Show how unstable the AI answer really is

Hook: Roll one buyer question ten times, then hand finance the stability PDF. Counting that one question by hand costs EUR 135 to EUR 270 of analyst time.

You are asking for budget to be visible in AI answers, and leadership hears the ask and pictures a search ranking: one fixed list that holds until somebody changes it. The request dies in the meeting where a colleague opens ChatGPT on a phone, types your category question once, sees your brand in the answer and decides the whole problem is imaginary. One roll proves nothing in either direction. Ten rolls of the same question against the same model give ten different lists, and the shape of that variation is the argument. Your competitor holds the answer whatever the wording; you appear in six rolls and vanish from four. That gap is a number as soon as somebody counts it.

What it costs today

Assumption: an analyst at EUR 90 an hour. Ten fresh chats, the same question pasted ten times, ten answers saved, then a tally of who got named and where.

Job, by hand	Time	Cost at EUR 90/h
One question, ten rolls	1.5 to 3 h	EUR 135 to EUR 270
A ten-question set	15 to 30 h	EUR 1,350 to EUR 2,700

The method drifts as well. The next analyst counts sentence repetition differently, so this audit and the last one are two measurements wearing the same label.

That table prices the asking and the tally, which is the work a team actually does. The step it leaves out is the one that produces the number. Ten answers hold dozens of separate claims, and the question worth money is which of them the model repeats in every roll and which it says once. By hand you can only match wording, so "we are the cheapest" and "nobody undercuts us" land in two different rows. Rankfor.AI compares the claims by meaning, files the ones that say the same thing into a single row, then splits those rows by how many rolls each survived. No hourly rate sits against that line, because almost nobody books the day it would take.

Dice Roller sits on Strategist at EUR 496 a month. At EUR 270 a question the plan is covered by the second question (2 x 270 = 540). At EUR 135 a question it takes four (4 x 135 = 540).

What the video shows

1. Dice Roller opens under Tools on the Stability roll setting. One card holds the whole run: the question, your brand name, and a slider for how many times to ask.
2. The setup goes in. Type the buyer's question the way a customer would type it, add your brand name so every mention gets counted, drag the slider from 3 to 10, pick Search mode and run it.
3. Three cards open the results page: consistency score, semantic overlap, number of rolls, with the sentiment split across all ten answers underneath.
4. The brand panel sorts every name the model reached for into four bands by how many of the ten rolls named it: Always at 80% of rolls and above, Frequent from 50, Occasional from 20, Rare below that.
5. Switch that panel to Table and the whole list arrives with counts beside it. A rival sitting in Always against your own name sitting in Occasional is the slide.
6. Message Stability splits the sentences: the ones the model says every time, the ones it varies, the ones that appeared once. Lines are grouped by meaning, so a point made three ways lands in one row.
7. Responses holds all ten answers in full, nothing summarised away, with brand names highlighted in their band colour inside the text.
8. Sources lists every cited page, how many rolls it fed, and one click back to the answer it produced. This tab exists because the run used Search mode.
9. Export writes the PDF, and the closing frame puts it in front of finance.

What you need

• Strategist, EUR 496 a month, or Enterprise.
• A project with a completed scan. Without one the run button submits nothing.
• The model set to Gemini, ChatGPT or Grok before you start. Changing it mid-run clears the form.
• Search mode for the source list. Memory mode has no Sources tab at all.
• Time. Ten rolls is the slow end of the slider, so start the run the day before you need the answer.
• One Dice Roller run at a time per scan.

Questions people ask

How many rolls can I ask for? Ten. The slider stops there, so this measures repetition across ten answers and makes no claim beyond them.

Which AI is answering? Gemini, ChatGPT or Grok. One run asks one model, the one set on your project, so a three-model picture means three runs.

Is the consistency score statistically significant? No. Both scores on that page are weighted composites between 0 and 1, shown as percentages. The argument you take to the board is the band gap and the ten answers sitting underneath it.

Does it keep watching the question for me? No. Each run is a manual click. There is no schedule and no alert, so a monthly rhythm is something you put in your own calendar.

Roll the question your board would type, then put the four bands on one slide.