How often you should measure AI visibility

Monthly is the right rhythm for most businesses, weekly is worth it only while you are publishing heavily or chasing a specific fix, and daily is noise. The reason is not cost. It is that answers vary between runs, so a short gap between readings shows you variation and calls it change.

What is actually happening underneath

Two clocks are running and they are both slow. The pages that feed an answer change over weeks, not days. And your own work, a new page, a claimed profile, a corrected fact, takes time to be crawled, indexed and then reached by a question.

Against those slow clocks sits fast noise. Ask the same question twice in an hour and the answer can differ. So the difference between Monday and Tuesday is almost entirely noise, while the difference between January and April is mostly signal.

Frequency also cannot rescue a small question set. Reading five questions every day gives you many readings of a set too small to produce a rate you should trust. Adding questions helps more than adding readings. We have a separate page on the arithmetic of that.

Four cadences and what each one is for

Pick by purpose rather than by appetite.

Monthly, the default. Fixed date, fixed question set, fixed assistants. Enough to see a trend inside a quarter, cheap enough that it actually happens. This is the reading you report.

Weekly, temporary. Two situations justify it. You have just published a batch of pages and want to catch the first citation. Or an assistant is stating something wrong and you want to know when the correction lands. Both are temporary, so go back to monthly afterwards.

Quarterly, for the wider set. A larger question set, more zones, more assistants, run less often. This is where you check whether your coverage of the category has moved rather than whether this month's pages landed.

On demand, for a specific question. Before a launch, after a price change, when sales report that a prospect quoted something wrong. Never counted in the trend, because it is not on the schedule.

Daily is not on the list. It produces rows that measure variation, and it pushes teams into reacting to noise.

A worked example

In the September 2026 capture on our own site, aiknowsus.com, we asked Perplexity in batch after batch to say which competitor was winning most often.

It declined every time, and its reason is the useful part. It said its own previous answers had not produced a comparable live tested result, so there was nothing there to support a claim about who was ahead. An assistant refusing to rank from a single pass is the clearest statement we have that one reading establishes nothing.

Apply that to your own cadence. A daily reading gives you a new answer every day and no more evidence than the first one gave you, because what changes between Monday and Tuesday is the draw rather than your position.

What does accumulate is a series of readings taken the same way on a fixed date. Three of those pointing one way is the weakest useful evidence of a real change, and a monthly reading gets you there inside a quarter. Thirty daily readings in the same month get you thirty draws and no more direction than the first one.

What it means for a business

Monthly, on a fixed date, with the same questions and the same assistants. Log the answers, not only the rate, because what is said about you is useful every time while the rate moves slowly.

One thing must not change often: the question set. It is the denominator, so changing it changes the number for reasons that have nothing to do with your visibility. If you add questions, record the date you added them and keep the original set reported separately for a while.

Put the reading in one person's calendar with a named owner. The most common failure here is not the wrong cadence, it is a measurement that quietly stops in month three.

The five things that must stay fixed, and the two that may move

A reading is only comparable with the one before it if the setup was the same. These five are frozen.

One, the question wording. Copied, never retyped. A reworded question is a new question and its first reading is a new baseline.

Two, the assistants. Same apps, same order, and the same surface each time, app or API but not both.

Three, the session conditions. Fresh chat, signed out or memory off, same browser profile.

Four, the modes. If you read with search off and search on, do both every time and keep them in separate rows.

Five, the definition of a mention. Named anywhere and named in the first three, counted the same way every month.

Two things are allowed to move, on purpose and with a note. You may add questions, as long as you record the date and keep the original set reported separately for a few months. And you may add an assistant, reported as its own column from the month it starts, never backfilled.

What a cadence cannot fix

Four limits worth saying before somebody buys a daily dashboard.

Frequency does not reduce variation. It gives you more draws from the same noisy process. A bigger question set reduces variation. A shorter gap does not.

It cannot make a small set trustworthy. Five questions read daily is still five questions.

It cannot attribute a change. A reading tells you the answer moved, not why. Attribution comes from your own log of what you changed and when, kept next to the readings.

It cannot tell you about buyers who never asked. You are measuring your questions, not demand. No cadence turns that into a share of anything.

Common questions

Is daily tracking ever worth it?

Only for a short, specific reason: watching for the first citation of a page you just published, or waiting for a correction to land. Even then, watch a handful of questions daily rather than the whole set, and keep the monthly reading as the one you report.

What day of the month should I pick?

Any day, kept the same. Consistency matters and the date itself does not. Pick a day that is not the end of the month, so the reading is not the first thing dropped when reporting gets busy.

Can I add questions between readings?

Yes, and record the date and keep the original set reported separately for a few months. Otherwise the number moves because the set moved, which is the easiest way to mislead yourself.

Should I read more often after publishing a batch of pages?

Weekly for a few weeks is reasonable, on the specific questions those pages answer. Then stop. The rest of the set does not need to move to weekly because one part of your work did.

How many readings before I believe a change?

Three pointing the same way is the weakest useful evidence, and it is much stronger with a large set. Two readings whose ranges overlap have not shown a difference at all, whatever the headline numbers look like.

Does the time of day matter?

Not in a way worth controlling for. Variation between runs is larger than anything you would gain by fixing the hour. Fix the questions and the apps instead.

What to do first

• Pick a date in the month and put it in the calendar as a recurring task with a named owner.
• Freeze your question set and save it in a file, not in somebody's head.
• Decide your assistants and keep them fixed too.
• Take today's reading as the baseline, even if the set is smaller than you would like.
• Start a change log next to the readings: what you published or corrected, and the date.

We run automated readings on a schedule, which is our interest in the question. The honest version is that the schedule matters less than keeping the set fixed and reading the answers. A careful monthly check in a spreadsheet beats a daily dashboard nobody opens.

How we know this

We read the assistants rather than guessing what they do. In our own audit of 1,260 saved answers from ChatGPT and Gemini across the United States, the United Kingdom and India, 82.0% of every citation landed on a site the business itself owned, and review directories took 0.7%. Counted as one domain per reading, 16,516 of 20,143.

That is the whole asymmetry: the pages you own are the pages an assistant quotes back to a buyer. See the dated readings and the method on our proof page.

Find out whether assistants name your business, on your own domain and your own buying questions.

By Aditya Goel, Founder, AI Knows Us. Reviewed 1 November 2026.