How Many Prompts and Runs to Track AI Visibility

To measure AI visibility for a small business, plan on 30 to 50 distinct prompts per topic on each engine you care about, run each three or four times a month, and judge the result once a month, never from a single day. That gives you roughly 90 to 200 answers per topic per engine each month, which is enough to see large movements and too few to trust a change of 10 or 15 points.

A topic here means one buying need, for one kind of buyer, in one market. "CRM software" and "CRM for dental clinics in Leeds" are two topics and need two prompt sets.

No study has tested this exact plan. It's our compromise between what the research below suggests and what a small business can afford.

Say you run a bookkeeping firm and track one topic, "bookkeeper for a small construction company", on ChatGPT. You write 30 questions a buyer might ask, from "what does a bookkeeper do for a contractor" to "best bookkeeping firms for builders near me", and mention your firm's name in only two or three of them. Each question runs four times during the month, for about 120 answers.

At the end of the month the report says something like: named in 38% of answers, with a likely range of about 25% to 50%. Next month it says 45%. At this size a move of less than about 20 points is within the noise, so you note it and wait. If it says 65%, or 15%, something real has probably changed, and that's when you look at which answers changed and why.

The rest of this post explains where those numbers come from.

Because the same prompt rarely gets the same answer twice. In a [SparkToro and Gumshoe study](https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/), 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI 2,961 times. ChatGPT and Google's AI almost never gave the same list of brands twice, under 1 in 100 by SparkToro's count, with Claude slightly more consistent, and the same list in the same order came up about once in a thousand.

Presence held up better than position. On ChatGPT, City of Hope, a Los Angeles hospital, appeared in 69 of 71 answers to one prompt but came first in only 25 of them, which is why the study's authors call ranking-position claims "full of baloney" and suggest visibility percentage over many runs instead.

Sources move as much as brands. Researchers at the University of St. Gallen ran 32 prompts, eight in each of four industries, every day on four engines from late January to late March 2026, and [found](https://arxiv.org/abs/2604.07585) that two runs of the same prompt on the same day overlapped far less than you'd expect: roughly half of one run's cited sources showed up in the other, and well under half on ChatGPT.

At least 30 to 90 per topic and engine, judging by the one study that measured it directly. In a [July 2026 preprint](https://arxiv.org/abs/2607.10341), Ron Sielinski of IQRush collected 125 answers for each of 10 topics on Perplexity, ChatGPT search and Gemini, each answer to a different AI-written question, and measured when the ranking of cited websites stopped changing. It took between 33 and 94 answers, and three of the 30 topic and engine pairs, all on ChatGPT search, hadn't settled after 125.

That is a ranking of cited sites, not a brand-mention rate, so read it as an order of magnitude rather than a target. Even after the rankings settled, the typical top-10 site's plausible range spanned about five positions, so "we're number 3" and "we're number 7" can describe the same reality.

Pooling runs over time also helps: the St. Gallen team's uncertainty for each brand fell steadily as daily runs were combined, and they recommend rolling windows of two to four weeks. That assumes nothing big changes inside the window. A model update breaks the assumption, so mark updates and don't pool across them.

At least 30 per topic, because real phrasing varies far more than most prompt sets assume. When SparkToro asked 142 people to write their own prompt for the same need, a pair of headphones, almost no two were alike. Who is asking changes the answer too: a [May 2026 study](https://arxiv.org/abs/2605.30207) found that adding a buyer persona to otherwise identical prompts changed which brands were recommended, and mid-market brands saw up to 75% of their recommendation set change while category leaders stayed put.

So for each topic, write prompts that cover the main jobs a buyer is trying to do, with and without a persona, at different stages from "what is" to "which one should I buy". Take them from real customer questions, sales calls and search queries rather than inventing them. Leave your brand name out of most of them, because a prompt that names you will name you back, and it tells you nothing about whether a stranger would find you.

Three or four times a month is a sensible floor on engines that search the web live, such as ChatGPT and Perplexity, because a single prompt's answer swings a lot from run to run. Beyond that, more phrasings usually buy more than more repeats.

The studies point both ways because they measured different things. St. Gallen, tracking brand presence with only eight prompts per industry, had to get its precision from repeats and recommends at least seven runs per prompt per day. A [variance study by Dmitrij Żatuchin](https://arxiv.org/abs/2607.13304) measured how positive answers were about 20 brands rather than whether they were named, and its model projects that 15 phrasings run once beat 5 phrasings run five times. Its answers came from low-temperature API calls, which vary less than the apps your buyers use, so read it as a hint that phrasing is a big source of noise rather than a rule.

SparkToro's own advice is to run each prompt 60 to 100 times before trusting a visibility figure. A plan of 30 prompts run four times a month reaches 360 answers per topic over a quarter, or 12 runs per prompt. That's well short of SparkToro's figure, which is why a quarterly trend is more trustworthy than any single month. Several trackers, including Peec, SE Visible and Otterly, run each prompt daily, which gives plenty of repeats and usually too few phrasings.

At around 120 answers a month, about 20 to 25 points. A 10 or 15-point move is within the noise.

The textbook margins, for a brand named in about half its answers and every answer independent, look like this:

| Answers per topic per engine, per period | Margin on one figure | Margin on a change between two periods | |---|---|---| | 100 | about ±10 points | about ±14 points | | 300 | about ±6 points | about ±8 points | | 800 | about ±3.5 points | about ±5 points | | 1,500 | about ±2.5 points | about ±3.6 points |

Real margins are wider, because four runs of the same prompt are partly the same answer. 30 prompts run four times behave more like 40 to 75 independent answers than 120, which puts the margin on a month-to-month change at roughly 16 to 22 points. Near zero the arithmetic changes again: if you were named in none of 100 answers, your true rate could still be about 3%, or higher if those answers came from only a couple of dozen prompts.

You can tighten it by keeping the same prompts every month and comparing each prompt with itself, which removes much of the prompt-to-prompt noise from the change. And don't read too much into any single line: if you check nine topic-and-engine lines every month, expect one or two to look like a real change each quarter purely by chance.

Here is one for a business that sells to other businesses and wants to track three topics on ChatGPT, Google's AI and one other engine its buyers use:

- 30 distinct prompts per topic, 90 in all, with your brand name in no more than a handful. - A fixed core of those prompts kept for at least a quarter. New prompts go in a separate group, so a change in the number reflects the engines and not your rewording. - Each prompt run about four times a month on each engine, spread across the month rather than on one day, for 1,080 answers a month and about 120 per topic per engine. - The same engine mode, signed-in state and location every run, and every spelling and short form of your brand counted as a mention. - Your three or four main competitors measured on the same prompts, so you can tell a market-wide shift from your own. - One monthly figure per topic and engine, reported with its range. At this size that's about plus or minus 12 to 15 points. It would be 9 if every answer were independent, and nearer 18 if each prompt gives much the same answer every run, because then four runs tell you little more than one.

Pick the engines your buyers use, not the ones a tool includes by default.

Our own setup covers fewer engines than that plan. cite-met.com tracks 71 prompts on ChatGPT and Claude, and in the week we wrote this, most of that week's run failed with errors on our side, leaving 10 answers. The dashboard still showed a clean 0% named rate, and zero out of 10 is consistent with a true rate as high as 30%. Ask any tracker, ours included, to show how many answers stand behind each figure.

What a plan like this costs on each tool is in our [guide to AI visibility tracking costs](/blog/post/ai-visibility-tracking-cost).

Check for these before you trust a dashboard:

- The number of answers and distinct prompts behind every figure, not just the percentage. - A range next to each rate, worked out by resampling whole prompts rather than single answers, or an export of the raw answers so someone can do it for you. - Whether the headline counts mentions or citations. An answer can name you without linking to you, and some engines leave out citations in many answers. - Separate rates for each engine, since they disagree with each other and your buyers may favour one. - Markers for model updates, because a new model version can move every line at once. - Branded and unbranded prompts reported apart.

A note on the research. Four of the five studies cited here are 2026 preprints without peer review, and the fifth, SparkToro's, is a practitioner blog post. Four have an author or data partner tied to a company that sells AI visibility measurement, as we do: Gumshoe co-produced SparkToro's study, Sielinski works for IQRush, one St. Gallen author is affiliated with Aurora Intelligence, and Żatuchin with Rankfor.AI. They used different methods and still point the same way, but all of us have a reason to tell you measurement is harder than it looks.

If you want answer tracking alongside crawler, visit and lead data on the same site, our [plans](/pricing) show what each includes, and the [home page](/) explains how the pieces fit.