Finding good prompts is the part everyone does. Choosing how many of each kind to track, and leaving the list alone long enough to compare it with itself, is the part that decides whether the resulting number is worth reporting.
This is the companion to GEO prompt research, which covers how to discover prompts in the first place. This one picks up where that leaves off: you have a few hundred candidates, and now you have to build the map.
Key Takeaways
- The same brands score 95% on branded prompts and 7% on non-branded ones. Any score that blends the two groups is mostly a readout of your prompt mix, not your market position.
- Mix ratios are the single biggest lever on the headline number. Branded prompts should stay at 10β15% of the map, reported separately; 55β65% non-branded and 20β30% competitive is the working split.
- A frozen core set is what makes two months comparable. Rewriting a prompt creates a new prompt; the history does not carry over.
- One day's number is close to noise. Re-running the same prompt one minute later changes the cited sources about as much as waiting a day does.
- Seven effective observations per prompt, and a 10β28 day rolling window, is where the error bars get small enough to act on.

Branded prompts measure something else entirely
Start with the number that breaks most dashboards. In a tracking sample of 3,989 brands and roughly 23 million answers over the 30 days to 2026-10-15, the median brand mention rate was 95% on prompts that contain the brand name and 7% on prompts that do not. In the same dataset, 98% of brands scored higher on branded prompts. (This is vendor-reported data from Searchable and cannot be independently verified, but three separate sources point the same direction, and the direction is what matters here.)
Read that mechanically. Add one more branded prompt to your map and the average goes up. Nothing about your actual standing has changed. You have simply added a question you were always going to win.
Carlos Barraza's worked example makes the arithmetic concrete. Take a brand mentioned in 95% of branded answers and 10% of category answers, in a single week:
Prompt list | Reported "visibility" |
|---|---|
10 branded + 10 category | 52.5% |
18 category + 2 branded | 19% |
20 category, no branded | 10% |
Same brand. Same week. Same engines. Three numbers, and the only thing that changed was the list.
So branded prompts do belong in the map β they are the only way to see whether the model describes you accurately, mixes you up with someone else, or repeats a complaint thread. But what they measure is accuracy and sentiment, not visibility. They go in their own group with their own reporting line, and they never get averaged into a single "AI visibility" figure. A report that blends them is not comparable to anyone else's, including your own from last quarter.
The mix: what actually goes in the map
Two published recommendations exist, and they disagree less than they appear to.

HubSpot suggests roughly 50β60% non-branded, 20β30% competitive, 10β20% branded. Conductor suggests 75/25 with branded capped at a quarter. The difference is bookkeeping: Conductor folds competitive prompts into the non-branded bucket. Converted to one caliber, both land on the same rule β branded stays under a quarter of the map.
Our working split is 55β65% non-branded, 20β30% competitive, 10β15% branded. For commerce brands, push competitive down toward 20% and move transactional intent into the non-branded side, because product-card competition happens on category and use-case questions rather than on "A vs B".
Each bucket answers a different question:
- Non-branded β when a stranger describes their problem, are you in the answer at all? This is the growth surface.
- Competitive β when you are already in the consideration set, are you treated as a credible option? This drives the "alternatives to X" answers.
- Branded β does the model get your facts right, and is the tone neutral or worse?
Six dimensions that must not have an empty row
Correct ratios still leave you free to put the entire map on one intent and one persona. The second test is coverage:
- Topic clusters. Start with 3β5 priority topics at 5β10 prompts each; enterprise maps run 10β30 topics. Collapse near-synonyms into one or two representative sentences rather than padding a topic with twenty rewordings.
- Intent. At least one prompt per intent type, with recommendation and comparison intent taking the majority of the non-branded half. Definitional prompts deserve a small share β they rarely produce commercial outcomes.
- Persona and constraints. The same question with a budget, a company size, a region or a use case attached. Constraints change which brands get named, which makes them a variable you have to cover rather than an optional flourish.
- Market and language. Full coverage in primary markets, a reduced subset elsewhere. The same question in another language pulls a different set of sources; translated conclusions do not transfer.
- Engine. Run the whole map on each engine you monitor, but set the baseline separately for each one. A shared threshold across engines is not a comparison.
- Bound asset. Each intent cluster names the page that should be getting cited. If you cannot name one, you have found a content gap, not a monitoring item.
Freeze the core set
A map that changes while you are measuring cannot tell you anything. The discipline is a split:
Stage | Prompts | Topics | Frozen / exploratory | Change rule |
|---|---|---|---|---|
Starting (0β1 month) | 25β50 | 3β5 | 100% frozen | Collect only; change nothing |
Growing (2β6 months) | 150β300 | 5β12 | 70 / 30 | β€20% change per quarter, added in dated batches |
Enterprise / multi-market | 500+ | 10β30 | 70 / 30 per market | Separate frozen set per market; no merged cross-market reporting |
Three rules make it work:
- The frozen set does not change for a quarter β not even punctuation. A rewritten prompt is a new prompt in the data; its history does not continue, and your chart needs a break marker.
- The exploratory set absorbs experiments. New phrasings, new segments and new markets live here until two consecutive months show they have both volume and intent.
- Retire anything that has not driven a decision in three months. It is not monitoring at that point, it is noise.
On wording: AI search questions run 70β80 words on average against 3β4 for a Google query. A map full of three-word phrases is measuring something other than the conversation you care about. Write whole questions, put the constraints inside the sentence, keep one intent per prompt, and never add a bare follow-up like "which is cheaper?" β out of context it tests nothing.
How many runs before a number means anything
This is the part with real evidence behind it, and the part most often skipped.
A University of St. Gallen study ran the same prompts daily for 45β46 days and, separately, re-ran each prompt up to ten times inside a single day. That second design is what makes it useful: it separates the model's own randomness from genuine movement in the answers.

The result is blunt. Cited-source overlap between consecutive days runs 0.34β0.42 (Jaccard) β about 65% of cited sources turn over daily. But re-running the same prompt within the same day gives 0.32β0.43. Statistically indistinguishable. Most of what a daily dashboard shows as movement is the model resampling, not your position changing.
Four operating rules follow:
- At least seven effective observations per prompt. Bootstrap convergence puts the standard error of brand mention rate under 0.10 at n=7 (95% CI Β±0.158) and under 0.08 at n=8. Cited-source coverage needs nβ₯8.
- Report on a 10β28 day rolling window. Standard error falls below 0.10 at d=10 days and below 0.05 at d=24. If you run each prompt once per engine per day, the rolling window is your sample β a daily figure is an alert, not a trend, and even a week is still noisy at brand level.
- Brand level for KPIs, URL level for diagnosis. Brand-set overlap (0.45β0.59) is markedly steadier than cited-source overlap (0.34β0.42). Watching an individual URL rise and fall is mostly reading noise; use it for attribution, not reporting.
- Set the baseline per engine. Citation concentration (Gini) averaged 0.715 but ranged from 0.671 to 0.782 across the engines tested. Where a few domains absorb most citations, 20% visibility is a much harder number than the same 20% on a flatter engine.
The study tested ChatGPT, Gemini, Google AI Mode and Perplexity in the Swiss-German market between 2026-01-24 and 2026-03-25. GEOly's own monitoring covers ChatGPT, Google AI Mode, Google AI Overviews and Gemini; the engine list differs, but the statistical behaviour is a property of how these systems sample, not of which four you pick.
One more finding worth carrying: 57.8% of ChatGPT runs in that study returned no citations at all, because definitional questions do not trigger retrieval. "No citation" and "not mentioned" are different results and need different columns.
Two caliber errors that quietly move the number
- Do not average daily rates. Each day has a different record count, so the daily rate is already weighted. Read the totals row, or re-weight by record count. A straight mean gives small days the same vote as large ones.
- A missing day is not a zero. It means nothing was collected, not that nobody mentioned you. Counting it as zero drags the window average down and invents a decline.
What has to travel with the number
Any AI visibility figure that leaves your team should carry three things, or it cannot be compared with anything:
- The composition β how many non-branded, competitive and branded prompts, and whether the map changed this period.
- The split scores β non-branded visibility and branded accuracy reported separately, never blended.
- The caliber β engines, market, rolling-window length, and effective observations per prompt.
If you want the measurement side defined precisely, the AI visibility metrics reference has the formulas and denominators, and Cited shows which domains are actually being used as sources. The full eight-stage method this article draws on β sourcing, fan-out stress testing, scoring and the feedback loop back into new segments β is written up on the Prompt Research Loop page.
FAQ
How many prompts do I need to start? Twenty-five to fifty, across three to five topics, fully frozen for the first month. The goal in month one is not insight, it is accumulating enough window length for the first comparison to be valid.
Should I track branded prompts at all? Yes β they are the only way to catch the model describing you wrongly or negatively. Just report them as their own group. The failure mode is blending them into a headline visibility score.
Why can't I compare this week to last week? You can, if the map did not change and you are comparing rolling windows rather than single days. If you added or rewrote prompts, the two numbers are measuring different things and the chart needs a break.
What if a competitor's score jumps? Check their prompt-set exposure before you check their content. Within a single map run under one caliber the comparison is fair; across two different tools with different prompt lists it is not.
Does this apply per engine or in aggregate? Set baselines per engine. Citation concentration and citation frequency vary enough between engines that one threshold applied to all of the ones you monitor produces false comparisons.
Written by GEOly AI. More on GEO and AI Visibility, or start with GEO prompt research if you are still building the candidate list.
Sources. Branded vs non-branded mention rates: Searchable tracking data, 3,989 brands, ~23M answers, 30 days to 2026-10-15 (vendor data, not independently verifiable). Mix ratios: HubSpot, "Which AI prompts should you track?"; Conductor Academy, "AI Prompt Tracking". Worked example: Carlos Barraza, "Your AI Visibility Score Is Only as Good as Your Prompt List" (illustrative figures). Measurement stability: Schulte, Bleeker & Kaufmann, "Don't Measure Once: Measuring Visibility in AI Search (GEO)", arXiv:2604.07585, 2026-04-10. Question length: SimilarWeb, via iPullRank.

