How to Track Your Citations in ChatGPT: A Weekly Workflow an Agent Can Run
A practical weekly routine for tracking ChatGPT citations, from choosing prompts to reading page-level data and queuing fixes.
Tracking ChatGPT citations by hand works exactly once. You type a question, scan the sources, feel informed, and then never repeat it the same way. The second attempt uses different wording, the third happens on a different day, and by week four you have a folder of screenshots that cannot be compared with each other.
A repeatable workflow fixes that. The version below is built so that most of it can be run by software on a schedule, with a human reviewing the result. It has five steps, and each one produces something you can check.
Step 1: Fix the prompt set
Citation tracking only makes sense against a stable list of questions. Start with 30 to 50 prompts that real buyers would ask. Split them into three types: organic questions about your category, brand questions that name you, and competitor comparisons. Each type behaves differently, so keep them labeled.
Do not write the prompts as keywords. Write them as full questions, the way a person would type them into a chat box. If you are short on ideas, mine your Search Console queries, the People Also Ask boxes for your main topics, and the pages on your own site that attract questions.
In Promptwatch, the Prompt Explorer suggests prompts from Search Console, ranking keywords, People Also Ask, or any page you point it at. Prompts can be tagged by topic and type, and each carries search volume and a difficulty score, so you can start with prompts that are worth winning.
Step 2: Run the panel the same way every time
Consistency beats volume. The same prompts, the same country, the same language, on a fixed cadence. Promptwatch runs prompt panels daily per model, country, and language, which removes the human drift that wrecks manual checks. If you track ChatGPT only, the free Explore tier covers 10 prompts and 310 responses on ChatGPT, which is enough to learn the shape of the data before you pay for anything.
One detail matters here. ChatGPT often runs several background searches for one question, called query fan-outs. The sources it cites come from those searches, not from your original wording. Tracking the fan-outs per model tells you which phrasings your pages need to match.
Step 3: Read citations at the page level
A domain-level count says your site was cited 40 times. It does not say which page earned the citations or which page was ignored. You want a table with one row per URL.
Promptwatch's citation analytics report page and domain citations with URL positions, and Page Tracker lets you pin specific URLs, yours or a competitor's, and watch them over time. Citation Trends classifies every citation by content type and source type, and for each page it reports first cited, ramp-up days, peak rate, and decay half-life. That last figure is the useful one for planning: a page whose citation rate decays quickly needs a refresh sooner than one that holds.
Step 4: Check the access side
If a page you expect to be cited never shows up, the cause is often boring. The crawler never fetched it, or fetched it and hit an error. ChatGPT uses three user agents you should know by name: GPTBot, OAI-SearchBot, and ChatGPT-User. There is no bot called "ChatGPTBot," so a robots.txt rule written for that name does nothing.
Agent Analytics in Promptwatch shows real-time, IP-verified crawler logs, so requests from spoofed user agents are filtered out. It also follows the crawl-to-citation path and reports a citation rate per page, plus error tracking with failure explanations. Crawler logs start on the Professional plan. The Essential plan has none, which is worth knowing before you buy.
Step 5: Turn the data into a queue
A report that nobody acts on is decoration. The final step is to convert findings into dated tasks. A good weekly queue has no more than five items, each with an owner and evidence attached.
The Action Board in Promptwatch does this directly. A Discovery agent investigates your data on a weekly schedule and suggests up to five evidence-backed actions per run. You can add custom actions, assign them, comment on them, and receive a weekly digest by email. If you want the loop closed further, the Content Agent on the Business plan plans, writes, and publishes against visibility gaps, with a review inbox in front.
A sample week
Monday: the panel has already run through the weekend. Open the citation view and sort by change since last week. Note any page that lost citations and any new page that gained them.
Tuesday: check crawler logs for the pages that dropped. Was there a spike of errors, a blocked path, or a drop in fetches?
Wednesday: open the Action Board. Accept, reassign, or decline the suggested items. Write the one you think is most important in the team channel.
Thursday and Friday: ship the fixes. Refresh the content, correct the robots rule, repair the broken path.
You can hand part of this to a scheduled agent. Scheduled Chats deliver an analyst-style summary in the app or in Slack, so Monday's reading step arrives as a message instead of a task.
What not to track
Skip vanity metrics. A single "visibility score" is fine as a trend line, but never as the goal. Skip one-off screenshots. Skip prompts you would never want to win. And do not compare numbers across tools without checking how each defines a mention, because definitions differ.
The recommendation
If you only do one thing from this post, run the same prompt set on a schedule and read citations per page, not per domain. For that, Promptwatch is the most direct way to get the whole chain in one place: prompts, page-level citations, crawler logs, and an action queue. Start on the free Explore tier to confirm the workflow suits you, then move to Professional when you need the crawler logs.