name, and how often you reach the first three, which is the set a buyer actually considers. Both count only questions asked at the point of buying. An informational question has no recommendation to win, so it stays in the set and is excluded from this score rather than manufacturing a zero nobody could act on.

A brand named first and then advised against is not counted as a recommendation. An answer that opens with the one most people buy and then says it would not is a Trusted problem wearing a Chosen result, and counting position alone would score the most damaging answer in the run as a win.

The work here is usually not on your own pages, for the same reason a competitor shows up first: an answer is assembled from what the assistant read, so the useful output names the third party pages that were cited on the buying questions you lost, and how many of those questions each one touched. Where one is a publisher or review site, a pitch to it arrives drafted from the same measurement, and the outlet decides what to do with it. Where it is a retailer, what arrives is a request to your own account team for that retailer. Where it is a community, what arrives is the thread and what to bring to it, never a post. A plan drafts for a limited number of pages, and a page Delphi could not classify gets no draft.

That refusal is deliberate. If an assistant tells a buyer a product is discontinued, the sale does not happen, and it makes no difference to that outcome whether your page is right and the assistant is stale, or your page is out of date and the assistant is reading it correctly. Either way you are losing the sale and either way there is work to do. Presuming your page is the truth would send half of these cases to fix the wrong end.

Where neither your home page nor your product names address a claim at all, that is recorded as nothing to compare rather than as agreement, and accuracy is scored only where something was genuinely comparable. A brand nobody could check is never reported as a brand that got it right.

Severity is weighted rather than counted, because one invented negative can undo a whole content campaign and a wrong pack size cannot. Whether this is the thing to fix first still depends on the sequence: an accurate description earns nothing while no answer is naming you.

Auditing a brand's own storefront is the obvious way to measure this, and for most brands it is wrong, because most brands do not sell on their own site. It scores a shop that does not exist, and a number that reads near zero for everybody has stopped discriminating between anything.

What the assistant says at the end of an answer is the transaction path as the buyer experiences it. Where it sends them. Whether the offer carries a price, a pack size or a named variant rather than the brand name alone. Whether it tells them the product is discontinued, out of stock, hard to find or not sold where they are. And, weighted lowest of the four, whether your own site is ready for an agent to transact against.

Where a comparison cannot be made at all, for instance where your pages publish no price, it is reported as unmeasured and its weight moves to the parts that could be read. Not knowing outscores knowing it is wrong, which is the right way round.

The weakest step sets the order of the plan. It does not decide the whole plan, and it should not. Whether a brand can move a given step depends on circumstances outside its control, so a product that points at the weakest number and stops has told a commercial director to go and do something they may not be able to do.

So the rest is presented as a ranked set rather than a queue. Within it, each piece of work is weighed by how strong the evidence is for that kind of work against how much effort it takes, so quick work on strong evidence comes ahead of slow work on thinner evidence.

The sequence is a claim about how the five steps interact, and it is checkable rather than asserted. If the first piece of work ships and the step it targeted does not move by more than the measurement's own margin of error, that is reported as no confirmed change rather than as a win.

The trigger is what makes it derived rather than templated. It is evaluated in code rather than asked of a model, and there is no close enough: a recommendation fires only when its condition is met, and a signal the run did not measure resolves to false. Where fewer than seven fire, the plan is completed with the ones the measurement came nearest to firing, each marked as not yet met and naming its number, below everything that did fire. Where more than fifteen fire, the plan keeps the fifteen that lead its sequence.

• Trigger. The condition that fired, or for work added to complete the plan, the condition it came nearest to, with the measured signal either way.
• Target. The page, the domain or the part of your own site the work acts on, named rather than described.
• Mechanism. Why it works, and which of the five steps it is trying to move.
• Effort. A stated band with its reason attached. Nothing here measures effort, so the reason travels with the band rather than a bare letter standing in for one.
• Expected lift. A range with its basis, never a point estimate.

Every recommendation also states what would count as it having worked: what a next measurement would have to see, and what a next measurement still could not see whatever happens. Work ordered this way respects its own prerequisites, so nothing is scheduled before the thing it depends on.

The same inputs give the same plan. There is no sampling and no clock in the part that chooses, which is what makes a comparison between one month and the next mean anything at all.