Monitoring from scratch, and keeping it alive.

Monitoring is the least difficult and most frequently abandoned part of this discipline. It is not hard; it is dull, repetitive and easy to skip during a busy week. Designing for that reality is what separates a routine that produces a trend line from one that produces two data points and a gap.

Building the prompt set

Twenty to thirty questions is the right range. Fewer misses variation; more becomes unsustainable manually.

Draw them from real sources: sales call questions, support tickets, search console queries and the phrasing customers actually use. Include recommendation questions, comparison questions, and problem questions, because assistants handle each differently.

Write them down and freeze them. A changed prompt starts a new series, so if you must change one, add the new version rather than editing the old.

Controlling for personalization and location

Assistants tailor answers to location, account history and settings, so decide how you will ask before the first run and keep it constant. Use the same accounts every time, with memory or personalization features turned off where possible, or logged-out sessions where the product allows it.

For local businesses, location matters most. If your customers are in one metro, run prompts from there, either by being there or by naming the place in the prompt (“in Round Rock, TX”). Record which approach you used; answers to “near me” and to a named town can differ noticeably.

The recording format

One row per ask: date, prompt, assistant, whether you were named, who else was named, sources cited, and a note on sentiment or accuracy.

That last column matters more than people expect. Discovering that an assistant describes a service you no longer offer is frequently the most valuable output of the whole exercise, and a yes-or-no visibility score would never surface it.

Making it survive

• Put it in the calendar as a recurring block with a named owner. Unowned recurring tasks stop.
• Keep it under an hour. If it takes half a day it will not happen. Twenty prompts across three assistants is roughly forty minutes.
• Ask each prompt several times. Five is a reasonable minimum, because a single ask cannot distinguish signal from variance.
• Accept imperfect runs. A fortnight missed is fine; abandoning because you missed one is not.
• Review quarterly, not monthly. Monthly review of a noisy metric encourages reacting to variance.

A template for the first run

For a first run, keep it simple and complete:

• Twenty prompts: eight recommendation, six comparison, six problem questions.
• Three assistants: ChatGPT, Google (AI Overview or AI Mode) and one of Perplexity, Gemini or Copilot.
• Three asks per prompt per assistant in the first run, to see how much answers vary.
• One row per ask in the sheet, with sources pasted in full.
• A short summary at the end: named in X of Y asks, top five competitors, top ten cited domains, inaccuracies found.

That is about 180 observations and a few hours of work. It is also a baseline most businesses never create, and the one every later decision will lean on.

When to move from manual to automated

Automate when the manual routine starts slipping, not before. Signs it is time: runs are being skipped, the prompt set needs to grow beyond thirty, you need separate tracking for several locations or products, or someone senior wants a weekly chart.

When you do, load the same prompts into the tool and run both methods in parallel for a month. The overlap shows how the tool's numbers relate to yours, so the trend line survives the switch instead of starting again.

Frequently asked questions

How do I monitor AI visibility manually?

Fix twenty to thirty buyer questions, ask each to ChatGPT, Perplexity and Gemini several times per period, and record date, prompt, assistant, whether you were named, who else was, sources cited, and any accuracy problems. Roughly forty minutes per fortnight.

How many prompts should I track?

Twenty to thirty. Fewer misses variation across question types; more becomes unsustainable manually and the routine stops, which is worse than tracking fewer consistently.

Why do in-house monitoring efforts fail?

Because the work is tedious rather than difficult, so it gets deprioritised during busy weeks and quietly stops, leaving no history. Naming an owner, keeping it under an hour and accepting occasional missed runs all help it survive.

Should I monitor AI visibility logged in or logged out?

Whichever you can keep consistent. Logged-out or memory-off sessions reduce personalization; the important thing is using the same setup every run.

How do I monitor Google AI Overviews?

Search your prompts in Google and record whether an AI Overview appears, what it says and which sources it links. AI Mode can be checked the same way. Results vary by location and time, so keep conditions constant.