How to Track AI Citations Without a Platform

Someone on your team asks whether your brand appears in ChatGPT, so you type in a question a buyer might ask, and there you are, named in the second paragraph.

You screenshot it and feel good about it, until a colleague runs the same prompt the next morning and gets an answer that never mentions you.

That's the trouble with trying to track AI citations by hand.

A single answer tells you little. And a paid AI visibility tracker can feel like a lot to buy just to find out.

A spreadsheet can do the job well enough, as long as you follow three habits: asking the same buyer questions in each AI engine on a schedule, logging what every answer names and cites, and scoring the results as rates instead of single screenshots.

Let's learn more about the sheet, the sampling rule and the scoring method, all built to stay manageable for a small prompt set. We use no Tenpoint Labs data here, because our tracked prompts and counts stay in-house.

Key takeaways

• Log one row per run, not one row per prompt
• Run each prompt several times per engine
• Score named and cited separately, with the number of runs shown
• Add Search Console and GA4 as free cross-checks

What Is a Manual AI Visibility Tracker?

A manual AI visibility tracker is a spreadsheet where you record what AI engines say and cite for a fixed set of buyer questions, run by hand on a schedule. A citation is a page that an AI answer lists as a source, and a mention is your brand's name appearing in the answer text. The sheet measures both: whether an answer names your brand, and whether it cites a page on your domain.

Those two outcomes don't always travel together, like a band and its drummer's opinions. An engine can describe your product accurately without linking to you, and it can link to your page without naming your brand in the text. Our GEO playbook for B2B covers how pages earn those citations in the first place.

The sheet records what you saw, in your session, on the day you ran it. That makes it a sample. The rest of this guide is about making the sample trustworthy.

How to Track AI Citations in the Sheet

The sheet needs one row per prompt run, with columns for the prompt, engine, date, run number, whether your brand was named, whether your domain was cited, the cited URLs, the competitors named and notes. This table shows what to record in each column and why.

Keep the prompt list fixed once you start. A sheet whose questions change every week can't show a trend, only a mood.

How Many Runs Does Each Prompt Need?

Run each prompt several times per engine, because AI answers change from one run to the next and a single run is a sample of one. Our rule of thumb is five runs per prompt per engine, spread across a few days.

It's a starting point you can raise, not a threshold from a study. On one engine, we'd take five runs of ten prompts over fifty prompts run once.

The reason is volatility. A July 2026 study by GetMentions, an AI-visibility vendor, ran the same prompts daily for a week in logged-out sessions from a fixed location. It found that the sources behind a typical answer changed substantially from one day to the next, most on Gemini, ChatGPT and Google's AI Mode and least on Perplexity.

That's one vendor's one-week dataset, and it measured changes in cited sources, not how often a brand is named. It supports repeating runs. It doesn't set a number.

A screenshot of one lucky answer is a sample of one with excellent lighting. Rank checks have the same problem, as our piece on why two trackers disagree explains.

Report every result with its base, such as "named in 3 of 5 runs", never just "named". Five runs is a small sample, so a result like "3 of 5" is a rough signal, and only large gaps between prompts or competitors deserve action.

How Do You Score the Results?

Score each prompt and engine as the share of runs where your brand was named and the share where your domain was cited, always shown with the number of runs behind it. In a hypothetical sheet, one prompt run five times on one engine names the brand in three runs and cites its domain in one, which reads "named 3 of 5, cited 1 of 5".

Keep named and cited as separate scores. Compare yourself with competitors only on the same prompts and the same engines, and judge change over weeks, not days.

We haven't set a cutoff for what counts as good, because we couldn't find a source that gives one. Any published threshold is an opinion until you see how it was measured.

What Free Signals Can You Add?

Two free Google reports sit outside the sheet: Search Console, which includes the Performance report and a newer generative AI report, and the Google Analytics AI Assistant channel. None shows what an answer says, but they show how Google's AI features and AI assistants relate to your site.

Sites appearing in AI features such as AI Overviews and AI Mode are included in overall search traffic in the Performance report's Web search type, according to Google's AI features documentation.

Search Console's generative AI performance report, which Google says it rolled out to all websites worldwide as of August 31, 2026, breaks out your impressions in AI Overviews and AI Mode by page, country and device. It shows impressions, not the answers or the sources listed.

Google Analytics now reports an AI Assistant channel in the Default Channel Group, covering visits whose referrer matches a recognized assistant such as ChatGPT, Gemini or Claude. It counts visits that arrive with a recognized referrer, so an answer that names you without a link won't appear in it, and a visit shows that a click happened, not what the answer said.

What Can't a Spreadsheet Tell You?

A spreadsheet can't tell you what other people see, because it records the answers your own sessions returned. These are the limits to state whenever you share the numbers:

• Personalization and location can change an answer
• Logged-in and logged-out sessions can differ
• Engines change models and features without notice
• A growing prompt set outruns what you can repeat by hand
• The sheet shows what was cited, not why

Say so in the report. A tracker that states its limits is easier to trust than a press release.

Each limit has a practical fix. Record the session type in the notes column, run prompts from the same location each time, and re-run a handful of old prompts whenever an engine announces a change, so you can tell a real shift in your results from a shift in the engine.

When Does a Platform Make Sense?

A platform makes sense when the number of prompts, engines and runs is more than you can repeat by hand every week. At that point the sheet has done its job: it showed you which prompts matter and gave you a baseline to test a tool against.

Our comparison of AI citation trackers shows which tools list the sources behind an answer. Check any tool's results against your sheet on a few prompts before you trust it.

FAQs

Start With One Sheet and One Question

So how do you track AI citations? One row per run, several runs per prompt, and rates shown with their base. If the sheet outgrows your week, that's the signal to test a platform against it, and not before.