Your weekly AI visibility routine — fully manual (about 45 min/week)

Monitor your AI visibility without any tools: five fixed questions asked in ChatGPT every week with the wording kept constant, a log sheet with seven columns, and a monthly follow-up with one fix at a time. About 45 minutes a week — plus an honest section on where the manual method runs out of road.

You can monitor your AI visibility entirely by hand: ask ChatGPT the same five questions every week, in a new conversation and with the wording kept constant, log the answers in a spreadsheet, and follow up once a month with one fix at a time. It takes about 45 minutes a week when done thoroughly. Here is the whole routine — the log sheet, the weekly steps and the monthly follow-up — plus an honest section on where the method runs out of road.

Why should the measurement be a routine — and not a one-off test?

Because a single answer is only a sample. AI models don't answer the same way every time, so a single check cannot tell a real change from random variation. What you actually want to know — is my business mentioned more often or less often than last month? — can only be answered by asking the same thing, in the same way, week after week. The measurement lies in the repetition, not in the individual answer.

And there is something worth watching: Ahrefs' analysis of 300,000 keywords measured a 34.5 % lower click-through rate for the top result in Google when an AI answer sits above it — and 58 % in the December 2025 re-run. A growing share of your customers' research now ends inside the answer, without a click to anyone. So that is where you need to keep watch — not just in your traffic statistics.

There is an encouraging reason to measure regularly too: visibility can be moved. The Princeton/Georgia Tech research (KDD '24) measured up to a 40 % visibility lift in generative engines from targeted optimisation of content. But you can only see whether a fix worked if the measurement stood still in the meantime: same phrasings, same method, before and after. That is exactly what this routine gives you.

What do you need before you start?

Three things, all of them free: five fixed questions, a spreadsheet and a fixed slot in your calendar. The five phrasings are the foundation — they should sound like your customers, not like the owner, and your company name must not appear in any of them.

• Five fixed questions: If you don't have them yet, build them with the bank of questions your customers ask ChatGPT, and run them through the 5-question test the first time. From now on the wording is sacred: change a phrasing, and the time series starts over for that particular question.
• A spreadsheet: Google Sheets or Excel — the structure is in the next section. Paper doesn't work: you need to be able to filter and count once three months have passed.
• A fixed time: for example Monday morning before the day's first task. The routine rarely dies from being difficult — it dies from being “something I do when I find the time”.

What does the weekly routine look like, step by step?

Expect about 45 minutes in total: twenty to ask and take notes, ten for logging, ten for Search Console and five to skim this week's line against the previous ones. The order is fixed, so you don't skip the boring part:

1. Ask the five fixed questions — one at a time, each in its own new conversation (approx. 20 minutes). The wording must be verbatim as in the spreadsheet, and you must not mention your own company name along the way. A new conversation per question is not pedantry: ChatGPT uses the conversation's context, and one contaminated answer makes the week's measurement useless.
2. Log every answer straight away (approx. 10 minutes). Were you mentioned — yes, partly or no? Who was mentioned instead? Which sources did the answer refer to? “Partly” covers things like incorrect details, old addresses or a name that is almost yours.
3. Check Google Search Console (approx. 10 minutes). Look under Performance for new queries in question form. They show how real people phrase things about your field right now — and they are raw material for future adjustments to your five, without you changing anything in the middle of a time series.
4. Skim this week's line against the previous ones (approx. 5 minutes). You are not looking for conclusions yet — only for deviations worth noting: a new competitor in the answers, a source you haven't seen before, a “yes” that turned into a “no”.

The log sheet is the routine's memory, and seven columns are enough — one row per question per week, so a month fills around twenty rows:

The “Model” column is there for a reason: models are updated continuously, and a switch from one version to another can move the answers without you having changed anything. With the model in the sheet, you can later see whether a kink in the curve coincides with a model switch — or with something you did yourself.

What should the monthly follow-up contain?

Once a month you add half an hour on top and switch roles: from collecting data to using it. Four weeks is the smallest window where patterns can be told apart from noise, and the follow-up has three fixed items — in this order:

1. Review the trend. Count yes/partly/no per question over the last four to five weeks. Is the picture stable, swinging, or moving in one direction? Look at the “Mentioned instead” column too: if the same competitors or directories keep coming back, that tells you what kind of sources the AI leans on in your industry.
2. Pick ONE fix. Maybe the page answering the weakest of your five needs a rewrite. Maybe your details need correcting where the answers fetch them. Which type of fix makes sense depends on the outcome — if you are not being cited as a source, it is typically technology and content; if you are not being recommended, reputation and mentions matter too. The difference between the two outcomes is worth an article of its own. But always one fix at a time: do three things in one month, and afterwards you won't know what worked.
3. Verify last month's fix. Did it move anything in the sheet? The honest answer is often “not yet” — AI engines pick up changes slowly, so give a fix at least one more month before judging it. But the question must be asked every month, otherwise untested “improvements” pile up without anyone knowing whether they work.

Where does the manual method run out of road?

The routine above is real and produces usable data — we used it ourselves before we built anything at all. But it has three limitations you should know before basing decisions on it:

• One model at a time. 45 minutes covers ChatGPT. But Gemini, Copilot and Perplexity answer from their own sets of sources and can paint a different picture of your exact industry — and every extra model adds almost the same amount of time again. Most people never get past the first one.
• Sample noise. Five answers a week is still not many. A single “no” doesn't mean you have dropped out, and a single “yes” doesn't mean you are in — even with constant wording, you need to see the same picture several weeks in a row before it is a pattern. Expect the first four to six weeks to mostly produce patience.
• Discipline slips. That is the experience — our own included: the first weeks are easy, then a busy Monday arrives, and “I'll do it tomorrow” becomes a gap in the sheet. Gaps are not cosmetic. They are exactly what makes it impossible to see afterwards when a change happened — and then the time series has effectively started over.

When does it make sense to automate?

When one of two points is reached: discipline slips, or you want to measure more than the 45 minutes can cover — more phrasings, more models, the competitor picture over time. The routine in this guide is exactly what Geoa automates: the same fixed phrasings, measured systematically, with a notification when something moves — from 299 kr/month, versus roughly 3 hours of your own work a month.

You can start without paying: a free account monitors your first 5 questions automatically, and paid plans cover 25, 100 or 300 depending on tier — all with a 14-day free trial, no card required. And either way, the work in this guide is not wasted: your five well-tested phrasings and your log sheet are exactly the baseline an automatic measurement should build on.

Frequently asked questions

How long does the routine realistically take?

About 45 minutes a week for five questions in one model, plus half an hour for the monthly follow-up — around 3 hours a month in total. The first week takes longer, because you also need to settle the phrasings and set up the log sheet.

Can I get away with every other week?

Yes — the routine also works every other week, as long as the wording and method are kept constant. The price is time: with half as many data points, it takes roughly twice as long to see whether something is a pattern or just noise. Better to pick a cadence you can keep than an ambitious one you drop.

Which AI should I test in?

Start with ChatGPT — it is the one most of your customers use. Feel free to add Perplexity: it shows its sources clearly, so you can see exactly where the information in the answer comes from. More models give a richer picture but cost correspondingly more time — build the habit with one first.

When should I switch to automatic monitoring?

The honest answer: when discipline slips — gaps in the log sheet ruin the time series — or when you want to measure more questions or more models than the manual time can cover. Until then, the manual routine is a fully valid place to start, and everything you log can be reused the day you automate.