How to Track Your Citations in ChatGPT: A Weekly Routine That Takes Under an Hour

Checking ChatGPT once and screenshotting the result is not tracking. Answers vary from run to run, so a single look tells you very little. Tracking means the same questions, asked on a schedule, with the sources written down in a form you can compare next week.

Here is a routine that works, whether you do it by hand for a month to learn the shape of it or hand it to software afterward.

Step 1: fix a prompt panel

Write down 20 to 40 questions a buyer would really type. Mix three kinds: category questions ("best tools for X"), problem questions ("how do I fix Y"), and comparison questions ("A vs B"). Add a handful of branded ones, but keep them a minority, because branded prompts flatter you and tell you nothing about discovery.

Freeze the wording. Changing a prompt changes the measurement, so edits should be logged and rare. Tag each prompt with a topic so you can see later which subject areas are strong and which are empty.

Step 2: record sources, not just mentions

A mention and a citation are different events. A mention means the brand name appeared in the answer text. A citation means a URL was listed as a source. You can be mentioned without being cited, and cited without being named, and each case points to a different fix.

For every prompt run, capture four things: whether you were named, whether any of your URLs were cited, which URL it was, and which other domains were cited alongside. The fourth column is where the insight lives. If the same two third-party sites show up in most answers, those are the pages you need to be present on.

Step 3: look at the page level

Domain-level counts hide the useful part. "We were cited 14 times" does not tell you that 11 of those were one old guide and the product pages never appear. Roll your log up by URL.

Two numbers per page help. How often is it cited across the panel, and in which prompts. A page cited for one oddly specific question is luck. A page cited across a topic cluster is a real asset, and the cluster's uncovered prompts show where to write next.

Step 4: compare with the same panel for rivals

Run your competitors through the exact same prompts and count the same way. Share of voice, meaning your mentions divided by all mentions among the compared brands, only makes sense when the denominator is stable. Keep the competitor list fixed too, and note any change in the log.

Step 5: separate noise from change

ChatGPT answers move around. Before you react to a drop, check whether it happened across several prompts or only one, whether it persisted for more than one week, and whether the cited sources shifted or only the order did. A one-week dip on a single prompt is noise. A multi-week decline across a topic is a signal.

When a signal appears, the next question is whether the page was actually reachable. That means looking at crawler activity: did OAI-SearchBot and ChatGPT-User fetch the page, and did they get an error? Citation tracking without crawl data leaves you guessing about causes.

What the by-hand version costs

Run 30 prompts, once a week, by hand: expect about 90 minutes to run and log them, longer to roll up by URL and competitor. That is fine for a month. By week six most people stop, which is the real problem with manual tracking. It is not that it is inaccurate. It is that it does not survive a busy quarter.

There is a second limit. A human running prompts sees one sample per prompt. ChatGPT may answer the same question differently each time, so a daily automated panel gives a steadier picture than a weekly manual one.

Where Promptwatch fits

Promptwatch does the logging for you. Monitors run your prompt panel daily per model, country and language, and the citation analytics report page-level and domain-level citations, with URL positions. If your question is "is this specific page being cited", the Page Tracker follows individual URLs, yours or a competitor's. Citation Trends goes further, classifying each citation by content type and source type, and showing for each page when it was first cited, how many days it took to ramp up, its peak rate, and its decay half-life.

Two parts of it map directly onto the routine above:

• The panel in step 1 becomes monitored prompts, tagged by topic and prompt type (organic, brand, competitor comparison).
• The competitor comparison in step 4 uses auto-discovered competitors and share of voice, so the denominator is not something you maintain in a spreadsheet.

Then there is the question tracking by hand cannot answer: why. Agent Analytics logs AI crawler visits in real time, IP-verified, with a citation rate per page and error tracking with explanations. If a once-cited page lost its citations, you can check whether the bots could still fetch it. Crawler logs start on the Professional plan, and Essential does not include them, which matters if diagnosis is the reason you are buying.

A reasonable way to start

Do the manual version for two weeks with 20 prompts. You will learn which of your pages get cited, which third-party domains dominate your topic, and how noisy the answers are. Then move the same panel into Promptwatch so the routine runs without you. The free Explore tier covers 10 prompts on ChatGPT only, which is enough to test the workflow on the questions you care about most. The monthly plans from Essential up include a 7-day free trial if you want to see the wider data first.

The recommendation is simple: keep the panel and the log structure from this post, and let the platform collect. Your hour a week should go into reading what changed and deciding what to publish, not into pasting URLs into a sheet. You can see how it works at promptwatch.com.