What an AI visibility score is, and why two tools give one brand two numbers

The main unit of measurement is an answer. Delphi puts buying questions for a brand's category to ChatGPT, Google Overview, Gemini, Copilot, Claude, Perplexity and Grok, stores every answer word for word, and reads those stored answers for what they say about the brand. The assistant that gave each answer is recorded beside it, with the exact model wherever the assistant names one, so a figure drawn from answers traces back to sentences a person can read. The first of the five gates is read from the brand's own pages instead of the answers, and the last reads those pages alongside them, because whether a machine can read a site, and whether the brand is ready to be bought from, are not questions an answer settles on its own.

So the score is a position rather than a volume. It says where a brand stands among the names an assistant offers a buyer. It does not count how many buyers asked. That is a different measurement, and it is not one this figure contains.

The five are the parts AI visibility is made of, in the order a buyer passes through them, and each answers one question:

• Found. Can AI systems access and read the brand at all.
• Seen. Does the brand appear in AI answers for category questions.
• Trusted. Is it described accurately and positively.
• Chosen. Is it recommended at or near the top.
• Bought. Is the brand ready to be transacted when an assistant sends a buyer.

The gates are weighted, and they are not weighted equally. The weights are fixed by the model rather than set by the customer, so two brands measured by Delphi are measured on the same scale, and changing one is a change to the method rather than a setting.

Questions with no purchase intent stay in the set and are excluded from recommendation scoring. A question about how a category works has no recommendation to win, so counting it would create a false zero on Chosen for a brand that failed at nothing.

An assistant that cannot read a site describes the brand from what other sites say, whatever the site itself says. A brand nobody names cannot be the recommendation, however good its product pages are. So work aimed at a later gate can be finished and correct and still reach few buyers while an earlier gate stays low. Each gate keeps its own score, so a low gate shows beside the higher ones, and the lowest is named so a team can see where buyers are being lost.

That is why the lowest gate is named as the binding constraint and leads the plan. On a tie the earlier gate in the chain wins, because a buyer meets it first. The one thing that comes ahead of it is a robots.txt refusing OAI-SearchBot, Claude-SearchBot, PerplexityBot, Google-Extended, Googlebot or bingbot, because an assistant that cannot read the site cannot pick up anything corrected on it.

Weighting an unmeasured gate at zero states a finding. It says the brand failed at something nobody looked at, and it caps the Index at a ceiling the brand can do nothing about. Excluding the gate and rescaling the rest keeps the number a reading of what was actually measured, and the disclosure is what stops that being a quieter kind of dishonesty: a number computed over a smaller basis is a different number, and hiding that is a lie of omission.

The same rule runs one level down. An Index is computed from the assistants that answered rather than from the assistants that were planned, and where answers were collected and could not be read, that shortfall is stated beside the figures. A gate that was not measured is also never named as the binding constraint, so a plan can never lead with work nobody looked at.

A band word can change without the number meaning anything new. A single point across a threshold moves a brand from Fair to Weak while sitting well inside the measurement's own margin of error, which is why the dashboard and the history page call a movement smaller than that margin unconfirmed. Wherever the Index appears, the band around it appears too.

A score is comparable only with another score produced the same way. Setting an Index beside a figure from a different tool compares two different question sets put to two different sets of assistants under two different counting rules, and the arithmetic will not reconcile. The comparisons that do hold are a brand against itself over time on the same question set, and a brand against the rivals named in the same answers.

• Which questions were asked. A score is a reading of one question set. A set weighted toward the corners of a category a brand already owns reads high, and a set built from what buyers actually type reads lower. Neither is wrong and they are not the same measurement.
• How many times each was asked. These models do not answer identically twice. Asked once, a question records one draw, and a movement between two single draws can be noise. Asked several times, each assistant's figure rests on more than one answer and is more reliable. Delphi's band around the Index counts the spread between the assistants, each on the questions it answered, and which questions were asked, and how many questions a measurement needs sets out how it is made.
• Which assistants were asked. The assistants disagree with each other on the same question, and they disagree about which sources they read. A score over a few assistants and a score over many are different measurements even on an identical question set.
• How position inside an answer is weighted. Being the one name an assistant gives, reaching the first three, and being mentioned somewhere in the paragraph are three different commercial outcomes. A tool that treats them alike and a tool that separates them will report different numbers from the same answer.
• What counts as a mention at all. An answer can credit a brand under a sub-brand, a product line or the legal entity. A tool matching the trading name alone scores those as absent. A tool matching loosely credits a brand for a name that was never theirs, which is the same error in the more flattering direction.

The fourth is worth an example. Take a hypothetical set of four brands: Ardent, Kestrel, Trailhead and Nimbus, all invented for this page. An answer that opens by naming Ardent and then warns the buyer away from it puts Ardent in first position. Counted on position alone, that is the strongest possible result. Read for what it tells a buyer, it is the worst answer on the page. Delphi treats being mentioned and being recommended as two readings, so it never counts a named-and-warned-against answer as a recommendation, and a tool that counts position alone scores it as a win.

None of the five is a fault. They are design decisions and each one is defensible. What separates a score worth acting on from a number is whether the tool publishes which decisions it made, so the reading can be checked instead of believed. How to compare AI visibility tools sets out the questions to put to a vendor before you buy.

• Not a count of buyers. Nothing here counts how often a question was asked, so no figure on this page reports demand.
• Not revenue. The score measures a position in answers. It does not measure money that moved because of that position. Where you give a revenue band, Delphi estimates how much is decided in answers you do not win, as a range, and that estimate never enters the score.
• Not proof of cause. A gate that rises after work shipped is what happened afterwards. Competitors move, assistants change, and the evidence they read moves with them.
• Not a census. It is the questions in your set, on the day it ran, put to ChatGPT, Google Overview, Gemini, Copilot, Claude, Perplexity and Grok. It is a designed sample of buyer intent rather than every question anybody ever asked.
• Not stable underneath you. A model update can move a score with nothing changing on your side. Every stored answer records the model that produced it wherever the assistant names one, so a movement of that kind can be seen for what it is.
• Not personalised. What is measured is the answer to a fresh question with no earlier conversation. An answer shaped by one person's own past use is not reproducible and is not modelled here.