Who built Delphi Monitor, and why

Delphi Monitor is built and run by Phantasm and Brain. The same company writes the software, runs the measurements and answers the one support address this product names.

Nothing in a Delphi reading is licensed from somebody else's panel. Every answer that is scored is one Delphi put to ChatGPT, Gemini, Claude, Perplexity or Grok itself, stored word for word, so a reading can be traced back to the answers that produced it rather than taken on trust.

Counting mentions does not answer the question a commercial director has. Being named and being chosen are different outcomes, and a brand can be named in almost every answer and recommended in none of them. An answer that recommends somebody else also leaves no session, no referrer and no line in any report a brand already owns, so the month the shortlist is lost looks exactly like the month it was won.

So the reading is sequential rather than a single verdict. Found, then Seen, then Trusted, then Chosen, then Bought. A buyer has to reach the brand before an assistant can name it, and the brand has to be described accurately before it is recommended, and recommended before a purchase can start. The five are weighted and combine into one figure out of 100, published with the band around it, and the weakest step caps every step after it. That step is named as the binding constraint, and the plan starts there rather than with whatever is cheapest to fix.

The methodology page ends in a list of things the measurement cannot see, and that list is the argument rather than modesty. A buyer who finds a limit after signing is worse off than one who finds it during an evaluation, and so are we.

Three of those limits shape what a reading looks like. These assistants do not answer identically twice, so the Index is published with the band around it, and a movement smaller than that band is reported as no confirmed change rather than as progress. A run is a sample of the questions in one set on the day it ran, not a census of everything buyers ask. And a check that could not run is reported as unmeasured rather than as zero, because a zero says a brand failed at something nobody looked at.

Delphi refuses four things, and every one of them costs something.

• It does not claim revenue caused by an answer. Delphi measures position in AI answers and will not convert one into the other, which is the sentence a buyer would most like to hear.
• It does not report a simulated number. A panel with no data states what is missing and the one action that fills it, so an empty screen is a finding rather than a hole.
• It does not read retailer-owned assistants. What happens inside a retailer's own answer box is not measured here, and no reading implies that it is.
• It does not promise an outcome. Projected lift is always a range with a stated basis, never a single figure, because a point estimate is a promise wearing arithmetic.

Dropping any one of them would make the product easier to sell and would make every other number on it worth less. That trade is the whole reason to build this rather than something louder.

One address, read by a person: contact@delphimonitor.com. A demo request reaches a person too, and every one of them is picked up by hand. A first measurement takes a little time to arrange, and the reply comes back with next steps within two business days.