Measuring AI traffic: how to tell if ChatGPT and Perplexity send you visitors

Your boss asks the reasonable question: “Is all this AI-visibility work actually bringing us anyone?” You open your analytics, look for traffic from ChatGPT, and find… almost nothing. A trickle from chatgpt.com, a few sessions from perplexity.ai, and a suspiciously large pile of “Direct” traffic you cannot explain. Here is the uncomfortable truth: AI referral traffic is real and growing, but it is systematically undercounted by standard analytics, and if you judge your AI strategy purely on the referral numbers you can see, you will badly underestimate what is happening. This is a guide to measuring it honestly — what the numbers show, what they hide, and how to read the whole picture.

The headline is this: for most sites, the visitors AI engines send you are a fraction of the influence AI engines have on you. People read an answer that mentions you and never click — but they remember the name, and some of them arrive later as a “direct” visit or a branded search. So AI shows up in your analytics in two ways: a small, visible stream of direct clicks from the AI tools, and a large, invisible wave of influence that surfaces as other channels. Measure only the first and you will conclude AI does not matter. Learn to see the second and the real story appears.

Where the visible AI traffic shows up

Start with what you can see, because it is worth seeing properly. In your analytics, AI tools appear as referral sources under their own domains: ChatGPT sends from chatgpt.com (and sometimes chat.openai.com), Perplexity from perplexity.ai, Google's Gemini from its own properties, and so on. To find them, filter your referral or source/medium reports for those hostnames, or build a segment that groups them into an “AI assistants” channel so you can watch it as one line over time. Most analytics platforms do not categorise these for you yet, so if you do not build the grouping yourself, the traffic gets scattered and looks like noise. Doing this is the baseline: it turns “I think we get some AI traffic” into a real, trended number you can report. It will usually be modest — and that is the point that trips people up, because modest-looking referrals coexist with large, invisible influence.

Why the numbers undercount reality

There are several reasons the visible referral number is lower than AI's real effect, and understanding them keeps you from drawing the wrong conclusion. The biggest is simply that AI answers are designed to satisfy the question in place — the whole user experience is “read the answer, do not click” — so a mention that shapes someone's opinion often produces no click at all. On top of that, referral data gets lost for ordinary technical reasons: some AI apps strip the referrer, some traffic comes through native apps that do not pass a clean source, and privacy settings and link handling can blank out the origin. The combined effect is that a meaningful slice of genuine AI-driven visits lands in your “Direct” bucket with no fingerprint at all. None of this means the influence is not real; it means the influence is real and the attribution is broken, which is a very different problem from “AI does not send traffic.”

The dark funnel, and why it matters most

The concept to internalise is the dark funnel: the stretch of a buyer's journey that happens where your analytics cannot see it. Someone asks ChatGPT for the best tools in your category, reads an answer that names you alongside two competitors, and forms an impression. They do not click. Three days later they type your name into Google, or your URL into the address bar, and convert. Your analytics proudly attributes that conversion to “branded search” or “direct,” and the AI answer that actually started the journey gets no credit whatsoever. For a growing number of businesses this dark funnel is where a large share of AI's value lives, and it is precisely the part a referral report will never show you. The mistake is to conclude it is not happening because it is not labelled. The sophisticated move is to measure it by its shadows.

How to measure influence you cannot directly track

You cannot put a tracking parameter on a sentence inside an AI answer, so you measure the dark funnel indirectly, through the traces it leaves. Three proxy signals do most of the work. First, watch your branded search volume over time — if more people are searching for your name, something is introducing them to it, and increasingly that something is AI answers. Second, watch your direct traffic, especially to deeper pages that people would not normally reach without knowing you; a rise there often reflects AI-driven awareness surfacing as untracked visits. Third — and this is the most direct handle on the whole thing — measure whether the AI engines actually name you for the questions that matter, because that is the input to the entire dark funnel. If your branded search and direct traffic are climbing while you are increasingly being named in AI answers, you are watching the funnel work even though you cannot click-attribute a single step of it.

Why being named is the leading indicator

Of those signals, the one worth building your measurement around is whether you get named, because it comes first in the chain and you can influence it directly. Branded search and direct traffic are lagging indicators — they move weeks or months after the awareness that caused them. Being named in the answer is the leading indicator: it is the cause, and the traffic is the effect. If you track only the downstream traffic, you are always reading yesterday's weather; if you track whether the engines name you, you see the front edge of the wave and can act on it. This is why we think the right primary metric for AI visibility is not referral sessions at all, but citation share — for the questions your buyers ask, how often are you named, and who is named instead of you. Get that rising and the traffic signals follow; obsess over the traffic signals alone and you are managing the symptom, not the cause.

A measurement setup that tells the truth

Put together, here is a setup that will not lie to you. Build an “AI assistants” channel grouping in your analytics so the visible referrals are tracked as one trended line rather than scattered noise — that is your floor, the influence you can prove. Then add the two proxy dashboards: branded search volume from Search Console, and direct traffic to non-homepage pages, both trended over months so you can see the shape rather than the jitter. Finally, and most importantly, track citation share on a schedule: ask the questions your buyers ask across the engines, log whether you are named and who is cited instead, and watch that line as your leading indicator. Read the three together. When citation share rises and, a few weeks later, branded search and direct follow, you have caught the whole mechanism in the act — and you can report it honestly, including the parts your referral report alone would have told you did not exist. We built OptiLink's AI-visibility tracking to own that leading indicator, because the clicks will always lag the citations, and the citations are the thing you can actually move.

Common mistakes when reading AI traffic data

A few errors come up again and again, and each one leads a team to the wrong decision. The first is judging the whole AI channel on raw session counts and concluding it is negligible — that ignores the dark funnel entirely and will talk you out of work that is quietly compounding. The second is the opposite over-correction: attributing every rise in direct traffic to AI without evidence, which is just as unrigorous and will embarrass you when someone asks for proof. The honest posture is to hold both truths — the visible number is a floor, the real impact is larger, and the way to bound the gap is the proxy signals moving together. A third mistake is measuring once and declaring victory or defeat; these channels are volatile and still maturing, so a single snapshot tells you almost nothing. Trend everything over months. A fourth is forgetting that AI traffic, when it does click through, often behaves differently — it tends to arrive later in the decision process, already informed, so it can convert at a higher rate even in small numbers. Segment it and look at qualified actions, not just sessions, or you will undervalue the visitors AI does send.

What to report upward, and how to frame it

When you present this to someone who controls budget, resist the temptation to show only the clean referral number, because it understates your case and sets a trap for later. Frame it as a system. Lead with the leading indicator — “here is how often the AI engines name us for the questions our buyers ask, and here is the trend” — because that is the thing your work actually moves and the thing that predicts everything downstream. Then show the corroborating signals: the visible AI referral line as a provable floor, and the branded-search and direct-traffic trends as the shadow of the dark funnel. Be explicit that attribution is imperfect and that you are reading a pattern across signals rather than a single tracked path, because honesty about the method is what makes the conclusion credible. Presented this way, AI visibility stops being a vague “we should probably do something about AI” and becomes a measured channel with a leading indicator you can influence and lagging indicators that confirm it worked. That is a story a budget-holder can act on, and it happens to be the true one.

Setting up an AI-assistants channel

The first practical step is to make the visible portion of AI traffic actually visible in your analytics, which by default it usually is not. Most analytics tools do not yet recognise AI assistants as a distinct channel, so referrals from them scatter across “referral” and “direct” and never add up to a number anyone looks at. The fix is to define a custom channel grouping that captures the known AI hostnames — the ChatGPT, Perplexity and Gemini properties and their variants — so they roll up into one line you can trend over time. It is a small configuration task with an outsized payoff: without it you are flying blind on even the measurable part of the channel, and with it you have a provable floor to build the rest of your measurement around.

The branded-search baseline

Branded search is one of your most reliable windows into the dark funnel, but only if you have a baseline to compare against. Before you can read a rise in people searching your name as evidence of AI influence, you need to know what your normal branded-search volume looks like, controlled for the obvious confounders like paid campaigns, PR spikes and seasonality. Establish that baseline first, then watch for sustained lifts that do not correspond to any other cause. A steady climb in branded search while your other channels hold flat is one of the clearer fingerprints of AI answers introducing you to people who later come looking — but the signal only means something against a baseline you took the trouble to establish.

Reading server logs for AI crawlers

A different and underused data source sits in your own server logs: the visits from the crawlers the AI engines use to retrieve content. While this does not tell you about end-user traffic, it does tell you whether and how often the engines are actually fetching your pages, which pages they favour, and whether changes you make affect their crawling behaviour. For a team serious about AI visibility, watching crawler activity in the logs is a leading-leading indicator — evidence that the engines are paying attention to your content at all, upstream even of citation. It requires some technical setup and careful interpretation, but it is first-party data the engines cannot obscure, which makes it a valuable complement to the softer downstream signals.

Why conversion rate is the tell

If you track only one thing about the AI traffic you can see, make it conversion quality rather than volume. The recurring finding is that AI-referred visitors, though fewer, often arrive later in their decision and better informed, and so convert at a notably higher rate than the raw session count would suggest. This matters for two reasons. It corrects the instinct to dismiss a small session number as negligible, because a small number of high-intent visitors can be worth more than a large number of casual ones. And it gives you a value-weighted way to talk about the channel: not “we got two hundred sessions” but “we got two hundred sessions that converted at three times our site average.” Measuring the quality, not just the quantity, is what reveals the channel's real contribution.

Benchmarking against competitors

Your own citation share means more when you know how it compares. Because the engines name a small set of sources for any given question, AI visibility is inherently competitive: the relevant question is not just whether you are named but who is named instead of you, and how that is changing. Periodically running the questions your buyers ask and recording not only whether you appear but which competitors do gives you a comparative benchmark that a bare count of your own mentions cannot. It tells you whether you are gaining or losing ground relative to the specific rivals you care about, and it surfaces the questions where a competitor owns the answer and you are absent — which are precisely the gaps worth closing first. Measurement without a competitive frame tells you how you are doing; with one, it tells you how you are doing against the people you are actually competing with.

A monthly measurement routine

All of this only works if it is a routine rather than a one-off. Volatile, maturing channels tell you almost nothing from a single snapshot, so the discipline is to measure the same things the same way on a regular cadence and watch the trends. A workable monthly routine: pull the visible AI-referral channel, re-run your core buyer questions across the major engines and record your citation share and who else is named, check branded search and direct against their baselines, and note the conversion quality of the AI-assisted visits. None of these steps is heavy on its own, and together they produce a coherent, trend-based picture that improves every month you keep it up. The value is in the consistency: the same measurements, repeated, are what turn noise into a signal you can actually manage against.

Tools, and their limits

A growing crop of tools promises to track AI visibility for you, and some are genuinely useful for automating the tedious parts — running your questions across engines on a schedule, logging citations, flagging changes. But it is worth holding them to the same scepticism as any measurement product. No tool has privileged access to how the engines actually choose sources, so all of them are, at bottom, sampling the same answers you could sample yourself and inferring from the outside. Use them to save time and add consistency, not as oracles, and be wary of any that present their numbers with more certainty than the underlying opacity of the engines can support. The fundamentals of good measurement — a baseline, a consistent method, a competitive frame, honest caveats — matter more than any particular tool you layer on top.

Frequently asked questions

How do I find AI traffic in my analytics?

Filter your referral or source/medium reports for AI hostnames like chatgpt.com, chat.openai.com, perplexity.ai and Gemini's properties, and group them into an 'AI assistants' channel so you can trend them as one line. Most analytics tools don't categorise these automatically yet.

Why is my AI referral traffic so low?

Because AI answers are designed to satisfy the question in place, so a mention that shapes opinion often produces no click. Referrers also get stripped by some apps and privacy settings, so genuine AI-driven visits frequently land in your 'Direct' bucket with no source attached.

What is the dark funnel in AI search?

The part of a buyer's journey analytics can't see — someone reads an AI answer that names you, doesn't click, then returns days later via branded search or direct and converts. The AI answer that started it gets no credit, so you measure it indirectly through proxy signals.

What should I actually measure for AI visibility?

Citation share — how often the engines name you for the questions your buyers ask, and who is named instead. It's the leading indicator that causes the downstream branded-search and direct-traffic rises, and the one thing you can directly influence.

How do I prove AI visibility is working to my boss?

Report it as a system: lead with citation share (the leading indicator you move), then show the visible AI referral line as a provable floor and the branded-search and direct-traffic trends as the dark funnel's shadow. Be explicit that it's a pattern across signals, not a single tracked path.