The CMO asked how ChatGPT describes us, here is how to answer properly

Answer it with three things on one page: the prompt set you asked, the sample of answers you got, and the caveat that says what the sample cannot prove. A single chat screenshot is not an answer, it is one draw from a non-deterministic system, and most executives will sense that quickly. The defensible version reads like this: "Across 60 buying-intent prompts run on two providers in the week of the 14th, we were named in 38 percent of answers, Northgate in 61 percent. Sentiment where we appear is mostly neutral. Week-over-week movement may reflect a provider model update rather than a real change, so treat one week as a reading, not a trend."

That is the whole shape. One metric with its denominator, one comparison, one sentiment note, one caveat. It takes an afternoon to build by hand the first time, and it survives the follow-up questions, which is the only test that matters. Below is how to build each piece, what to say when the CMO pushes on it, and where the honest gaps are.

Why the screenshot fails

Someone on the team opens ChatGPT, types "best project management software for mid-market," screenshots the answer, and pastes it into Slack. The brand is missing. Panic follows, or relief follows if the brand is there. Both reactions are unearned.

Three problems, in order of how fast they get exposed in a meeting:

• Sample size of one. LLM answers are non-deterministic. Ask the same question twice and you can get different lists. A single response tells you the answer was possible, not that it is typical.
• No denominator. "We are missing from ChatGPT" is a claim about a population. The screenshot is a claim about one prompt. The obvious next question is "out of how many," and there is no answer.
• No time anchor. Models get updated. An answer from March and an answer from May came from different systems. Without a date and a model version, you cannot separate a competitor's content push from a silent model change.

The screenshot is still useful. It is the illustration, not the evidence. Keep it on page two.

Step one, fix the prompt set before you look at anything

The biggest credibility risk in this work is choosing prompts after you have seen the answers. If you write the prompt list, run it, then quietly drop the four prompts where you look worst, you have built a marketing asset rather than a measurement. It reads as one in the room.

So write the list first, freeze it, and keep it. Aim for enough prompts that one strange answer cannot move the number much. Somewhere between 50 and 100 is a reasonable working range, which is also the range Seenby is built to auto-generate during onboarding from your brand name, description and up to three competitor names, with a curated fallback set if generation returns too few prompts (mvp:1).

Build the list across four intent types, so the number means something specific:

1. Category prompts. "Best CRM for a 200-person company." No brand named. This is where discovery happens and where absence hurts most.
2. Comparison prompts. "Northgate vs Aria for mid-market teams." You want to know whether you are in the comparison at all.
3. Attribute prompts. "Which analytics tools have SOC 2 and SSO." Absence here is usually a content gap you can work on.
4. Branded prompts. "What does Aria do." Tests whether the description is right, not whether it exists.

Write the intent type next to each prompt. When the CMO asks what kind of questions these were, you point at the column instead of improvising.

Step two, take a sample, not a look

A sample has four properties you must be able to state: how many prompts, which providers, which dates, and which model versions. Miss any of those and the number is not reproducible, including by you next month.

Providers matter more than people expect. ChatGPT and Perplexity behave differently because one is answering largely from a model and the other leans on retrieval plus a model. A brand that is faint in one can be prominent in the other, and the mix of providers you sampled changes the headline. Seenby queries exactly two providers at MVP, OpenAI and Perplexity, behind a provider adapter interface that degrades gracefully if one provider is disabled (mvp:2). State which ones you used. If you ran only ChatGPT, write "ChatGPT only," not "AI search."

Model versions matter because they explain drift. Tag every response with the model version at the time you collected it, which is what Seenby is built to do on every stored response (mvp:9). This is the difference between "our share dropped nine points" and "our share dropped nine points in the same week the underlying model changed, so we are holding the conclusion until next week."

Doing this by hand means a spreadsheet: one row per prompt per provider per run, with the full response text, the date, the model version, and then your reading of it. It is tedious and it works. If you are storing response text anywhere shared, strip personal names before it becomes searchable, because LLM answers occasionally name executives, employees or customers. Seenby runs a PII-scrubbing step in the response-storage pipeline before any response is queryable, and flags every stored response with whether it ran (mvp:5).

Step three, turn answers into three numbers

Read each response and record three things about it. Not ten. Three.

• Named or absent. Did the brand appear at all. This is the base of your share metric.
• Position. Where in the answer. First in a list of five is not the same as a parenthetical at the end.
• Sentiment. Positive, neutral, negative, or absent. Most mentions are neutral and that is fine. Negative mentions are rarer and worth reading in full.

Seenby's response parser extracts brand and competitor mentions, classifies sentiment on that four-value scale, and records ranking position within each response (mvp:3). By hand, you are doing the same three passes with your own eyes, which is workable at 60 prompts and unpleasant at 600.

From those three columns you get one headline number: the share of sampled answers that named you. Seenby computes this as Share of AI Voice, a single headline metric per brand per week, stored with its methodology caveat text attached to every computed record so the number does not travel without it (mvp:4, mvp:8). If you are building the page yourself, replicate that discipline manually. The caveat goes in the same visual block as the number, not in a footnote nobody reads.

Then add the comparison. A share figure with no benchmark invites the question "is 38 percent good," which has no answer on its own. A share figure next to two or three named competitors answers it structurally: you are behind Northgate, ahead of Vellum, and the gap to Northgate is 23 points. Seenby's dashboard is built to carry 30-day trend lines for Share of AI Voice, the sentiment breakdown, the top prompts where the brand is and is not mentioned, and a competitor benchmarking view in one place (mvp:7).

One warning on charts you build yourself. If you draw share as bars, do not scale the bars so the leader always fills the track. That makes a 4 percent share look like most of something, and the first person to notice it will distrust the whole page. Scale to 100 percent of the sample, always.

Step four, write the caveat before you write the summary

This is the step most teams skip, and it is the one that makes the page defensible. Write the caveat first, in plain language, and let it constrain what you claim above it.

A caveat that works has four clauses:

• What the sample was. "60 prompts, two providers, one week."
• What non-determinism does. "Repeat runs of the same prompt can return different answers."
• What model updates do. "Week-over-week movement may reflect a provider model update rather than a change in the market."
• What the number is not. "This measures how often we are named in sampled answers. It does not measure traffic, pipeline or revenue."

That last clause is the one that protects you six months out. The moment a presence metric is read as a revenue metric, someone will ask it to explain a bad quarter. Say what it is early and repeat it.

Keep the caveat wording identical every time you report. Changing it reads as hedging. Seenby takes the same approach, surfacing the methodology disclosure on every score display and on a dedicated methodology page, so the caveat is the same sentence everywhere it appears (mvp:8).

The one-page brief, in order

Here is the layout that holds up under questions. One page, this sequence.

1. The headline. "We were named in 38 percent of 60 sampled AI answers in the week of the 14th. Northgate: 61 percent."
2. The caveat, directly under it, same block, four clauses.
3. Method in three lines. Prompt count and intent mix, providers, dates and model versions.
4. Sentiment. "Of our mentions, 8 percent positive, 89 percent neutral, 3 percent negative." Quote the negative mention in full if there is one.
5. Where we are absent. The five to ten prompts with no mention, verbatim. This is the most actionable part of the page and usually the part the CMO reads twice.
6. One illustration. Yes, the screenshot. Labelled with prompt, provider and date.
7. What we are doing next, and what evidence would tell you it worked.

That is it. No appendix, no thirty-slide deck. If your CMO wants the raw responses, have them in a linked sheet and say so.

The four follow-up questions, and what to say

"Is this statistically significant?" No, and do not claim it is. Say what you have: 60 prompts is enough to see a large gap and not enough to call a two-point move. The movements worth acting on are the ones that persist across consecutive weeks and survive a model-version check.

"Why did it change?" Check the model version first, the competitor's visible activity second, your own content changes third. If the model version changed in the same window, say the drift is unattributable this week. Naming that uncertainty buys you credibility for the weeks when you do have an explanation. Seenby's weekly email digest is built to carry top insights, drift alerts tied to model-version changes when applicable, and the methodology caveat in the email itself (mvp:6).

"Can we just fix it?" Partly. You can publish comparison and attribute content, correct third-party descriptions of your product, and get your category language into the sources these systems draw on. You cannot buy a slot, and no one can guarantee an LLM will name you. Treat any promised ranking as a claim about a system the seller does not control.

"How often will you report this?" Weekly is the natural rhythm because it is short enough to catch drift and long enough to average out some noise. Monthly rollups for the leadership meeting, weekly readings for the team.

Where this is going

The reason to build the brief properly the first time is that you will be asked again. Once a CMO has seen a number with a denominator and a caveat, the screenshot era is over, and the standing question becomes "where are we this week." That is a reporting cadence, not a one-off.

Seenby is built for that cadence: a frozen prompt set, scheduled runs against two providers, a single weekly Share of AI Voice score with its methodology caveat attached, competitor benchmarking, and a weekly digest that says what moved and flags when a model update could explain it. It is in waitlist today, not launched, so the honest position is that the method above works with a spreadsheet now, and the tooling is built to take the spreadsheet off your hands when it launches. Either way, build the page in the order above. The caveat is not a weakness in the argument, it is the reason the argument holds.