LLM Citation Tracking Tool: AI Citations Tracking in ChatGPT
How we log ChatGPT footnotes for clients: mention vs citation vs self-citation, then Promptwatch for the source archive and the crawl path.
Clients who ask us for an LLM citation tracking tool are rarely asking for a GEO category tour. Someone on their side opened ChatGPT, expanded the sources under an answer, and saw a competitor URL or a listicle. They want to know if that happens on the questions that actually produce pipeline, and whether any of those footnotes point at pages we can edit. We treat that as a logging problem first. The measurement layer we run is Promptwatch. The page work is still ours.
This is the playbook, not a case study. We are not going to invent a client, a footnote share, or a traffic lift.
The question on the call
We ask them to paste two ChatGPT answers from last week, sources visible. One branded ("is [company] worth it"). One category ("best [category] for [use case]"). If they cannot produce those, we sit with them and run the questions live. ChatGPT has 820M+ weekly active users. Pretending the engine is a side channel is how teams delay the log for another quarter.
We then write three labels on a shared note, in plain language, because "visibility" is how this meeting goes wrong.
Mention: ChatGPT used the brand name in the paragraph.
Citation: ChatGPT attached a source URL.
Self-citation: that URL lives on a domain the client controls.
A typical messy answer looks like this. The model names the client in a shortlist of four. The footnotes are a review site, a YouTube demo from a partner, and a competitor's comparison page. That is a mention with zero self-citations. Celebrating the mention in a monthly deck is how a CMO thinks the site is "in ChatGPT" while every clickable source belongs to someone else.
The inverse happens on informational prompts. ChatGPT cites the client's glossary page and never names the product. That is a self-citation without a mention. Useful. Incomplete. We do not merge the two into one percentage for the slide.
How we log a ChatGPT answer
Explore is free: 10 ChatGPT prompts, one project, one seat. We use it in the first workshop so a skeptical partner can see the workflow without a purchase order. We do not run a live program on ten prompts.
Essential is $95 a month with a 7-day trial, 50 prompts, 6,000 responses, 200K visitor events. That is enough for a single brand that already knows which questions matter. Professional is $245 a month: 150 prompts, 2 projects, 25M crawler logs. We move a client to Professional when we need the crawl-to-citation path next to the ChatGPT log, or when they trade under two names. Kick-off is $199 a month for agency work with unlimited projects and prompts, 10 seats, 10,000 responses, 10M crawler logs. That is the plan we use when we are the ones operating the workspace across a roster.
Paid plans read the real ChatGPT UI, not an API sample. That matters because the footnote list in the product is what the customer's buyer sees. We will not pretend an API transcript is the same object.
What we actually open once the prompts are loading:
Citation analytics at page level and domain level. Page level is the URL we might rewrite. Domain level is the publisher we might need to earn, or the competitor who owns the footnote. Reddit and YouTube reports when those surfaces appear as sources. Offsite mentions when a cited article names the client without linking. Citation-type breakdowns so a news spike is not treated like a durable docs citation. Citation trends so we are not briefing off one Wednesday.
We do not screenshot ChatGPT by hand and file it as "tracking." Promptwatch is not sold to us as a screenshot product. We store the citation objects and the dates.
ChatGPT depth, then the other rooms
We start the log on ChatGPT because that is the engine in the brief. We do not leave it there.
Gemini (650M+ monthly active users) and Perplexity (22M+ monthly) reuse different source mixes on the same commercial question. We have sat in reviews where ChatGPT footnoted the client's docs and Perplexity footnoted a Reddit thread the client had never seen. If we only measured ChatGPT, we would have sent writers at a page that was already winning. Paid Promptwatch coverage includes those engines plus Claude, Grok, Llama, DeepSeek, Mistral, Copilot, and Google AI Overviews and AI Mode, still from the real UIs. Same citation objects. Same weekly note structure.
Self-citation is the line we protect in that note. If the share of ChatGPT sources on owned domains is low, the content ticket is "ship a page that can be the footnote," not "get mentioned more." If owned URLs already appear and the mention is missing, the ticket is positioning on that page, not another URL.
Crawl path and visits
On Professional (or Kick-off, when we are running the account), Agent Analytics shows whether ChatGPTBot, ClaudeBot, PerplexityBot, GoogleOther, or the Meta AI crawler fetched the URL we expected to be cited. Cloudflare, CloudFront, Fastly, Vercel, Netlify, Akamai, Google Cloud CDN, or custom HTTP. The crawl-to-citation path is the adult version of "did we publish." If the bot never arrived, we do not workshop adjectives.
Visitor analytics go on via a script or GTM. AI-referred sessions and conversions sit next to the citation list. A footnote that never produces a visit is still a source fact. It is not a pipeline fact. We refuse to add them into one vanity total.
Promptwatch is based in Amsterdam, founded April 2025. 4.7/5 on G2, 1,840+ brands and agencies. When a cited competitor page is simply more complete than ours, Content Agents can draft for Webflow or Framer with a review inbox. We still edit. We do not autopilot pricing claims.
What we tell them to buy, and what we will not promise
AthenaHQ from $295 a month is a reasonable Action Center if their problem is that SEO tickets die in Slack. Eight engines, GEO score. Not the citation archive we run. Scrunch from $250 a month annual can serve crawlers a different page than humans see. Pair it with a log if they already bought it. Profound's public Starter is $99 a month yearly and ChatGPT-only; Enterprise is around $40k a year and sold as marketing agents. We do not start a citation program there.
Otterly and Peec are mention-first tools. We skip them for this brief.
We will not attach a euro or dollar value to a ChatGPT mention. We will not claim Search Console generative reports cover ChatGPT. We will keep mention, citation, and self-citation on separate lines, start the log on ChatGPT, and expand engines before we greenlight a large content batch.
If you want the same setup, the platform is Promptwatch. We will still write the pages. The tool holds the footnotes.