How to track citations and sources in AI answers, and which tools to use
A practitioner guide to tracking citations and sources in AI answers: what to measure, how to read it, and the tooling we run client programs on.
Clients rarely ask us "are we cited?" in those words. They ask why a competitor keeps getting named when they have the better product. Nine times out of ten the answer sits in the sources the AI engine pulled from, and you only see those if you track citations on purpose.
Here is how we do it, and the tools that make it practical. One scope note first: this is about watching which sources AI engines cite for your buyer questions. Verifying whether a reference inside a single AI-written document is real is a separate job, and we do not cover it here.
Mentions and citations are different data
A mention means the brand name appeared in the answer. A citation means the answer linked to or leaned on a specific page. You can be mentioned without being cited, and the reverse. Mentions tell you about reputation. Citations tell you where the engine found its evidence, which is the part you can act on.
Most simple trackers report mentions and stop. For client work we need at least four layers:
- Which URLs get cited, and in what position in the answer.
- Which domains dominate the source mix, including Reddit, YouTube and third-party review pages.
- Whether a page keeps its citations over weeks or fades.
- Whether the AI crawler could fetch the page at all.
How we set up the tracking
We start from the client's real buyer questions, not a keyword list. Sales calls, support tickets and the questions prospects ask in discovery tell us more than a tool's suggestions. We write a first panel of 30 to 50 prompts, grouped by topic and intent, and add the clearly branded and competitor comparison prompts separately so they do not distort the picture.
Then we run them daily across the engines that matter for that client. In Promptwatch, monitors run prompt panels per model, country and language, and a brand book decides what counts as a mention, which matters more than people expect once a brand has aliases and product names.
Reading the citation reports
The first report we open is the domain view. It shows who the engines trust for the topic. If two forums and a comparison site make up most of the sources, no amount of polishing the client's homepage will change the answer. We plan outreach, honest participation and content for those places instead.
The second is the page view. Promptwatch's citation analytics list page-level and domain-level citations with URL positions, and the Page Tracker lets us follow a specific URL, including a competitor's. When a competitor page is cited where ours is absent, we read it side by side and ask what it answers that ours does not.
The third is Citation Trends. It classifies each citation by content type and source type and shows first cited, ramp-up days, peak rate and decay half-life per page. That last measure changed how we plan refreshes. A page that peaks and decays quickly needs a different update schedule from one that holds.
We also look at offsite mentions, where the brand is named on a third-party page without a link. Those are often the easiest relationship wins, because the author already likes the brand.
The crawl check most teams skip
When a page is not cited, the lazy explanation is "the content is not good enough." Sometimes it is. Often the page was never fetched, or the fetch failed. Agent Analytics in Promptwatch gives IP-verified, real-time AI crawler logs, so spoofed user agents are filtered out. We can see GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot and the rest, a crawl-to-citation path with a citation rate per page, and error tracking with failure explanations.
Crawler logs start on the Professional plan ($245/mo). The Essential plan ($95/mo) has citation analytics but no logs, so we flag that to clients who ask which tier to start on. If your site sits behind Cloudflare, Vercel, Fastly or a similar CDN, the integrations are documented and the setup is usually a short job.
Tying citations to traffic
A citation that sends no one to your site is nice for brand awareness and useless for pipeline planning. Visitor analytics in Promptwatch tracks AI-referred visits and conversions through a lightweight script, a GTM template or the WordPress plugin. We use it to decide which cited pages deserve more investment, and which are cited often but never visited.
Other tools we have seen
We do not pretend the market is empty. AthenaHQ, Scrunch AI and Profound all report citations in some form, and Ahrefs and Semrush have added AI modules for teams that already live in them. They are reasonable choices for specific situations, such as an enterprise agent program or a team that wants one login. What we have not found elsewhere in one place is the join between crawler logs, citation data and visits. For our own programs that join is the deciding factor.
What no tool can promise
Citation data shows where the evidence is. It does not guarantee a recommendation, and answers vary from run to run. We report trends over several weeks and say so in every client report.
A first week plan
Day one: write the prompt panel and set the brand book. Day two: connect crawler logs if the plan allows. Days three to five: read domains, pages and gaps, and pick three fixes. After that, change one thing at a time so the next report means something. You can try the free Explore plan or a 7-day trial on a paid Promptwatch plan to see your own source mix before committing to anything.