AI Prompt Tracking: How Many Prompts and Which Ones (2026)
AI prompt tracking: which questions to track in ChatGPT and other assistants, how many give a stable number (60 to 100), and how to avoid a flattering set.
- prompt tracking
- ai visibility
- measurement
- geo
AI prompt tracking only gives a usable number when the questions are ones buyers ask, never name your brand, and are numerous: Sistrix found 20 prompts measure to about ±9.5 points and roughly 70 are needed for ±5. Self-chosen sets also flatter: the own brand came out ahead in 63 to 80 percent of Sistrix's comparisons.
What goes wrong when you pick the questions yourself?
Questions you write yourself drift toward the corner of the market where you already win. The founder of a time-tracking app writes 15 prompts: her niche ("time tracking for design agencies that bill hourly"), her integration ("time tracker that syncs with Notion") and a few category questions. Her brand shows up in most answers. A wider set written from buyers' real questions shows her in a small fraction. (Invented example; the effect itself is measured below.) Nodpick, the tracker we build, is designed around this problem: it writes the question set from your category and keeps your brand name out of every scored question.
Sistrix compared brands' self-chosen prompt sets with a neutral base. Without a selection effect, the own set should come out ahead about half the time. Instead the own brand came out ahead "in drei von vier Fällen", 63 to 80 percent depending on the competitors compared (published 27 October 2026). Sistrix's verdict: "kein Messfehler, sondern ein Auswahleffekt", not a measurement error but a selection effect.
The bias has a second, quieter cost. In self-chosen sets, Sistrix found the own brand stable, while competitor figures from the same small sets are the ones to read with caution. You end up confident about the wrong number. Fixing it starts with size.
How many prompts should you track for AI visibility? Our test table
For a mention rate you can act on, plan for 60 to 100 different questions; 20 only give a rough picture. Sistrix measured a typical deviation of ±9.5 points with 20 prompts and ±4.2 with 100, and states that halving the uncertainty takes four times as many prompts. We used that square-root rule to fill in the sizes in between.
| Different questions | Typical deviation | 95 % band |
|---|---|---|
| 10 | ±13.4 points | ±26.9 points |
| 20 | ±9.5 | ±19.0 |
| 60 | ±5.5 | ±11.0 |
| 70 | ±5.1 | ±10.2 |
| 100 | ±4.2 | ±8.5 |
| 150 | ±3.5 | ±6.9 |
| 200 | ±3.0 | ±6.0 |
Method. Starting point: Sistrix's ±9.5 points at 20 prompts, scaled by the square root of 20 divided by the number of questions; the 95 % band is twice the typical deviation, as Sistrix defines it. Computed 31 October 2026. Check against the source: the table reproduces Sistrix's own published points: ±4.2 at 100 prompts, "rund 70 Prompts" for ±5 and "rund 200" for ±3. Limits. An extrapolation from one study's averages; your brand's spread depends on your market and engines.
Rerunning the same few questions does not close that gap (the post on why scores jump around shows the simulation). Sistrix found that 20 prompts never go below ±7.6 points however often they are repeated. So the next question is not "how often?" but "which ones?".
Which prompts should you track in AI search?
A good AI prompt tracking set covers the main ways buyers ask, in their words, without your brand name. Nodpick's method suggests questions across four angles: the best tool for the buyer's kind of company, the problem they want solved, alternatives and comparisons. A fifth angle, constraints such as location, price or team size, is often where local and niche brands win or lose.
| Angle | Example question (fictional brands) | What it tests |
|---|---|---|
| Category | "What is the best CRM for a 10-person B2B team?" | Are you in the default shortlist? |
| Problem | "How do I stop deals slipping through in a small sales team?" | Are you suggested before the buyer knows the category? |
| Alternatives | "What are good alternatives to Northwind CRM?" | Do you replace the market leader in answers? |
| Comparison | "Northwind CRM vs Contoso: which is better for a startup?" | Are you added when buyers compare others? |
| Constraint | "Which affordable CRM hosts its data in the EU?" | Do your real strengths decide the answer? |
Real buyers phrase the same need in very different ways. SparkToro measured an average semantic similarity of 0.081 across human-written prompts with the same intent (in a study run with Gumshoe.ai, an AI-tracking startup). Do not track five near-copies of one question; write distinct questions per angle. Google adds a reason for breadth from its side: AI Overviews and AI Mode "may use a 'query fan-out' technique - issuing multiple related searches across subtopics and data sources", so an answer is built from many sub-questions anyway (see query fan-out).
With a mix in place, one temptation remains: adding your own name.
Should any tracked question name your brand?
Not the questions that feed your score. A question that names you measures whether the assistant knows you, not whether it suggests you to a buyer who has never heard of you. Brand questions almost always "find" you, which inflates the rate.
Keep exactly one branded question outside the score, to check what assistants say about you: product, pricing, who you are for. Nodpick handles it that way: scored questions never name your brand, so every mention counted is one the engine made on its own, and one extra question about your brand by name feeds the "what AI knows about you" view, never the score. The prompt tracking glossary entry covers the basic definition, and the post on tracking brand mentions in ChatGPT shows the weekly setup for one engine.
Comparison questions that name competitors are fine, because they test whether you are added. The next risk is subtler: a set that is unbranded but still tilted.
How do you check your question set for flattery?
Compare your set against questions you did not write. Ask someone outside the team, ideally a salesperson or a recent customer, to write ten buyer questions. Run both sets on the same engines in the same week. If your mention rate on your own set is far above the outside set, your set is tilted toward you.
Three warning signs worth checking before the first run:
- Your differentiator in the question. "Time tracker with Notion sync" tests a feature only you advertise.
- Only bottom-funnel questions. Comparisons and alternatives without category and problem questions miss early buyers.
- Your language, not theirs. Product names for features, internal jargon, or the wrong country's spelling.
Then freeze the wording. A tracking set is only comparable week to week if the questions stay the same: Nodpick's change rule compares the same questions week to week, which only works when the questions do not move.
When should you change the question set?
Change questions in planned batches, not one by one. Every reworded question breaks its own history. Replace a batch when your market or product changes, mark the date, and read the trend from that date on.
Doing all this by hand is real work: a five-angle set of 60 questions, written in each market's language, run on six engines every week, compared question by question. The manual route in the tracking guide works for a handful of questions; at 60 questions on six engines it means reading 360 answers a week.
Nodpick writes the first set for you from your category, country and language across those angles, or takes the questions you write, keeps your brand name out of every scored question, and asks ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity and Claude every week in 7 languages for 16 European markets plus the US. Every score shows its 95 % range, so you see whether your set is big enough, and every answer that names you is sorted into recommended, just mentioned or negative. A question set built this way measures your market, not your own hopes. Before you pick a tracker for it, run it through the 10-question buyer checklist.
Check your own brand
See what a clean question set shows for your brand: the free AI visibility checker writes five unbranded buyer questions from your website, asks ChatGPT, Perplexity, Gemini and Claude, and shows whether they name you, who they name instead and which sources they cite. Five questions are an early sample, and the report says so. The checker opens on launch day; join the waitlist on that page to get one email when Nodpick opens. For the weekly version with 25, 70 or 220 questions, see the pricing page.
Common follow-up questions
How many prompts should a small brand track for AI visibility?
Start with 20 if that is all you can write well, and read the result as rough, about ±9.5 points by Sistrix's figures. Move toward 60 to 100 distinct buyer questions once you rely on the number for decisions. More different questions narrow the range faster than repeating the same ones.
Should you track AI prompts in several languages?
Yes, if you sell in several markets. An assistant told the buyer's country and language can name local providers, so an English-only set can miss how buyers in Germany or France see you. Track each market as its own set and compare markets only by trend.
Can you use Search Console queries as AI tracking prompts?
They are a good source of wording, especially queries that already read as questions, such as those starting with "best", "which" or "how". Rewrite them as a buyer would ask an assistant, in full sentences with context, and drop any that contain your brand name.
How often should I run my tracked prompts?
Weekly suits most brands, as long as the set stays fixed between runs. What breaks a trend is not the rhythm but reworded questions, because each question is compared with its own earlier answers. If you have budget for more runs, spend it on more different questions rather than more frequent ones.
Sources
- Sistrix, "Breite schlägt Wiederholung: Wie man AI-Sichtbarkeit verlässlich misst" (27 October 2026), https://www.sistrix.de/news/breite-schlaegt-wiederholung-wie-man-ai-sichtbarkeit-verlaesslich-misst/, Last checked: 1 November 2026
- SparkToro, "Why everyone prompts AI differently", https://sparktoro.com/blog/why-everyone-prompts-ai-differently-even-when-they-want-the-same-thing/, Last checked: 1 November 2026
- Google Search Central, AI features and your website, https://developers.google.com/search/docs/appearance/ai-features, Last checked: 1 November 2026
- Nodpick, How we measure AI visibility, https://nodpick.com/method, Last checked: 1 November 2026
- Nodpick, Pricing (plans and FAQ), https://nodpick.com/pricing, Last checked: 1 November 2026
- Precision table (extrapolated from Sistrix's published figures), computed 31 October 2026
Checked 1 November 2026.