# Prompt Tracking: How Many Prompts Are Enough? > Most prompt tracking programs report a number their sample cannot support. Here is the run math, per engine, and what each budget can honestly prove. Canonical URL: https://cite.solutions/blog/prompt-tracking-how-many-prompts Source: Cite Solutions (cite.solutions) Published: 2026-08-06 --- [AI Visibility](/category/ai-visibility)12 min read # Prompt Tracking: How Many Prompts Are Enough? [Subia PeerzadaFounder, Cite Solutions · August 6, 2026](https://www.linkedin.com/in/subia-peerzada-75025764/) Key takeaways ## Key takeaways for AEO optimization Treat AEO as a measurement system, not a one-off publishing sprint. 1. 01Track prompt clusters that sit close to revenue, not vanity questions. 2. 02Compare your brand against competitors by source type, recommendation presence, and page type, not just mentions. 3. 03Turn every gap into a concrete content, PR, or technical fix with a weekly review cadence. Every prompt tracking setup starts with the same question and answers it with a guess. How many prompts should we track? Someone says 25\. Someone else says 100\. The set gets built, the dashboard turns on, and nobody asks the second question, which is the one that decides whether the dashboard means anything. The second question is how many times each prompt gets run. That is where most programs lose the plot, because the unit they budget is not the unit that carries the evidence. ## How many prompts do you need for prompt tracking? Track 25 to 40 prompts per topic, then run each one repeatedly rather than once. Published convergence research puts stable rankings between 33 and 94 collected answers per topic-engine pair. A 25-prompt set sampled once a week reaches that band in two to three weeks. Sampled once a month, it never does. The prompt-run budget The same prompt set buys 21% less evidence on Gemini than on Google AI Mode, before anyone touches your content A run only becomes evidence when the answer carries citations. Engines cite at different rates, so the number of runs needed to reach the same evidentiary weight differs by engine. 42 vs 51 runs needed to collect 40 cited answers on Google AI Mode versus Gemini A tracker that runs every engine the same number of times is under-sampling the one that cites least. Google AI Mode · cites in 97.4% of answers 42 runs about 5 sources per cited answercheapest engine to sample ChatGPT (web search) · cites in 92.5% of answers 44 runs about 4.6 sources per cited answer2 extra runs to match Google AI Mode Gemini · cites in 79.1% of answers 51 runs about 5 sources per cited answer9 extra runs to match Google AI Mode What each run count can actually prove, at a 20% mention rate 50 runs±11.1ppCannot separate 20% from 30% 100 runs±7.8ppDirection only, no magnitude 200 runs±5.5ppLarge moves visible 400 runs±3.9ppQuarter-over-quarter reporting 900 runs±2.6ppSmall lifts become defensible Sources: engine citation rates from The CITE Index, Cite Solutions, 90,132 answers, 500 prompts, 10 categories, May 19 to July 21 2026\. Convergence band from Sielinski, arXiv 2607.10341\. Run counts and margins of error are our arithmetic. Corpus covers Indian consumer categories, not AI Overviews. ## Prompt tracking counts prompts and pays for runs A prompt is a question you decided to ask. A run is one answer an engine actually produced. Only the second one is data. This distinction sounds pedantic until you look at what happens to a report built on the first one. ### A weekly refresh gives you one draw, not a trend Most tools run each tracked prompt once per refresh cycle. If your brand appeared last Tuesday and vanished this Tuesday, you have two coin flips, not a decline. Generative engines are stochastic by design. The same prompt on two occasions produces different text citing different sources, with nothing in the world having changed between them. ### The convergence band is published, and one run per week is nowhere near it On July 11, 2026, Ronald Sielinski published [a convergence framework for AI visibility measurement](https://arxiv.org/abs/2607.10341) testing exactly this. Across 30 platform-topic combinations on Gemini, SearchGPT, and Perplexity, at 125 queries per pair, rank stability fired between 33 and 75 responses. Adding a precision test raised the requirement to between 33 and 94. Three of the 30 combinations never converged at all inside the collection window. Median convergence landed at 42 responses on Gemini, 37 on SearchGPT, and 51 on Perplexity. Those are per topic, per engine, not per program. ### Keyword tracking and prompt tracking ask different questions The habits carried over from rank tracking are what break prompt tracking, because rank tracking had no sampling problem to solve. **Keyword tracking asks:** * •What position did this URL hold today? * •Did the position change since yesterday? * •How many keywords are in the top ten? * •Which competitor outranked us? **Prompt tracking asks:** * •How often were we named across repeated runs of this prompt? * •Is that rate different from last period, or inside the noise band? * •How many runs produced the rate we are reporting? * •Which sources did the engine read to build the answer? > A rank is an observation. A mention rate is an estimate. Estimates come with error bars or they come with nothing. ## 5 reasons your prompt tracking number moved without your work moving Each of these produces a change on the dashboard that has no cause inside your marketing. Rule them out before you open a root-cause investigation. ### Reason #1: You resampled a probabilistic surface The most common one, and the least investigated. Two single runs a week apart differ because generation is not deterministic. At 200 runs and a 20% mention rate, your 95% confidence interval is still roughly ±5.5 points. At 50 runs it is ±11.1 points, which means a reported move from 20% to 30% is inside the error of the instrument. ### Reason #2: Your prompt set drifted while you were not looking A set that gains five prompts a month produces a trend line partly made of the additions. New prompts enter at whatever rate they enter at, and the aggregate moves. Freeze the set for at least 90 days, log the reason each prompt is in it, and version the list the way you would version a schema. ### Reason #3: You mixed intent types into one average Conductor ran [14,000 API calls across 10 industries, seven intent types, four models, and five personas](https://www.conductor.com/academy/ai-recommendation-consistency-analysis/) and found consistency is predicted by query intent rather than industry size. Brand overlap between runs ranged from 40% on purchase-intent prompts to 63% on comparison prompts. A set weighted toward purchase intent is structurally noisier than one weighted toward comparison. Blend them and you get an average whose variance nobody can account for. ### Reason #4: You compared engines that cite at different rates Two engines given the same number of runs do not return the same amount of usable data, because they do not cite at the same rate. The engine that cites least produces the noisiest line on your chart, and it will look like the engine where your visibility is least stable. It is not. It is the engine you sampled least effectively. The next section works the arithmetic on our own corpus. ### Reason #5: Your category simply moves that much Some categories churn and some do not. Without a category baseline, every arrow on the dashboard looks equally meaningful. We worked through the baseline problem in detail in [what an AI overview tracker misses](/blog/ai-overview-tracker-what-it-misses), and the churn ranges there are wider than most teams expect. > Before you explain a movement, establish that there was one. ### Find out whether last quarter's movement was signal or resampling We rebuild your prompt set, size the run budget per engine, and report a change threshold calibrated to your category instead of a weekly arrow. First findings inside 14 days. [Book a Discovery Call](/contact) ## What 90,132 answers say about the run budget each engine costs you A run only becomes evidence when the answer carries citations. An uncited answer tells you the engine responded, not which sources it trusted. Engines differ sharply on that, and the difference has a direct price in runs. ### Gemini returns a citable answer four times out of five Our [CITE Index study](/state-of-ai-india/final-report) ran 500 unaided buyer prompts through ChatGPT, Gemini, and Google AI Mode every night for 63 days between May 19 and July 21, 2026, collecting 90,132 answers across 10 consumer categories. Across the whole corpus, 89.6% of answers carried at least one citation. The per-engine split is where the budgeting problem lives: Google AI Mode cited sources in 97.4% of its answers, ChatGPT with web search in 92.5%, and Gemini in only 79.1%. All three attached about five sources when they cited at all. The full set of figures sits on our [AI search statistics](/ai-search-statistics) page. ### The engine you track changes the budget, not just the coverage Work the arithmetic forward. To collect 40 cited answers, which is the low end of the published convergence band, you need 42 runs on Google AI Mode, 44 on ChatGPT, and 51 on Gemini. | Engine | Cited-answer rate | Runs for 40 cited answers | Budget consequence | | -------------------- | ----------------- | ------------------------- | ---------------------------------------- | | Google AI Mode | 97.4% | 42 | Cheapest engine to sample to convergence | | ChatGPT (web search) | 92.5% | 44 | Roughly parity with AI Mode | | Gemini | 79.1% | 51 | 21% more runs for the same evidence | A tracker that runs every engine an equal number of times is not treating them equally. It is under-sampling the one that cites least, then reporting all three on the same axis. > Equal runs across engines is not a fair test. It is three tests of unequal strength on one chart. ### Your own domain is a small share of what the answer is built from Across the same corpus, reddit.com drew 14,698 citations, more than any other source, and appeared in 13.6% of all 90,132 answers. Four of the twelve most-cited domains were brand-owned sites. If your prompt tracking only records whether you were mentioned, you are discarding most of the usable output. The cited-source list is the part that tells you where to work next. ## Step 1: Decide what decision the number has to support Write down the decision before the budget. "Is our mention rate up quarter over quarter" and "did the pricing page rewrite work" need very different sample sizes, and only one of them is answerable on a small set. A directional read needs far fewer runs than a causal claim about a specific change. ## Step 2: Group prompts into topics and treat the topic as the unit Convergence research reports thresholds per topic-engine pair, not per program. Twenty-five prompts spread across five topics gives you five prompts per topic, which will not converge on any engine. Build 25 to 40 prompts inside each topic you actually compete in, then report at topic level. Our guide to [selecting prompts for LLM tracking](/blog/how-to-select-prompts-for-llm-tracking) covers how to source and filter the list itself. ## Step 3: Size the run budget per engine, not per program Set a target of at least 40 cited answers per topic-engine pair, then divide by that engine's cited-answer rate to get the run count. Gemini needs roughly a fifth more runs than Google AI Mode to land in the same place. Record the run count alongside every number you report. A mention rate without a denominator is not a measurement. | Runs per topic-engine, per period | Margin of error at a 20% mention rate | What you can honestly claim | | --------------------------------- | ------------------------------------- | ---------------------------------------- | | 50 | ±11.1pp | Presence or absence, nothing about trend | | 100 | ±7.8pp | Direction, if the move is large | | 200 | ±5.5pp | Meaningful quarter-over-quarter change | | 400 | ±3.9pp | Board-grade reporting on a topic | | 900 | ±2.6pp | Small lifts become defensible | ## Step 4: Run a no-change period before you judge anything Run the frozen set on schedule for four to six weeks with no content or off-page changes at all. Record the spread. That spread is your noise band, and it is category-specific. Anything inside the band is not news. Anything outside it earns a root-cause pass. Setting the threshold after you see the number is not measurement, it is explanation. ## Step 5: Report cited-source share alongside your mention rate Log every cited domain on every run, not only whether you appeared. That list is your off-page target map and your early warning when a model swap changes the grounding pool. Two lines beat one: how often you were named, and how the cited-source pool turned over. The second one moves first. We covered the weekly churn pattern in [citation drift](/blog/citation-drift-why-your-ai-visibility-changes-weekly), and the weighting question in [measuring share of voice in AI search](/blog/share-of-voice-ai-search-measurement). If the run budget this implies is beyond what your current tool or team can carry, that is a real constraint and worth naming out loud. It is also the point at which [a managed GEO agency](/geo-agency) becomes cheaper than a seat license nobody has time to configure. ## FAQ ### How many prompts do you need to track AI visibility? Twenty-five to 40 prompts per topic, not per program. Convergence research puts stable rankings between 33 and 94 collected answers per topic-engine pair, so the prompt count only matters in combination with how often each prompt is run. Five prompts per topic will not converge no matter how long you track them. ### How many prompts should you use to test AI visibility? For a before-and-after test on a specific change, budget by runs rather than prompts. Detecting a lift from a 20% mention rate to 30% at conventional confidence takes roughly 290 runs per period per engine. Detecting 20% to 25% takes closer to 1,100\. Smaller effects are not cheaply provable. ### What is prompt tracking? Prompt tracking is running a fixed set of buyer questions through AI answer engines on a schedule and recording whether your brand was named, which sources were cited, and how the rate changes over time. It replaces rank tracking for surfaces that generate an answer instead of ordering a list. ### How often should you run prompt tracking? Daily for the first four to six weeks to establish your category's noise band, then weekly for the frozen set. Monthly refreshes cannot reach the convergence band on any engine, so a monthly cadence produces a chart of draws rather than a trend. ### Why does ChatGPT give different answers to the same prompt? Generation is probabilistic, and the retrieval pool behind it refreshes independently of your content. Two runs of the same prompt sample different sources and produce different text. This is why a single run is an observation rather than a measurement, and why repeated runs are the only way to get a rate you can defend. ## The bottom line Prompt tracking fails quietly. Nothing errors out. The dashboard fills, the weekly email sends, and the number it reports carries an error bar nobody printed. Three changes fix most of it. Budget runs instead of prompts, because runs carry the evidence. Size the run count per engine, because Gemini costs a fifth more than Google AI Mode to reach the same confidence. Establish your category's noise band before you set any threshold for what counts as a change. The standard advice in this category is to track 20 to 40 prompts. [SE Ranking recommends 20 to 40 across journey stages](https://seranking.com/blog/how-to-choose-prompts-to-track/), and [MaxAEO's prompt-run math](https://maxaeo.ai/blog/how-many-prompts-to-test-ai-visibility/) puts the default at 40 to 100 prompts run two to three times a week. Neither is wrong. Both answer the cheap half of the question and leave the half that decides whether the number holds up. An [AI visibility audit](/ai-visibility-audit) will tell you what your current sample can and cannot support before you commit another quarter to reporting it. ### Get a prompt tracking program sized to the claims you want to make Cite Solutions builds the topic-level prompt set, sizes the run budget per engine, establishes your category noise band, and runs the content and off-page work that moves cited-source share. [Book a Discovery Call](/contact) Tags [GEO](/tag/geo)[AEO](/tag/aeo)[AI visibility](/tag/ai-visibility)[AI citations](/tag/ai-citations)[ai search optimization](/tag/ai-search-optimization)[prompt tracking](/tag/prompt-tracking)[b2b ai visibility](/tag/b2b-ai-visibility) ## Continue the brief [01AI VisibilityOtterly AI: Is It Watching Your Market?Otterly AI monitors 50+ countries, wider than any tracker at its price. Country tracking covers four of seven engines, and each market costs you prompts.Aug 4, 2026Read→](/blog/otterly-ai-multi-country-tracking)[02AI VisibilityWhat Does Peec AI Actually Track?Peec AI includes three AI engines on every self-serve tier, out of six on offer. Here is what the three you drop would have told you, and what they cost.Aug 3, 2026Read→](/blog/peec-ai-what-it-tracks)[03AI VisibilityAI Overview Tracker: What It Sees and What It MissesAn AI overview tracker samples one surface, from one location, on a query set you picked. Five things it cannot see, and how to read the number anyway.Jul 31, 2026Read→](/blog/ai-overview-tracker-what-it-misses) [FrameworkLearn the CITE framework behind our GEO and AEO workSee how Comprehend, Influence, Track, and Evolve turn AI visibility into an operating system.](/framework)[ServicesExplore our managed GEO services and AEO execution modelAudit, prompt discovery, content execution, and ongoing monitoring tied to AI search outcomes.](/services)[AuditStart with an AI visibility audit before executionUnderstand prompt coverage, recommendation gaps, source mix, and where competitors are winning.](/ai-visibility-audit) On this page On this page ## Work with us on this [GEO AgencyManaged generative engine optimization for B2B brands.Explore→](/geo-agency)[AEO ServicesAnswer engine optimization: be the answer AI quotes.Explore→](/aeo-services)[AI Visibility AuditMeasure how AI engines cite and recommend you today.Explore→](/ai-visibility-audit) ## Ready to become the answer AI gives? Book a 30-minute discovery call. We'll show you what AI says about your brand today. No pitch. Just data. [Book a Discovery Call](/contact)