---
url: 'https://seoafterai.com/analytics-data/why-twenty-prompts-beat-twenty-days-of-one/'
title: Why twenty prompts beat twenty days of one
author:
  name: Priya Raman
  url: 'https://seoafterai.com/author/p-raman/'
date: '2026-11-05T06:48:25+00:00'
modified: '2026-11-05T06:48:29+00:00'
type: post
summary: 'SISTRIX analysed 10,826 daily measurement series across four AI models and eight countries. With a budget of twenty queries, spending them on twenty different prompts is more than three times as accurate as asking one prompt for twenty days. We checked the same shape in our own crawler logs.'
categories:
  - 'Analytics & Data'
tags:
  - AI Mode
  - AI Overviews
  - ChatGPT Search
  - Perplexity
image: 'https://seoafterai.com/wp-content/uploads/2026/11/cover-349-c23c5262.png'
published: true
---

# Why twenty prompts beat twenty days of one

SISTRIX analysed 10,826 daily measurement series across four AI models and eight countries. With a budget of twenty queries, spending them on twenty different prompts is more than three times as accurate as asking one prompt for twenty days. We checked the same shape in our own crawler logs.

## Table of contents

- [One question, twenty answers, no pattern](#one-question-twenty-answers-no-pattern)
- [Ten thousand measurement series](#ten-thousand-measurement-series)
- [Deviation by prompt count](#deviation-by-prompt-count)
- [Twenty-three days of our own logs](#twenty-three-days-of-our-own-logs)
- [Holes in both measurements](#holes-in-both-measurements)
- [What to buy with a fixed query budget](#what-to-buy-with-a-fixed-query-budget)
- [FAQ](#faq)

## Key takeaways

- Across 10,826 daily series, 70% of brands appeared in a prompt’s answers on fewer than one day in five. Only 6% appeared on four days in five or more. The pattern held on ChatGPT, AI Overviews, AI Mode and Perplexity.
- For a brand whose true appearance rate is 25%, one prompt over twenty days lands anywhere between 0% and 59%. Twenty different prompts on one day land between 16% and 35%.
- The gap between prompts is 2.4 times larger than the day-to-day wobble of one prompt, so repetition only attacks the smaller half of the noise.
- Prompt sets chosen by the brand flatter it: a better score than a neutral pool in three cases of four, by 63% to 80% depending on the comparison.
- In our own logs over 23 days, one day’s view of Bingbot would have put its article fetches anywhere from 1 to 10. The twenty-day mean is 3.4.

## One question, twenty answers, no pattern

Ask an answer engine the same question twice and you get two different answers, with different brands in them. That makes any single check close to worthless, and it raises a practical question: if you can afford twenty queries, do you spend them asking one question twenty times, or twenty questions once?

## Ten thousand measurement series

Johannes Beus published the analysis on 28 October 2026, built on 10,826 daily measurement series across four AI models and eight countries.

The instability is the starting point. Of all brands that appear in a prompt’s answers at all, 70% show up on fewer than one day in five; only 6% appear on at least four days in five. The split is the same on ChatGPT, AI Overviews, AI Mode and Perplexity. Beus cites a smaller experiment pointing the same way, run by Mike Sonders and written up by Mike King at iPullRank: twelve prompts, a hundred runs each, 72% of brands appearing in fewer than one answer in five.

The mechanism sits in two kinds of variation. The same prompt wobbles from day to day. Different prompts disagree far more sharply — a brand is almost always in the answer to one question and almost never in the answer to the next. Beus puts the between-prompt differences at 2.4 times the within-prompt wobble, which is why repetition hits a floor: twenty prompts carry ±7.6 points that no amount of repeating gets below.

One more finding changes how any vendor dashboard reads. In prompt sets chosen by the brand itself, the brand scores better than it does in a neutral pool in three cases out of four — between 63% and 80%, depending on the competitors compared against. Beus calls it selection rather than error: people build prompts around questions their own brand answers well.

## Deviation by prompt count

| How the budget is spent | Typical deviation | Measured range when the true rate is 25% |
| --- | --- | --- |
| 1 prompt, 20 days | ±34 points | 0 – 59% |
| 20 prompts, 1 day | ±9.5 points | 16 – 35% |
| 100 prompts, 1 day | ±4.2 points | 21 – 29% |

*SISTRIX, published 28 October 2026, from 10,826 daily measurement series across four AI models and eight countries. “Typical deviation” means roughly two measurements in three fall inside the range, and 95% inside twice it. With twenty prompts the deviation sits between ±9 and ±10 points in every model and every country measured.*

| Weekly stability in a neutral pool | Fluctuation, week to week |
| --- | --- |
| Brands with fewer than 100 prompts | ±5.7% (median) |
| Brands from 10,000 prompts | ±1.9% |
| Very large brands in AI Overviews | ±3.1% |

*Same source, measured on 60 brands per country across Germany, the US, the UK, France, Italy and Spain, as deviation from each brand’s own trend adjusted for changes in the prompt pool. Between 20% and 55% of the prompts mentioning a brand change every week, while the net total moves only 4–10%.*

## Twenty-three days of our own logs

We have no brand tracker, so we cannot reproduce the study. We can test the claim underneath it — that one observation of a noisy process misleads — on the one noisy process we do record: which of our articles the crawlers fetch.

We parsed origin logs from 13 October to 4 November 2026, 23 days, excluding our own two IP addresses, and counted the distinct article URLs each crawler fetched on each day.

Only one crawler gave us enough days. Bingbot appeared on 20 of the 23. Its daily article count ran from 1 to 10 against a 20-day mean of 3.4, so a single day’s reading would have been wrong by a factor of three either way. Of the 33 article URLs any crawler touched, Bingbot fetched 27 of them on no more than one day in five — 81%, next to the study’s 70%.

For the AI crawlers there is nothing to compute, and the reason is worth a number rather than a shrug: GPTBot appeared on one day of the 23, Googlebot on one, OAI-SearchBot on three, ClaudeBot on four. PerplexityBot came on seven days and touched two articles. One observation day leaves no spread to measure.

## Holes in both measurements

**Our CDN hides repeat visits.** Pages sit on the edge for four hours and a job refreshes them every three, so crawler requests served from the edge never reach our server. Our counts are a floor. The error runs in the direction that weakens our own comparison, which is the version we would rather publish.

**Thirty-three URLs is a census of a small site, not a sample.** The resemblance to the study’s 70% is a resemblance; at this size it could be coincidence.

**The study’s base is partly undisclosed.** Beus gives the series count, the models, the countries and the statistical meaning of his deviations, which is more than most. Missing: the prompt set, the window those 10,826 series cover, and whether answers came through each product’s interface or an API. Three of the seven method points we require, so these are SISTRIX’s measurements rather than constants.

**Self-selection cuts both ways.** The 63–80% finding says brand-chosen prompt sets flatter their owner. It comes from the vendor selling the neutral pool it recommends instead. Plausible mechanism, interested party.

## What to buy with a fixed query budget

**Put the budget into prompts, not repeats.** Twenty prompts gives roughly ±10 points, seventy gives ±5, two hundred gives ±3; halving the uncertainty takes four times as many prompts. Tracking fifteen prompts daily costs about what a hundred prompts weekly costs, and measures far worse — [the same arithmetic that decides whether a test is readable at all](https://seoafterai.com/technical-seo/the-seo-test-you-cannot-run-what-split-testing-needs-before-it-works/).

**Compare weeks, not days.** On a fixed set of twenty prompts a weekly average carries ±1.8 points against ±4.8 for one day. Daily values jump five to ten points for no reason, which is how a flat campaign reads as a win on Tuesday and a loss on Thursday.

**Keep two questions apart.** How visible you are needs a pool nobody built around you; whether your work changed anything needs a fixed set read as weekly averages. One set used for both gives a flattering number and an unreadable trend — and neither survives contact with [the rest of the measurement problem after AI answers](https://seoafterai.com/analytics-data/measuring-search-after-ai-answers/), where [citation counts move for their own reasons](https://seoafterai.com/ai-search/ai-citations-churn-daily-the-page-is-rarely-why/).

## FAQ

How many prompts do I need to track to get a usable number?
By the study’s figures: twenty prompts gives about ±10 points, seventy gives ±5, two hundred gives ±3. Halving the error takes four times the sample, so there is no cheap route to precision. If the budget stops at twenty, report the number as a range and refuse to discuss moves smaller than it.

My brand tracker shows a drop this week. Is it real?
Check the move against the noise first. In a neutral pool, brands with fewer than 100 prompts drift by a median of ±5.7% week to week with nothing behind it; brands tracked across 10,000 prompts drift ±1.9%. On a self-selected set of twenty prompts a single day carries ±4.8 points. A drop smaller than your own noise floor is not an event, and you learn your floor by measuring the weeks when nothing happened.

Why do different tools report different visibility numbers for the same brand?
Partly because they ask different questions. A prompt set assembled around your brand beats a neutral pool in three cases of four, by this study’s measurement, with the gap running 63% to 80%. Before comparing two tools, ask who chose the prompts and how many there are: two different pools measure two different things, and both can be internally consistent.