A 30-Minute Weekly AI Visibility Review

A visibility programme dies in one of two ways. It becomes a monthly slide nobody reads, or it becomes a daily anxiety where somebody checks the chart, sees noise, and reacts to it. Both failures come from the same root: no agreed cadence, and no agreed list of things that are allowed to trigger action.

Thirty minutes a week fixes it. Not because thirty minutes is enough time to do deep work, but because the review is not for doing work. It is for deciding what work is worth doing, and that decision takes about half an hour if you know the order to ask the questions in.

Before the numbers: did the scan actually run?

Three minutes, and it goes first because every number below is conditional on it. An empty-looking chart has four causes and they are not interchangeable.

First scan in progress. Queued or running. Scans trickle across their window rather than firing all at once, so results arrive over several days. Nothing to do but wait.

No scan covered this window. Nothing ran in the range you selected. This is not a measurement of zero. It is the absence of a measurement. Widen the range.

Scans ran, but nothing was scored. Answers came back and no confirmed brand was set up to look for. "Nobody told us what to look for" is a thirty-second fix and a completely different sentence from "we looked and found nothing."

Scan completed with gaps. Some engine calls did not land. The rates are still correct for what ran; they simply cover less evidence than a healthy scan, and a rate over fewer runs has a wider interval.

Get in the habit of naming which of the four you are looking at before you look at anything else. It takes seconds and it prevents the single most common wasted hour in this discipline: diagnosing a content problem that was a scheduling artefact.

The agenda

Block 2: real movement, and the discipline of finding none

Five minutes, and most weeks the honest output is "nothing moved."

The rule is simple and it has to be held: a change counts only when the confidence intervals do not overlap. A move from three mentions in ten runs to four looks like a 33% improvement and is statistically indistinguishable from a coin flip. If your product renders such a move as flat, believe it.

The value of saying "nothing moved" in a meeting is underrated. It is a real finding, it takes five seconds, and it stops the group inventing an explanation for randomness, which is what happens by default when six people look at a chart with nothing on the agenda after it.

Block 4: one prompt, read properly

The largest block, and the one people skip because it does not feel like analysis.

Pick one prompt you lose consistently. Not a different one each week for variety: the same one, until you have done something about it. Open the stored answers and read them in full.

You are looking for four things, and you will usually find at least two:

What the model thinks the category is. Often narrower or broader than your positioning assumes.

Who is presented as the default. There is frequently one brand that appears first in most answers. That is the incumbent position in the model's view of the category, and it is not always the market leader.

A stale fact about you. Pricing that changed, a feature you shipped, a limitation you removed. Models repeat what they read, and what they read may be two years old. This is the most fixable finding in the entire review and it only surfaces by reading.

Which page won. The answer's sources list the URLs that got used. Open the one that beat you and read it as a retriever would: does one passage answer the question completely, on its own, with the vendor named inside it? That is usually the whole difference, and it is a paragraph edit rather than a strategy.

Block 6: crawl coverage, the leading indicator

Five minutes on the least exciting number in the review, because it is the one that moves first.

Take the pages that should be answering your tracked prompts and check when a search-index crawler last fetched them. Not a training crawler: those tell you nothing about being citable. If those pages have not been fetched recently, that is your entire finding for the week, and everything downstream is explained by it.

The reason this belongs in a weekly cadence rather than a monthly one is the lag. Crawl access breaks on a Tuesday and shows up in your traffic six to eight weeks later. Catching it in the week it happens is the difference between a fix and an incident.

What not to do weekly

Three things, and each one is tempting precisely because the review makes them feel urgent.

Do not change the prompt set. Every reworded prompt breaks the series. The urge to fix a prompt peaks in week three and the correct response is to write it down and revisit it quarterly. If a prompt is genuinely wrong, add a corrected one and keep the original running rather than editing in place.

Do not commission content off one week of data. A single week is a handful of runs per prompt. A content brief written on that evidence is a bet on noise, and the cost is a quarter of somebody's writing capacity.

Do not report the trend to leadership weekly. Weekly numbers move within their intervals. Reporting them creates a rhythm of explaining moves that are not moves, and it burns the credibility you will need when something real happens. Monthly, with the interval visible, is the right cadence upward.

Monthly and quarterly

The weekly review is not the whole programme. Two slower loops sit above it.

Monthly, add the numbers that need a longer window to be readable: attributed revenue from AI sources, the conversion rate against your site baseline, the crawl-to-refer ratio per operator. These need lag-aware windows and are meaningless week to week. Monthly is also when the trend goes to leadership, with its interval attached.

Quarterly, do the two things the weekly cadence is forbidden from touching. Review the prompt set against what your buyers are actually asking (sales calls, support tickets, the questions in your own search logs) and decide whether to start a second set. And review the category: is the roster of brands beside you drifting toward an adjacent market, and if so, is that a framing to contest or to accept?

Keep a one-page log

The last piece, and the one that makes the whole cadence compound rather than repeat.

Every week, write four lines: the date, the scan state, whether anything moved beyond its interval, and the one action taken. Nothing longer. Nobody reads longer.

The value shows up in month four, when a rate moves for real and somebody asks what changed. A log of thirteen four-line entries answers that in about a minute. Thirteen weeks of memory answers it with a story, and the story will be wrong, because a model version changed in week seven and nobody wrote it down.

That is the entire argument for cadence over intensity. The programme that wins here is not the one that looks hardest at the chart. It is the one that is still running in six months, with a record of what it did, so that when something finally moves you can say what caused it.