# Why can tracking AI visibility every day be misleading?

> A visibility line drawn from 10 questions a day can swing 20 to 30 points on its own. What is daily tracking for, and over what interval should you read it?

Published: 2026-10-07
<https://geomono.com/en/blog/yapay-zeka-gorunurlugunu-her-gun-takip-etmek>

Tracking AI visibility every day as a single line can be misleading, because the same question produces different answers even on the same day, and in a small question set this randomness creates a bigger swing than any change in the brand's real position. A line drawn from 10 questions a day can rise and fall by 20 to 30 points without anything having changed. Daily tracking is valuable for catching sudden events; the visibility rate, by contrast, should be read with the same questions, with repetition and period by period.

## Same question, same day, different answer

AI engines do not give the same answer to a question every time. The answer changes with the model's sampling at that moment, the web search it runs at that moment and the sources it reads. This is not a fault; it is how these systems work.

In the [research](https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/) SparkToro published with Gumshoe on 28 January 2026, 600 volunteers asked 12 questions to ChatGPT, Claude and Google's AI a total of 2,961 times. According to the research, the chance that ChatGPT or Google's AI gives the same list of brands in two answers to the same question is less than one percent, and the chance of the same list arriving in the same order is about one in a thousand. By contrast, the number of answers a brand appeared in, that is, how often it appeared, was much more consistent. The researchers write that rankings are not reliable, while an appearance rate measured over dozens or hundreds of repetitions is a reasonable measure.

The ["Don't Measure Once"](https://arxiv.org/html/2604.07585) study published on arXiv in April 2026 (Julius Schulte, Malte Bleeker, Philipp Kaufmann; a preprint that has not been peer reviewed) repeated the same questions in two different ways. ChatGPT, Gemini, Google AI Mode and Perplexity were asked 32 questions across four sectors, first every day between 24 January and 20 March 2026, then many times within the same day between 21 and 25 March. A similarity measure showing how much the source lists of two answers overlap (0 means no overlap, 1 means identical) came out between 0.34 and 0.42 from day to day, and between 0.32 and 0.43 within the same day.

That these two ranges are almost the same is an important result. The authors' interpretation is that most of the observed swing comes not from change over time but from the randomness of the answers themselves. In other words, the difference between yesterday and today is often no bigger than the difference between asking twice at the same hour.

## A single response is a roll of the dice

Whether a brand is mentioned for a particular question is, when you look at a single response, like a roll of the dice. Say your brand would be mentioned in exactly half of buyer questions if measured for months: a true mention rate of 50%. Today you asked 10 questions. The number in which you are mentioned does not have to be 5; it could be 3, or 7.

The size of this swing can be calculated. With a rate of 50%, the standard deviation of a single response is 0.5; the standard error of the mean over n questions is 0.5 divided by the square root of n. For 10 questions this is √(0.5 × 0.5 / 10) ≈ 0.158, or about 16 points. The 95% margin of error is 1.96 times this: 1.96 × 0.158 ≈ 0.31, or about ±31 points.

When you compare two separate days, the uncertainty of the two days stacks up. The standard deviation of the difference between two independent days is √2 × 0.158 ≈ 0.22, or about 22 points. This means that even when nothing has changed, on roughly one day in three you will see a difference of more than 22 points from the previous day. That is why it is not surprising for a line of 10 questions a day to swing 20 to 30 points purely by chance.

| Questions per day (n) | 95% margin of error (at a 50% rate) | Typical size of the difference between two days |
|---|---|---|
| 10 | ±31.0 points | 22.4 points |
| 40 | ±15.5 points | 11.2 points |
| 100 | ±9.8 points | 7.1 points |
| 400 | ±4.9 points | 3.5 points |

The calculation in the table rests on the simplest model, in which every response behaves like an independent roll of the dice, and it is a guide. In reality some questions give the same answer every time, while others change almost every time; the real size of the swing depends on how many questions in your set are unstable. What the table really says is this: quadrupling the number of questions halves the margin of error. The way to shrink uncertainty is not to look more often but to collect more responses.

> [!NOT]
> The margin of error is at its largest when the rate is 50%. If your rate is close to 10% or 90%, the interval narrows somewhat for the same number of questions, but with 10 questions it is still not enough to interpret a one-day difference.

## Then there are the days that cannot be measured

There is a second thing that misleads a daily line: responses where the engine did not answer or the answer could not be read. If these responses are counted as zero on a day with a service outage, the line drops as if your brand had lost visibility.

This is not a small share. In our own measurement, between 16 and 25 August 2026, for three corporate brands (a pharmaceutical company, a dairy brand and a technology group), 189 (10.3%) of the 1,835 planned responses to questions asked to ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews without the brand name could not be measured. These were not counted. This figure comes only from the six measurements of these three brands; it is not a market average.

We explain step by step how to keep unmeasured responses separate in [our guide to how brand visibility in AI is measured](/en/blog/yapay-zeka-marka-gorunurlugu-nasil-olculur).

{{cta}}

## Why do measurement tools aggregate results?

Measurement methods that take this randomness into account read the result not from a single day's answers but from the total of answers over a time window. The methods differ, but the direction is the same.

| Source | What the method says |
|---|---|
| ["Don't Measure Once" (arXiv, April 2026)](https://arxiv.org/html/2604.07585) | For brand visibility, it recommends asking each question at least 7 times a day, at least 8 times if source coverage matters, and reading the result over a rolling window of two to four weeks. |
| [Ahrefs Brand Radar methodology page](https://ahrefs.com/blog/brand-radar-methodology/) | Refreshes ChatGPT, Perplexity, Gemini and Copilot data once a month and uses a 90-day reporting window; Google AI Overviews and AI Mode are updated continuously along with keyword data. |
| [Profound, "The half-life of an AI citation is 11 days"](https://www.tryprofound.com/blog/the-half-life-of-an-ai-citation-is-11-days) | Smooths a page's daily citation share with the average of the last 14 days, to suppress daily fluctuation. |

These three methods show that the daily result is not read on its own. Ahrefs' 90-day window and Profound's 14-day average serve different purposes, but both dissolve a single day's noise into the total.

The finding in the same Profound article also shows why aggregation is necessary: in the company's data for one year (September 2025 to September 2026, seven engines), the median cited page fell to half of its peak citation share within 11 days. If sources change quickly, a single day's measurement carries both noise and real change at the same time.

There is also data suggesting that a weekly comparison looks more stable. In a [note it published](https://www.brightedge.com/resources/weekly-ai-search-insights/ai-search-citations-week-to-week-changes) in February 2026, BrightEdge writes that across the nine sectors it tracks on its own platform, 96.8% of the domains cited by five engines did not change at all from one week to the next. This is a finding based on the company's own platform data and has not been independently verified; it also compares a week's set of citations, not individual responses.

## What is daily tracking good for?

Looking daily is not wrong; what is wrong is reading the daily rate as a trend. What is worth watching every day are sudden, binary events: either it happened or it did not.

- **The site being closed to AI crawlers.** A robots.txt change or a setting in a content delivery network can make your site unreadable overnight. With new settings [announced](https://blog.cloudflare.com/content-independence-day-ai-options/) on 1 July 2026, Cloudflare said that from 15 September 2026, for domains newly joining Cloudflare, training and agent crawlers would be blocked by default on pages that show ads, while search crawlers would remain allowed. Noticing a change like this in a visibility line takes weeks; checking crawler access takes a few seconds.
- **The measured model changing.** Consumer apps change the model they serve from time to time. If the measurement is still being made with the old model, the results do not represent the answer the user sees.
- **A new false claim.** Wrong information about your brand beginning to appear in answers is also a sudden event. It makes sense to check a few questions you know to be risky by hand more often. In GeoMono, the wrong-information check is not a check that runs every day.

What these events have in common is that they are states, not rates. The answer to "is crawler access open?" can be measured the same way every day, and when it changes it tells you something definite. We explain how to manage AI crawlers in your robots.txt file in [our robots.txt guide](/en/blog/robots-txt-yapay-zeka-botlari-rehberi).

## How we work

We built our measurements around this distinction: the visibility rate is checked period by period and with repetition, while sudden events are checked every day as states. The details are on the [How we measure](/en/yontem) page.

### Visibility: same questions, repetition and interval

A registered measurement is repeated with the same question set, and two measurements are compared only if they were made with the same questions. If the question set changed, the two results are not placed side by side.

In scan-based analyses, the questions on which the engines disagree are asked 3 more times to every engine that responded, and a majority decision is used. Questions on which all the engines agree are not repeated, because that is not where the uncertainty is: asking again a question whose answer comes out the same every time does not narrow the interval.

When two periods are compared and a rate has changed, the report also states whether the difference can be distinguished from sampling noise and shows the likely interval. If the interval includes zero, no change is claimed. How often a scan-based analysis is refreshed is chosen in the brand settings: off, weekly or monthly.

### Sudden events: a daily status check

Sites for which technical readiness tracking has been started are rechecked every day, starting with the site that has gone longest without a check. The check re-measures whether the site is open to AI crawlers and how readable its pages are. If the picture has changed, a new dated row is added to the change log; if it has not, no new row is added. What comes out is not a chart but a log showing what changed and when. If the check itself fails to run, our team gets an alert.

Every day we also check whether the models used in the measurement are still the same as the models the engines serve in their consumer apps, and whether the providers have released a newer model in the same family. If a drift is seen, our team gets an alert. This check asks the engines no questions; it only reads model information.

> [!IPUCU]
> If you want to see right away whether your site is open to AI crawlers today, you can use the [free scan](/en/tarama). No sign-up is needed.

## Frequently asked questions

### Is a daily visibility chart completely meaningless?

It is not meaningless, but it should not be read on its own. On a line drawn from a small number of questions a day, a one-day drop or rise is mostly chance. It is more reliable to aggregate daily data over weeks and read it that way, or to limit the daily view to sudden events such as crawler access.

### How many questions and how many repetitions are needed?

There is no one-size-fits-all answer. When the true rate is around 50%, the 95% margin of error is about ±10 points with 100 responses and about ±5 points with 400 responses. The April 2026 arXiv study recommends asking each question at least 7 times a day and reading the result over a window of two to four weeks.

### My visibility fell 15 points in a week. Should I be worried?

First check how many responses these two numbers were calculated from, and whether they were measured with the same questions. Between two measurements of 40 responses each, a 15-point difference can easily arise by chance alone. Also check whether unmeasured responses were counted as zero; a service outage can look like a decline.

### Why do different brands come up for the same question every time?

The engines resample the model for every answer, run different searches and read different sources. The differences between engines add to this. We cover this topic in detail in [our article on why the same Turkish question brings up different brands](/en/blog/ayni-turkce-soruya-farkli-markalar).

## Sources

- [NEW Research: AIs are highly inconsistent when recommending brands or products; marketers should take care when tracking AI visibility](https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/), SparkToro, 28 January 2026
- [Don't Measure Once: Measuring Visibility in AI Search (GEO)](https://arxiv.org/abs/2604.07585), arXiv (Schulte, Bleeker, Kaufmann), 8 April 2026
- [Ahrefs Brand Radar Methodology: How we collect and model AI visibility data](https://ahrefs.com/blog/brand-radar-methodology/), Ahrefs, updated 26 February 2026
- [The half-life of an AI citation is 11 days](https://www.tryprofound.com/blog/the-half-life-of-an-ai-citation-is-11-days), Profound, 30 September 2026
- [AI Search Citations: How Much Do They Really Change Week to Week?](https://www.brightedge.com/resources/weekly-ai-search-insights/ai-search-citations-week-to-week-changes), BrightEdge, 6 February 2026
- [Your site, your rules: new AI traffic options for all customers](https://blog.cloudflare.com/content-independence-day-ai-options/), Cloudflare Blog, 1 July 2026
- Share of responses that could not be measured: questions asked without the brand name, 2026-08-25: 6 measurements between 16 and 25 August 2026, 3 corporate brands (a pharmaceutical company, a dairy brand, a technology group), ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews; 1,646 of the 1,835 planned responses were measured, 189 could not be measured. Our own client and comparison brands, not a market sample.
- How we measure: the GeoMono method, 2026-09-28: published method; the repeat-question rule and the period comparison are described on this page (https://geomono.com/tr/yontem)