The Quellan Method.

Every CMO asks the same question in some form: when someone asks ChatGPT about my category, am I in the answer?

Nobody has a shared way to answer it. The result is that the conversation runs on anecdote. Someone at the agency runs one prompt, shows a screenshot, everyone nods. A competitor runs a different prompt, gets a different answer, takes a different screenshot, draws the opposite conclusion. Without a method, you can argue both sides of the same question forever.

This document is the method we use at Quellan. It exists so that when we say a brand is visible or invisible in AI answers, the claim is inspectable. It is the protocol behind every Quellan analysis. We are putting it in the open because a method that cannot be examined is not one.

Search results come back the same way every time for the same query. Ten blue links, sometimes with a featured snippet on top. AI answers are generated in the moment. Ask the same question twice, you get two variations. Ask it through Claude and through ChatGPT, you get two different frames. Ask it with the word "best" versus "recommended," you get two different shortlists.

That means two things. First, looking once and screenshotting it proves nothing. Second, looking a thousand times without structure proves little more. To say anything reliable about how a brand surfaces in AI answers, you need a protocol: a repeated, defined, comparable procedure. Otherwise you are collecting vibes.

Five prompts. Five models. Three runs each. Seventy-five observations per brand. From those, the four dimensions are scored.

The prompts. Five prompts per brand, designed to probe different points in the customer journey. The set catches both the moment someone already knows the brand and the moment they do not.

The models. Five models, current as of the publication date of each analysis. As of April 2026 that is ChatGPT, Perplexity, Gemini, Claude, and Copilot. The set changes as the model landscape changes. Each model is queried through its free-tier web interface, not through paid API endpoints, because that is what the majority of users actually experience when they ask an AI system a question. Each model is queried without personalization, in a fresh session, with no context carrying across prompts.

The runs. Each prompt is run three times per model to smooth sampling noise. Results are aggregated. We look at what the model says on average, not on its best day.

What we do not do. We do not query models through their API with system prompts designed to extract favourable results. We do not use jailbreaks, role-play instructions, or leading phrasing. The goal is to replicate what a normal user gets when they ask an AI system a normal question. If the method is not reproducible by someone with an account and five minutes, it is not measuring what we claim it measures.

Worth saying out loud, because methods that do not name their own limits are hiding something.

The method does not predict revenue impact. We measure visibility in AI answers. We do not claim, on present evidence, that a ten-point score improvement produces a particular increase in purchase. That link is the next piece of research. We will publish it when the data supports a claim.

The method does not stay static. Models change, prompts that worked in April do not work in October, new models enter the top tier. Every Quellan analysis is dated, and we re-benchmark the prompt set twice a year. A score from March 2026 is not directly comparable to a score from March 2027 without adjustment, and we flag that in every report.

The four dimensions describe what is being measured. The five metrics are the specific readings we publish from them. Every Reading cites at least one. Each sits downstream of the protocol above, so anyone with an account on the five systems can reproduce the number.

How we read it. One brand-specific query, one comparison query, one discovery query. Three runs across the five systems, forty-five observations. For each observation, the descriptor the model returns is compared to the brand's own current positioning. A match scores 1.0, a partial overlap 0.5, a mismatch or absence 0.0. Representation Gap is one minus the mean score, reported as a decimal between 0.00 and 1.00.

Worked example, Oatly. The brand's current thesis is a functional-drinks and dairy-replacement argument. AI returns "oat milk" in forty-two of forty-five observations. Score 0.07. Representation Gap = 0.93. Present, mis-framed, at the altitude that counts.

How we read it. One event-linked query ("what is the latest with [brand]?") and one status query ("who is the current [role] at [brand]?"). Run weekly from the date of the event until at least three of five models reflect the change on two of three runs. Signal Lag is the days elapsed to first week of threshold, reported in weeks. If twelve weeks pass without the threshold, the reading is logged as indefinite.

Worked example, Loewe. Jonathan Anderson's Loewe exit was announced 17 March 2025, and he was confirmed at Dior Men in the same window. As of April 2026, five of five models still describe Loewe as under Anderson. Fifty-six weeks, threshold not met. Signal Lag = indefinite pending refresh.

How we read it. One mention query ("what are the major [category] brands?") and one recommendation query ("I want to buy [category product], what should I get?"). Three runs per query across the five systems, sixty observations. Presence Rate is the share of mention-query observations where the brand appears. Preference Rate is the share of recommendation-query observations where the brand lands in the top three. Preference Gap is the gap in percentage points.

Worked example, Nike. Presence Rate 91 percent, named in nearly every athletic-footwear query. Preference Rate 28 percent, top-three in roughly one in four. Preference Gap = 63pp. Seen, not chosen.

How we read it. One category-leadership query, rephrased five ways ("best," "top," "most respected," "leading," "iconic"), run across the five systems. Twenty-five observations. Compression is the share of observations where a single brand is named first. 1.0 is full compression, one winner every time. 0.2 is fully fragmented.

Worked example, niche perfume. Le Labo named first in twenty of twenty-five observations, Aesop three, Diptyque two. Compression = 0.80 to Le Labo. The category has a consensus answer before the reader asks.

How we read it. Not prompt-based. Each reading is timestamped on publish. We then track the first trade-press piece that covers the same move in the same frame over the following ninety days. Lead Time is the days from our reading to that citation. If trade press beats us to the frame, the reading is disqualified from the metric. Baseline publishing once ten readings close.

Note. Lead Time is the only metric in the family that a reader cannot reproduce directly, because it requires the Quellan publish log. Each measurement is timestamped and the cited trade pieces are dated, so the individual reading is inspectable even when the full protocol is not.