How I track whether AI is citing my content (and what I do when it isn’t)

AI citation tracking is the measurement half of AEO/LLMO (answer engine optimisation and large language model optimisation, which is structuring content so AI tools can find, extract and cite it). I wrote about the strategy side in my AEO/LLMO framework and barely touched the uncomfortable question underneath it: how do you know any of it is working? Restructuring a page with an extractable summary and a self-contained FAQ is the easy part, while proving an AI tool cited you because of those changes is hard, and nobody has a perfect answer, me included. What I built gave me enough signal to make editorial decisions and show leadership the work was paying off, and the dedicated tools that have appeared since haven’t changed the principles behind it.

Why doesn’t traditional SEO work for AI citation tracking?

Traditional SEO has a clean feedback loop (optimise a page, watch rankings, clicks and traffic move) and AI answers break it, because the user often gets what they need without visiting you. When ChatGPT or Perplexity answers from your content, there’s frequently no click and no referral, so GA4 has no idea it happened.

The click is disappearing on Google too: SparkToro’s 2026 zero-click study found 68% of US Google searches ended without a click in early 2026, and Pew Research Center found people clicked a traditional result on 8% of visits when Google showed an AI summary, against 15% when it didn’t. Neither figure counts the questions that never reach Google because someone asked an AI tool first, and Semrush’s 2025 study projected that, for digital marketing and SEO topics, AI search could send more visitors to websites than traditional search by early 2028.

So if you judge AEO/LLMO purely on organic traffic, you can be cited more than ever while the dashboard stays flat or drops, because fewer people click once the AI has given them the summary. I’ve written about what that looks like from the traffic side in organic traffic decline, and the measurement fix is a different set of signals.

What did I build in Amplitude for AI citation tracking?

I built three proxy signals in Amplitude at a B2B SaaS company serving 200+ higher education institutions, because Amplitude handled event-level tracking and custom funnels more precisely than GA4 did at the time.

Referral monitoring for AI tools. Some visits from AI tools carry a referrer that identifies where they came from, and Perplexity, ChatGPT and Microsoft Copilot each leave different ones. I flagged those visits as their own segment, separate from organic search, direct and social. It only ever catches a fraction of AI traffic, since most AI answers never send anyone to your site, so treat it as a floor. GA4 has since added a native AI Assistant channel, and I’ve covered how to track AI traffic in GA4 properly, including the sessions that still land in Direct with no referrer at all.

Zero-click behaviour proxies were the second signal: patterns that suggest someone arrived with a narrow question an AI answer had already set up. I watched for short sessions that landed on a specific deep page (never the homepage), read it and left; high bounce rates on content that was strong on every other metric; and visits to FAQ sections and product documentation that arrived by direct URL entry rather than through site navigation. None of those proves a citation, but together they describe people who already knew what they wanted and came to check or expand on an AI tool’s answer, which looks different from organic discovery, where people arrive on broader queries and browse more widely.

Content structure correlation was the most useful of the three. Whenever I changed a page’s structure (adding an extractable summary, restructuring the FAQ, attributing key claims), I logged the date and watched that page’s referral and engagement patterns over the following weeks. When the same shift showed up across several restructured pages, I treated it as signal, while knowing it was correlation and never proof. Freshness helps here too: Ahrefs’ analysis of 17 million citations found AI assistants cited content 25.7% fresher than what ranked in organic results, with Google’s AI Overviews the exception, so a structural update also earns a recency boost on most engines.

Proxy signal	What it suggests	What it can’t prove
AI referral segment	An AI tool sent a visitor	Anything about the answers that sent nobody
Short direct visits to deep pages	Someone arrived with a question an AI answer set up	That an AI was involved at all
Structure correlation	A restructure changed how a page gets used	That the change caused a citation

What signals indicate AI systems are pulling from your content?

No single metric says “an AI cited this page”, so I looked for a cluster of signals moving together:

• A changing referral mix. I tracked AI referrals as a share of total traffic rather than raw visits, because the raw figures early on were too small to mean anything by themselves.
• Direct traffic to deep pages that rises after a structural change, with no matching change in internal linking or promotion, can mean an AI tool pointed people at that page.
• Search Console shifts. When AI tools cite your content, some people follow up by searching for your brand or for long-tail phrases that match your extractable summaries, so a rise in branded and very specific queries is worth watching. I’ve since written up how to measure AI visibility without analytics, where branded search lift and self-reported attribution do most of the work.
• Sales and prospect feedback is the least technical signal and sometimes the most concrete. Prospects repeating a claim they “read somewhere” that matches your content closely, without being able to name the source, have often met it through an AI tool, and I logged those mentions through the same sales feedback process I used for pipeline attribution.

What does the iterative feedback loop look like?

When the signals suggested a page wasn’t getting picked up, I ran it through four steps and repeated them until something moved.

First came diagnosis against my AEO/LLMO framework: do the first 150 words answer the core question, are the H2s questions people ask, does the FAQ stand on its own, are claims specific and attributed? The problem was almost always obvious once I read the page as an AI extracting from it would, rather than as a human reader.

Then I restructured: rewrote the opening so it could be lifted whole, turned headings into questions, attributed key claims and made each FAQ answer self-contained. Rewording a sentence for style doesn’t change whether AI tools cite it, while restructuring a paragraph so its core claim survives without the surrounding context does, and I’ve laid out those tactical patterns in detail and in a checklist for ChatGPT, Perplexity and AI Overviews.

Monitoring ran for two to four weeks per page. If the referral mix or the engagement pattern shifted, the change probably worked; if nothing moved, the page went back to diagnosis.

Last, I documented the pattern every time a structural change produced a measurable shift, which built a library of the changes that correlated most strongly with AI visibility. That library decided how I structured new content from the start, so fewer pages needed fixing later, and it fed the editorial standards every contributor worked from, human or AI. Dedicated platforms now automate much of this loop; mine was manual because those tools didn’t exist yet.

What should you measure for AI visibility now?

Start with a fixed set of buyer prompts run across engines on a schedule, report how often you appear as a percentage of runs, and keep the proxies as supporting evidence.

1. Prompt tracking, as a visibility percentage. Build a list of the questions your buyers ask AI tools, run them across ChatGPT, Perplexity, Claude, Gemini and Google’s AI Overviews, and log whether you’re named, whether you’re cited as a source, and who else is. At a market research company, the first baseline from 20 buyer prompts had the company named in 38% of answers and cited as a source in 37%. I’ve written up my prompt research process step by step.
2. Treat AI answers as a sample, never as a ranking. In a SparkToro study with Gumshoe.ai, around 600 volunteers ran the same 12 prompts through ChatGPT, Claude and Google’s AI nearly 3,000 times, and the odds of getting the same list of brands twice were under 1 in 100. Rand Fishkin’s verdict, as Search Engine Journal reported it:

“any tool that gives a ‘ranking position in AI’ is full of baloney.”

3. AI referrals in GA4, using the native AI Assistant channel plus your own filters, which is the cheapest place to start if you already run GA4.
4. Share of voice in AI answers (your mentions as a share of all brand mentions across your target prompts) gives you a competitive benchmark that rank tracking never could.
5. Citation rate by page is the most actionable, because a page that never gets cited goes straight back into the diagnose and restructure loop.

Dedicated platforms such as Profound, Otterly.AI and Gauge now automate prompt tracking across engines, and they save time once the prompt list is right, but before that they just automate the wrong questions.

How precise can AI citation tracking get?

Not very, and anyone selling precision is selling something else. AI tools give different answers to the same prompt, so every visibility figure is a sample, and it’s only useful averaged across runs and compared month to month on the same prompts, the same engines and the same day. Every metric in this piece, from my manual Amplitude setup to the dedicated platforms, involves inference rather than certainty.

That’s still better than the alternative: a dashboard that tracks organic click-through rates beautifully while ignoring everyone who never clicked because an AI already answered them, which is precise and useless. The teams building baselines now will have the pattern library and the fluency to act fast when the tools improve, and as I argued in why AEO won’t save your content strategy on its own, the advantage comes from pairing structural work with measurement, because a team doing only one of them is guessing.