Hook: One question rolled three times, split into what the model repeats and what it invents once. Doing that by hand costs EUR 45 to EUR 90 of analyst time per question.
What You'll Be Able To Do
- Roll the question you dread three times and read the claims that came back every roll apart from the ones that surfaced once
- Read sentiment on every mention of your brand, with the exact sentences behind each tag
- Open all three raw answers in full, each with the pages it cited underneath
- Export the run as PDF, Markdown or CSV, saved as
hallucination-report-<timestamp> - Keep an audit trail in History: the question, the persona, the mode, the number of rolls and the date
Runtime: 5:36 | Difficulty: Advanced
You own what the public knows about the company, and the models now answer questions about your pricing, your service and whatever went wrong two years ago, out loud, to anyone who asks. Some of those answers are wrong. Most comms teams find out which ones when a journalist reads one back to them on a call. The method that settles it is repetition: ask the same question several times, then compare the answers. What the model says every time is your public position now, whoever wrote it. What it says once is a soft claim, and the wrong ones live there.
What it costs today
Assumption: an analyst at EUR 90 an hour. One question means asking it N times, pasting each answer into a sheet, counting your mentions, and tagging sentiment on every mention.
| Job, by hand | Time | Cost at EUR 90/h |
|---|---|---|
| One question, three rolls | 30 to 60 min | EUR 45 to EUR 90 |
| One question, ten rolls | 1.5 to 3 h | EUR 135 to EUR 270 |
| Ten questions, three rolls each | 5 to 10 h | EUR 450 to EUR 900 |
The video films three rolls, so the like-for-like figure on screen is EUR 45 to EUR 90. Marketer is EUR 196 a month. At EUR 90 a question the third question clears the plan (3 x 90 = 270). At EUR 45 a question the fifth clears it (5 x 45 = 225).
What the video shows
- Persona Prompts opens with the questions each buyer asks, written by the scan. Pick the persona whose questions would hurt most if the answer came back wrong.
- Add prompt takes the question that keeps you awake, worded the way a customer would word it. It lands at the top of the list with a Custom badge.
- The run window sets the shape of the audit: three rolls, and Search mode so every roll is grounded in live results and keeps its sources.
- The wait. Behind the spinner the model answers the same question three separate times, and every other run button on the screen dims while this one is in flight.
- The Overview tab holds the split. Core Repetitive Messages came back on all three rolls, Variable Messages showed up on one and vanished on the next, Opportunity Gaps appeared once. Claims are grouped by meaning, so a point made three ways lands in one row.
- Brand Sentiment counts every mention of your brand across the three answers: the total, the range between the best and the worst roll, and the sentences behind the negative group.
- Raw Responses holds the full text of all three answers, unedited, each with its own citations block underneath. This is the part legal asks for.
- Sources lists every cited page, its title, how many rolls used it, and which ones. One click jumps back to the answer that page produced.
- Export writes the report as PDF, Markdown or CSV, and History logs the run with its question, persona, mode, roll count and date.
- The closing frame holds the saved PDF, open on the split between the repeated claims and the ones that flickered.
What you need
- Marketer, EUR 196 a month.
- A project with a finished scan, so the personas and their questions already exist.
- The model set to Gemini, ChatGPT or Grok before you open the run window. One run asks one model.
- Search mode if you want the source list. Memory mode answers from training data and returns no sources at all.
- One run at a time per scan. A second one is refused while the first is in flight.
- Any wording change made before you run, because editing a prompt gives it a new identity and the old result no longer attaches to it.
Questions people ask
Which AI does this ask? Gemini, ChatGPT and Grok. Those three, and one of them per run, taken from the model set on your project. A three-model picture means three runs.
Does it warn me when a new wrong answer shows up? No. Every run is a manual click. Rankfor.AI has no watcher and no schedule on this screen, so the audit happens when you run it, and a monthly rhythm is something you put in your own calendar.
Why three rolls and not one? One answer cannot repeat, so one answer produces no split at all. Three rolls give you the first honest read on which claims are stable. Ten rolls sharpen it and take longer to run.
Is the Confidence Score a statistical confidence level? No. It is a weighted score between 0 and 1, shown as a percentage. Read it as a summary of the run, then read the raw answers for the detail a comms director has to act on.
Run it on the one question you would hate to see in print, and read what comes back.