Why You Can't See Your AI Search Traffic

The problem: AI is answering your prospects, and you can't see it

A growing share of insurance shoppers no longer type questions into Google. They ask ChatGPT, Perplexity, or Copilot — "how much is a Medigap plan in Sioux Falls?", "can I get life insurance with diabetes?" — and get a written answer with a handful of source websites listed beneath it. Those source mentions are citations, and they are how a business gets found in AI search: the AI either names your website as a source, or it names someone else's.

Here's what makes this a business problem rather than a curiosity: it's invisible. When you lose a Google ranking, you can see it happen. When an AI engine answers a hundred prospects this week using a competitor's page, nothing anywhere in your reporting even hints that it happened. You can't fix — or take credit for — what you can't see.

Why the tools you already use can't see it

The tools an agency already has were all built for the old kind of search, and each one misses AI search for a specific reason:

The result: the fastest-growing way customers find businesses has no dashboard. That's the gap this guide closes.

How we found the fix: we had the same blind spot

We run five insurance websites of our own, and this year we rebuilt them around content designed to be cited by AI engines — hundreds of pages that answer one question each. Which immediately raised the obvious question: is it working? We had no way to know. Same blind spot as everyone else.

The breakthrough came from an unglamorous place: Bing Webmaster Tools, a free dashboard most agencies have never opened. Buried in it is a report of how often your pages are cited in Copilot and ChatGPT answers — and one of our sites was showing 60–70 citations per day that appeared in no other tool we owned. That told us two things. First, AI search was already finding us more than we knew. Second, the measurement data exists — it's just scattered: one piece lives in Bing's report, one piece can be pulled from the AI engines' own APIs, and one piece has to be captured on your own website. No single product connects them, so we connected them ourselves. That three-part assembly is the fix.

The fix: three tools, one for each missing answer

Everything you'd want to know about your AI search performance reduces to three questions, and each of the three tools exists to answer exactly one of them:

Detection, volume, payoff. Miss any one and you're guessing: without Tool 1 you don't know why numbers moved; without Tool 2 you only see the questions you thought to ask; without Tool 3 you can't connect any of it to revenue. The next three sections take each tool in turn.

Tool 1 — the daily question test: are we in the answers?

What it is: a small script that asks the AI engines the same ten to twenty questions every day — the exact questions a ready-to-buy prospect asks — and writes down which websites each engine names in its answers.

Its job in the fix is detection. Rankings had rank trackers; AI answers have nothing — so this recreates that, through the engines' own APIs (OpenAI's returns the web sources it used; Perplexity's returns its citation list). One dated log row per question per day. The day your Medigap page starts being cited for "does South Dakota have a Medigap birthday rule," the log shows it — and on every day you're not cited, it shows who is, which is competitive intelligence you'd otherwise pay a subscription service five figures a year for. Consistency is the method: same questions, every day, so a change in the log means the answer changed, not your wording. (Our monitoring-loop post covers the weekly review ritual that sits on top.)

Tool 2 — Bing's citation report: how often are we shown?

What it is: a free report inside Bing Webmaster Tools showing how often your pages are cited in Copilot and AI answers — the report where we first discovered our own 60–70 citations a day.

Its job in the fix is volume. Tool 1 samples the questions you chose; this counts everything real people actually asked where your pages appeared — including the thousands of phrasings you'd never think to test. It works because ChatGPT's web search runs substantially on Bing's index, making this the closest thing to an "impressions" report that exists in AI search. Setup is minutes: verify your site (you can import it straight from Google Search Console), submit your sitemap, and turn on IndexNow so new pages are pushed to the index instead of waiting to be discovered. If you take a single action from this guide, take this one.

Tool 3 — the visitor counter: is it sending us customers?

What it is: about fifteen lines of code on your own website that count one thing — visitors who arrived from an AI chat. Every browser reports which page a visitor came from (the "referrer"); this checks that one field, and if it matches an AI engine like ChatGPT or Perplexity, it records three facts to your own server: which site, which page, which engine. For every other visitor it does nothing at all.

Its job in the fix is the payoff: connecting citations to actual humans. Citations are visibility; this is the cash register — which engine sends people, and to which pages. It stores no cookies, no IP addresses, no personal data of any kind, which is not a detail but the design (next section). Ours, verbatim, free to copy:

<script>
// AI-referral counter: counts visits arriving from AI engines. Cookieless and
// PII-free — sends only site id, path, and the referrer's hostname.
(function(){try{var r=document.referrer;if(!r)return;var h=new URL(r).hostname;
var E=["chatgpt.com","chat.openai.com","perplexity.ai","www.perplexity.ai",
"copilot.microsoft.com","gemini.google.com","claude.ai","you.com",
"search.brave.com","duckduckgo.com"];
if(E.indexOf(h)<0)return;if(!navigator.sendBeacon)return;
navigator.sendBeacon("https://YOUR-COLLECTOR-ENDPOINT/hit",
JSON.stringify({s:"yoursite",p:location.pathname,r:h}));
}catch(e){}})();
</script>

The receiving end is deliberately boring: a tiny service that appends each hit to a log — timestamp, site, page, engine, nothing else — and serves totals by day, by engine, and by section of your site.

Why insurance sites need the consent-safe version

If you sell Medicare, ACA, or health coverage, you've watched the OCR tracking guidance, the FTC's GoodRx and BetterHelp fines, and the wiretapping suits — and many of you correctly stripped analytics off your health pages, trading measurement for compliance. Our scan of 1,286 agency websites found nearly half of health-topic sites with lead forms still running client-side tracking; on those pages the right answer is usually removal. Tool 3 is built for exactly that constraint: because it stores nothing about the visitor, it can run on the pages where the defensible analytics list is short — measurement without the liability you just removed.

What the three show together

Connected, the three tools draw the funnel that has never existed for AI search: named in answers (Tool 1) → shown at volume (Tool 2) → visitors arriving (Tool 3). Content decisions stop being faith-based. Which answer pages earn citations. Which money questions you're losing, and to whom. Which engine sends humans who actually request quotes. Per page, per engine, per day — for pennies of API cost and code your pages don't feel.

Start with a baseline: our own day-zero numbers

One discipline makes all of it credible: start the log before you publish the content it will measure. We practice this. The day we turned our own question test on, we were cited on zero of fourteen questions — and the log showed the engines citing the state legislature's statute pages and the NAIC's own PDFs, the primary sources our pages are built on. That's not failure; that's the starting line, dated. The next day, Perplexity cited us for the first question. When your flip comes, you'll have its date, its question, and its engine — evidence a principal or a carrier will actually believe, instead of a marketing anecdote.

Questions agents ask

Doesn't Google Analytics already show traffic from ChatGPT and Perplexity?

Partially, and only where you can run it. AI-engine referrals arrive as ordinary referrer traffic, so most GA4 setups bury them in the referral bucket unless someone builds segments for each engine domain. Worse, on health-topic pages many agencies have correctly removed GA4 entirely for compliance reasons, and consent banners drop a large share of the sessions that remain. A first-party, cookieless counter measures those pages without reintroducing the liability.

What does Bing Webmaster Tools have to do with ChatGPT?

ChatGPT's web search runs substantially on Bing's index, and Bing Webmaster Tools now reports how often your pages are cited in Copilot and AI answers. It is the closest thing to an 'impressions' report that exists for AI search, it is free, and it only requires verifying your site — which most agencies have never done.

How many questions should the daily question test track?

Ten to twenty is plenty. Use the questions a ready-to-buy prospect actually asks — the ones your answer pages target — plus one or two brand and category queries. The value comes from asking the same fixed set every day, so a change in the log means the answer changed, not your question.

Is the visitor counter compliant on Medicare and health pages?

It is designed for exactly that constraint: it sets no cookies, stores no IP address, no user agent, and nothing about the visitor — only that some visit arrived from a given engine to a given page. That is categorically different from an advertising pixel that transmits a visitor's identity to a third party, which is where the OCR guidance, FTC actions, and wiretapping suits concentrate. As always, run your specific setup past your counsel.

Why measure before publishing new content?

Because the before-and-after is the proof. A dated log that shows zero citations on day zero, then the first citation for a specific question with a specific engine on day N, is evidence your principals and carriers will believe — and it tells you which content earned the flip.