Tracking AI chatbot citations

This is the most direct method and also the most tedious. Pick five to ten questions your target customer would plausibly ask, phrased the way a person types into a chat box rather than the way they'd type into Google. Run each one in ChatGPT, Perplexity and Gemini, and read the answer looking for three things: is your site named, is your site linked, and is your site's information used without any credit at all.

Perplexity is the easiest of the three to check because it shows numbered source citations by default, so you can see at a glance whether you're source [2] or absent entirely. ChatGPT only cites sources when it's using its browsing/search mode, which depends on the query and sometimes on which model tier is answering, so the same question can come back with links one day and a flat, uncited answer the next. Gemini behaves similarly, citations appear inconsistently depending on whether it decided to ground the answer in a live search.

Log the results in a spreadsheet: date, engine, prompt, cited (yes/no), linked (yes/no), which competitor showed up instead if you didn't. Re-run the same prompts monthly. A single check tells you almost nothing, because these answers are not deterministic, ask the same question twice in the same session and you can get a different set of sources. What you're building over months is a trend, not a snapshot.

When someone clicks a link that ChatGPT or Perplexity gave them, that click carries a referrer, and it does show up in standard web analytics, this part is genuinely useful and underused. In GA4, look at Acquisition reports for source/medium and filter or search for perplexity.ai, chatgpt.com, gemini.google.com, copilot.microsoft.com and bing.com/chat. Traffic from these will typically bucket into 'referral' unless the platform strips its own referrer, which some do inconsistently.

Two caveats worth taking seriously. First, a meaningful share of AI-assisted traffic arrives with no referrer at all, if someone reads an AI answer, then opens a new tab and types your name into Google or goes to your site directly, that visit shows up as direct or organic, not as AI referral, and there's no way to separate it out after the fact. Second, referrer strings vary and change without notice as these platforms update their apps, so a referral segment you build today can quietly under-count in six months. Check it periodically against a fresh list of hostnames rather than trusting a saved report indefinitely.

If you have real traffic volume, a rising trend in AI-hostname referrals, even a small one, is a more trustworthy signal than manual prompt testing, because it's actual users clicking through rather than your own occasional check. If you have low overall traffic, the numbers will be too thin to read anything into for months.

A citation with a link is the best outcome, but AI answers frequently name a business or describe it without linking anything, especially in conversational follow-up turns rather than the first search-grounded answer. That's still worth knowing about, because it means the model has some representation of you even if it isn't sending traffic yet.

The practical version of this, short of expensive enterprise monitoring platforms, is folding AI engines into whatever brand-mention routine you already run. If you already Google your own business name periodically, add the same query to ChatGPT and Perplexity. Ask direct questions like 'who is [your business] and what do they do' and read how the model describes you, since that description is effectively what it will hand to a prospective customer who asks a related question, and if it's wrong or outdated, no one will correct it for you.

This space is genuinely immature, and pretending otherwise sets people up for bad decisions. There's no share-of-voice metric with a defined denominator, no impressions count, no click-through rate you can trust the way you'd trust one from Google Search Console. Two different manual checks of the same prompt can disagree. Two different AI-visibility tools can report different citation counts for the same domain because they're built on different sampling methods and different assumptions about what counts as a citation.

What you actually have is directional evidence from three imperfect sources: manual spot checks that sample a moving target, referral data that undercounts because most AI-influenced clicks leave no trace, and mention monitoring that tells you about model awareness but not about outcomes. Used together over months, they tell you whether things are moving in a useful direction. None of them, alone or combined, gives you a number you can put confidently on a slide and defend. If a vendor sells you one that sounds too precise, ask what it's actually sampling and how often it's validated against the real engines, and be prepared for an unsatisfying answer.