How to track your citations in ChatGPT without checking by hand

Most teams begin tracking ChatGPT citations the same way. Someone opens ChatGPT, types a question, scrolls to the sources, and screenshots the result. It feels like evidence. It isn't, because the same question asked an hour later, from another account or country, can produce different sources.

Tracking citations properly means turning that one-off look into a repeatable measurement. Here is the routine we set up in the first fortnight of an engagement.

Decide what a citation is for you

ChatGPT shows sources in a few ways. A link can appear inline, in a source list, or in a news block. Some answers name a brand with no link at all. Those are different events and they need different handling.

We separate three things from the start:

• Cited page: ChatGPT links a URL on your domain.
• Cited domain: any URL on your domain appears, which is useful for trend lines.
• Offsite mention: your brand is named in an answer or on a third-party page, with no link to you.

If you lump these together you'll get a number that moves for reasons you can't explain.

Build a prompt panel first

Citations only exist in relation to prompts. Before touching any tool, write down 30 to 50 questions your buyers ask. Split them into brand questions, category questions ("best X for Y"), and competitor comparisons. Our guide on choosing the prompts worth tracking covers the selection in detail.

Keep the panel stable for at least a month. If you swap prompts every week, a rise in citations might just mean you picked easier questions.

What a good tracker records per response

For each prompt run, you want the full answer text, the list of cited URLs in order, the model and country, and the date. With that you can answer questions that spot checks can't:

• Which of our pages gets cited, and for which prompts?
• Which domains are cited instead of us?
• Is our citation rate rising, flat, or falling across the panel?
• Do we show up on the first source, or are we buried?

Position matters. Being the third link under a long list isn't the same as being the primary source.

Look at pages, not just the domain

A domain-level citation count hides the story. Two pages may account for nearly all of your citations while twenty others get none. That's useful to know, because it tells you what ChatGPT considers worth quoting from you.

In Promptwatch, citation analytics gives page-level and domain-level views, URL positions, and a Page Tracker you can point at specific URLs, yours or a competitor's. Citation Trends goes further and classifies each citation by content type and source type. It shows the citation rate over time and, per page, when it was first cited, how many days it took to ramp up, its peak rate and its decay half-life. That last metric is a good prompt for a refresh: a page whose citation rate decays quickly is telling you something has gone stale.

Check the cause as well as the count

When citations drop, the reason is often upstream. Possibilities worth checking:

1. A crawler block was introduced. ChatGPT-related bots include GPTBot, OAI-SearchBot and ChatGPT-User, and each does a different job.
2. A page moved or changed URL and the cited address now errors.
3. A competitor published a stronger answer to the same question.
4. A third-party page that used to mention you changed.

Promptwatch's Agent Analytics shows IP-verified AI crawler logs in real time, with error tracking and failure explanations. Spoofed user agents are filtered out, which matters because fake bot traffic can make a page look fetched when it wasn't. The crawl-to-citation path gives a citation rate per page, so you can see whether a fetch ever led to a citation.

We wrote about the access side in can AI crawlers reach your site. Run that check before assuming the content is the issue.

Set a review rhythm

Citation data is noisy day to day. We look at it weekly and judge it monthly. The weekly view catches breakage, such as a sudden loss of a page that carried many citations. The monthly view is where you judge whether the content work is paying off.

A simple weekly check:

1. Open the citation view for the panel.
2. Compare this week with the previous period.
3. Note pages that dropped out or entered.
4. Check crawler errors for any page that dropped.
5. Log one action.

The Action Board in Promptwatch can hold that action. A Discovery agent investigates the data weekly and suggests up to five evidence-backed actions per run, and you can add custom ones with assignees and comments. It saves us from keeping a separate to-do sheet that goes stale.

Tie citations to visits

A citation that sends no one to your site is still a brand impression, but you should know which is which. Visitor analytics in Promptwatch records AI-referred traffic and conversions from ChatGPT and other engines, using a lightweight script, a GTM template or the WordPress plugin. Put the citation data next to it and you'll see which cited pages earn clicks and which are quoted but never visited.

That's the number clients care about once they've seen the mention counts.

What not to over-read

Avoid concluding anything from a single day. Don't compare citation counts across different prompt panels. Don't assume that being cited in ChatGPT means being cited in Perplexity or Google AI Mode, as each engine pulls from its own mix of sources. Track each model separately even when the dashboard offers a blended view.

The setup we'd use

For a team starting from nothing, the shortest path is Promptwatch: load the prompt panel, let it run daily per model and country, and watch the citation analytics and Page Tracker for the pages you care about. Connect crawler logs if your plan includes them (Professional and above for brands) and add the visitor script so every citation can be tied to a visit.

Do that for four weeks before you change any content. You'll then have a baseline, and every later edit can be judged against something real instead of a screenshot from a Tuesday afternoon.