For the past several weeks I have been running a live experiment on helpnewyork.com: using Claude-in-Chrome to interrogate Google’s Analytics Advisor inside GA4, session by session, until I had a complete behavioral profile of every AI platform sending traffic to the site.

What came out of it is not what I expected. I expected traffic data. I got a content strategy.

The Setup

Claude-in-Chrome is Anthropic’s browser extension that lets Claude operate directly inside your browser — reading pages, clicking elements, filling inputs, capturing output. Analytics Advisor is Google’s Gemini-powered chat interface built into GA4, available to English-language accounts since December 2025. It answers natural language questions about your property data with charts, tables, and narrative interpretation.

The combination is unusual. You are using one AI (Claude) to systematically interrogate another AI (Gemini) about your site’s data, then synthesizing what comes back into strategy. The token budget for the heavy data reasoning stays inside Google’s infrastructure. Claude handles the query architecture, the capture protocol, and the synthesis.

I ran four structured sessions across two sittings, using a specific sequence of queries built to extract progressively deeper signal. Session 1 established baseline traffic. Session 2 closed gaps and confirmed AI referral data existed. Session 3 was the AI deep dive. Session 4 was velocity and geography.

What the Data Showed

Three AI platforms were sending meaningful traffic to helpnewyork.com during the 28-day window: ChatGPT, Claude, and Copilot. The behavioral profiles were so different from each other that treating them as a single “AI traffic” segment would have produced wrong conclusions.

Claude.ai traffic showed a 64% engagement rate and an average session duration of over 3 minutes. The dominant landing page was an NYC Summer Internships guide, accounting for over 60% of all Claude sessions. Geographic concentration was academic: Ithaca (Cornell), State College (Penn State), Washington DC. The users arriving from Claude were reading to act — they needed specific information, they found it, they stayed.

ChatGPT traffic showed a 21% engagement rate and an average session of 24 seconds. The top landing page was a cherry blossom guide. The users were fact-grabbing: they asked ChatGPT where to see cherry blossoms in New York, got a citation, clicked through, confirmed the location, and left. The content served its purpose in under half a minute.

Copilot traffic was between the two: 46% engagement, roughly 2-minute sessions, desktop-heavy, concentrated in New York’s suburbs. The top pages were civic services — SNAP benefits, tenant rights, transit discounts. These users were in planning mode, researching before they decided or applied.

The Finding That Reframes GEO

The cross-AI page overlap query was the most important one in the entire four-session arc. I asked Analytics Advisor which pages appeared in the top landing pages for more than one AI source. Only one real content page appeared in all three: the cherry blossom guide.

The obvious interpretation is that the cherry blossom guide was “AI-optimized.” The actual interpretation, once you look at the full traffic breakdown, is the opposite. Bing drove 59 sessions to that page. Yahoo drove 16 at 75% engagement and a 3-minute 46-second average session. DuckDuckGo drove 35. The combined AI traffic to that page was 32 sessions — 17% of total. The AI platforms were citing it because traditional search engines had already validated it as the highest-quality answer in the index.

AI citations are downstream of search quality, not upstream. The path to getting cited by ChatGPT, Claude, and Copilot is not to optimize for AI retrieval patterns. It is to build pages that win on Bing and Yahoo with enough depth that AI models treat them as authoritative sources. The GEO play is a traditional SEO play with better content.

The Content Strategy That Follows

Once you have the per-AI behavioral profiles, you have a content variant framework. The same article can be written in three structural architectures, each tuned to how one AI model retrieves and presents information.

The Claude variant is dense and process-oriented. Headers, eligibility criteria, numbered steps, official program names. Built for the student or researcher who arrived with a specific question and needs a complete answer they can act on.

The ChatGPT variant is a scannable list. Named items, one specific detail per item, direct answer in the first two sentences. Built for the user who will spend 24 seconds on the page and needs the answer immediately or they’re gone.

The Copilot variant is comparison and planning framing. What to know before you go, Option A versus Option B, cost context, logistics. Built for the desktop user doing research before they make a decision.

The core article is the same. The architecture is different. The AI that cites you depends on which structure you used.

The Methodology Is the Product

The query sequence I developed across these four sessions is a repeatable extraction methodology. It works on any GA4 property with Analytics Advisor enabled. The intelligence it produces — per-AI audience profiles, geographic signals, velocity trends, cross-AI content overlap — is not available through DataForSEO, SpyFu, or GSC. It requires Gemini’s reasoning layer operating on top of your property data, orchestrated by a structured query architecture.

I have packaged the complete methodology as a downloadable kit: the full query architecture across all four sessions, the capture protocol, the content variant framework, and the flags to escalate before your next content sprint. It is called Books for Bots: GA4 AI Referral Audit Kit.

The free version covers Session 3 alone — the AI deep dive queries that surface your ChatGPT, Claude, and Copilot traffic split. That alone will show you something most site owners have never seen: which AI is sending them traffic, to which pages, and how engaged those users actually are.

The full kit covers all four sessions and includes the content variant framework that translates the behavioral data into a writing system.

Both are available at tygartmedia.com. What you do with the data after that is yours.

Short version: In the last 29 days, Claude, ChatGPT, Perplexity, Microsoft Copilot, Gemini, NotebookLM, and Kagi collectively sent at least 94 new readers to tygartmedia.com — a site whose #1 content vertical is explaining Claude. AI assistants are now our #4 traffic source, ahead of Facebook, ahead of LinkedIn, ahead of every search engine except Google and Bing. The product is citing the publication that covers the product. That’s the loop. Here is what it looks like when you can actually measure it.

The finding that made me stop scrolling

I built a Claude-powered browser agent to poke around our GA4 account and surface “interesting stuff” a human analyst would miss. One of the first things it flagged was our Source/Medium report. Here is the top of the list, unedited:

Rank	Source / Medium	New Users (29 days)	Notes
1	(direct) / (none)	738	Mystery bucket
2	google / organic	289	Standard Google SEO
3	bing / organic	70	1m 20s average session — high intent
4	claude.ai / referral	63	Claude itself
5	m.facebook.com	43	Mostly 4-second bounces
6	duckduckgo / organic	41	1m 02s average
13	chatgpt.com / referral	9	ChatGPT
15	perplexity.ai / referral	5	Perplexity
21	copilot.com	3	Microsoft Copilot
24	gemini.google.com	2	Google Gemini
28	notebooklm.google.com	1	Google NotebookLM
35	kagi.com	1	Kagi AI results

Add up everything with an AI-assistant referrer and the combined count is at least 94 new users in 29 days — roughly 6.7% of all new users on the site. Claude alone, at 63 referred users, is our #4 traffic source. It is ahead of Facebook. It is ahead of LinkedIn. It is ahead of every search engine except Google and Bing. And we have been cited, at least once, by every major AI surface in the English-speaking internet: Claude, ChatGPT, Perplexity, Microsoft Copilot, Gemini, NotebookLM, and Kagi.

Why this is different from “we show up in Google”

Generative Engine Optimization (GEO) is the practice of structuring content so that large language models cite it as a source inside their answers. It is the younger, messier cousin of SEO. Most publishers cannot yet prove it is working. The feedback loop is long, the data is hidden inside a chat window, and the traffic that does leak through often lands in a “(direct)” bucket with no attribution at all.

We can see ours. GA4, for reasons that are probably accidental, already records claude.ai, chatgpt.com, perplexity.ai, copilot.com, gemini.google.com, notebooklm.google.com, and kagi.com as discrete referral sources when a user clicks a citation link. That means AI-assistant traffic is measurable as a first-class channel right now, today, with the free version of Google Analytics, on any site that happens to get cited.

The poetic layer of what we are looking at: Claude is the top AI referrer to a website whose #1 content vertical is explaining Claude. The product is sending readers to the publication that covers the product. If that is not a GEO moat, I do not know what one looks like.

These are not bounced visitors. They are readers.

The single biggest worry with any new traffic source is that it might be garbage — bots, previews, accidental clicks. The engagement data says the opposite. Users arriving from claude.ai spend 23 seconds on average and produce 0.56 engaged sessions per user. ChatGPT referrals average 21 seconds and 0.44 engaged sessions per user. For context, the site-wide average engagement time is dragged down hard by in-app social browsers; the Facebook mobile webview, for example, sits at about 14 seconds with 4-second bounces.

People arriving from an AI assistant are not scrolling past. They clicked the citation because the AI told them this was the primary source, and when they got here they read. That is a qualitatively different kind of traffic than Facebook or a random Google search. These are the highest-intent non-search users we have.

The secondary finding: Seattle is reading for three minutes

The same GA4 pass surfaced a city-level pattern we were not expecting. Seattle readers — 61 of them in 29 days — spent an average of 3 minutes and 6 seconds on site at a 61.3% engagement rate. The site-wide average session is roughly 40 seconds. Seattle readers are spending about 4–5x longer on the page than the typical visitor, at nearly twice the engagement rate.

City	Active Users	Engagement Rate	Average Time
Seattle	61	61.3%	3m 06s
The Dalles, OR	31	0%	1s
Shelton, WA	26	27.6%	15s
Des Moines	24	37.5%	10s
Beijing	31	6.5%	0s
Singapore	28	21.4%	5s

A few things jump out. The Dalles, Oregon at 31 users / 0% engagement / 1 second is almost certainly Google’s data center there returning preview requests — ignore it. Shelton, Washington is a real Mason County hyperlocal beachhead; 26 actual humans in our home county in 29 days is a legitimate foothold for the local desk. Beijing at 31 users / 0 seconds has the classic signature of cloud-hosted scrapers. And Seattle at 3 minutes is the single most valuable city in our data and it is not close.

The browser split confirms an unusually technical audience

Browser	Users	Engagement Rate
Chrome	850 (60%)	31.3%
Safari	232 (16%)	32.7%
Edge	99 (7%)	62.3%
Firefox	33 (2.3%)	60.5%

Edge at 62.3% engagement and Firefox at 60.5% engagement are not normal consumer numbers. A typical general-interest site sees those two browsers hovering in the 5–15% range. Microsoft Edge is the default on corporate-managed Windows machines. Firefox is the dev-preferred privacy browser. The combination of high Edge engagement, high Firefox engagement, and a Claude-heavy referral list all point at the same audience: developers and technical professionals at real companies, reading on managed workstations.

How to measure AI-assistant referrals in your own GA4

If you publish anything technical and want to see your own version of this number, the fastest path is a custom GA4 exploration with one segment. Open GA4 → Explore → Free Form. Add a segment with this condition:

Session source contains one of:
  claude.ai
  chatgpt.com
  perplexity.ai
  perplexity
  copilot.com
  gemini.google.com
  notebooklm.google.com
  kagi.com
  you.com
  phind.com

Break it down by landing page, engagement rate, and average engagement time. That is your AI-Referral dashboard. Watch it weekly. A non-trivial number of sites will discover they already have measurable AI traffic and never bothered to look.

Frequently asked questions

What is a GEO referral?

A GEO referral, or AI-assistant referral, is a visit to your site from a user who clicked a citation link inside an answer generated by a large language model such as Claude, ChatGPT, Perplexity, Microsoft Copilot, Gemini, NotebookLM, or Kagi. In Google Analytics 4 these visits appear as referral traffic from the assistant’s domain — for example claude.ai / referral or chatgpt.com / referral.

How many AI-referred users did tygartmedia.com receive in 29 days?

At least 94 new users across seven distinct AI assistants: 63 from Claude, 14 from ChatGPT (9 attributed + 5 unassigned), 10 from Perplexity (5 attributed + 5 unassigned), 3 from Microsoft Copilot, 2 from Gemini, 1 from NotebookLM, and 1 from Kagi. That is roughly 6.7% of all new users on the site for the period.

Are AI-assistant referrals real readers or bots?

Real readers. Average engagement time from claude.ai is 23 seconds and from chatgpt.com is 21 seconds, with engagement rates of 0.56 and 0.44 engaged sessions per user respectively. Those numbers are qualitatively higher than in-app social browser traffic (Facebook mobile webview averages about 14 seconds) and indicate a deliberate click-through from an AI citation, not a scraper.

Can any publisher measure AI-assistant referrals in GA4?

Yes. GA4 records visits from claude.ai, chatgpt.com, perplexity.ai, copilot.com, gemini.google.com, notebooklm.google.com, and kagi.com as discrete referral sources by default. Build a Free Form exploration with a segment that filters Session source on those domains and you will see the channel immediately if it exists for your site.

What is GEO in marketing?

GEO stands for Generative Engine Optimization. It is the practice of structuring web content, schema markup, and publishing signals so that large language models cite the content as a source inside AI-generated answers. GEO is to AI assistants what SEO is to search engines — the discipline of being the answer the machine hands to the reader.

The loop, and why it matters

The most interesting thing about this data is not the traffic. It is the feedback structure. Tygart Media publishes explainers about Claude. Claude crawls and cites those explainers. Readers click through from Claude’s answer back to tygartmedia.com. We publish more. Claude cites more. The site becomes, in effect, training data and a recommended source for the next iteration of the product it covers. That is the recursive loop that makes AI-native publishing a different business than search-era publishing.

I do not think every site can build this loop. It requires a narrow, technically-defensible topic — something an AI assistant would rather cite than paraphrase — and the patience to publish at a cadence LLMs reward. What I do think is that any publisher can check, today, whether the loop has quietly started forming underneath them. Most have not bothered. This post is partly a flex and partly an invitation: go look.

What happens next at Tygart Media

Three things. We are standing up a permanent AI-Referral channel in our GA4 so the number can be watched weekly instead of rediscovered quarterly. We are writing the playbook — the one this post hints at — for publishers who want to do the same. And we are building the browser agent that found this in the first place into a repeatable audit any publisher can run against their own GA4 in an afternoon. If that last one sounds useful, the newsletter is the place to follow along.

Claude sent us 63 readers last month. It will send more next month. We will be counting.

Part 2 of 2. In the first post I showed that Claude, ChatGPT, Perplexity, Copilot, Gemini, NotebookLM, and Kagi collectively sent tygartmedia.com at least 94 new readers in 29 days — and that Claude alone is our #4 traffic source. That is the headline. What follows is the interesting part: when you filter the landing-page report one AI model at a time, the three major assistants cite completely different kinds of pages, and the pattern is actionable.

Claude cites a small number of pages, a lot of times

Claude.ai sent 79 sessions across 63 users to 16 distinct pages. Two pages ate more than half of it:

#	Page	Sessions	% of Claude traffic	Avg Time
1	/claude-student-discount	22	27.9%	35s
2	/anthropic-console	21	26.6%	11s
3	(not set)	13	16.5%	5s
4	/claude-edu	4	5.1%	6s
5	/claude-pro-vs-chatgpt-plus	4	5.1%	7s
6	/claude-code-on-vertex-ai-gcp	3	3.8%	3s
7	/claude-desktop	2	2.5%	40s
8	/how-to-install-claude-code	2	2.5%	2s
9	/claude-4-deprecation	1	1.3%	1m 07s
10	/claude-managed-agents-pricing-cost-analysis	1	1.3%	1m 38s

The two biggest pages, /claude-student-discount and /anthropic-console, are 54.5% of all Claude-referred traffic to the site. Those are extremely specific query shapes — “how do students get Claude Pro free” and “how do I access the Anthropic Console” — and Claude has apparently decided our pages are the canonical answer for both.

The engagement twist is worth staring at. The two biggest Claude-referred pages have the worst time-on-page: 35 seconds and 11 seconds. The two pages that got a single Claude visit each — /claude-managed-agents-pricing-cost-analysis and /claude-4-deprecation — got 1 minute 38 seconds and 1 minute 7 seconds of real read time. The pattern is clean. When Claude can extract the answer directly into its chat window, users click through briefly to verify and leave. When the answer is deeper than Claude can summarize, readers stay to actually read. Both behaviors are valuable and both are measurable.

ChatGPT cites broadly, favors “X vs Y” content, and (oddly) sends geographic traffic

ChatGPT’s footprint is shaped differently. 16 sessions across 14 users to 13 distinct pages — almost every page received exactly one visit, which is the signature of a model citing a wide range of sources once each rather than reaching for a favorite.

Page	Sessions	Avg Time
/claude-student-discount	3	15s
/claude-computer-use-tutorial	1	2m 07s
/grok-vs-claude	1	15s
/opus-4-7-vs-gpt-5-4-vs-gemini-3-1-pro	1	0s
/claude-pro-vs-chatgpt-plus	(cross-model)	—
/claude-for-nonprofits	1	30s
/everett-waterfront-visitor-guide…	1	0s
/hood-canal-shellfish-season-2026…	1	0s
/rakuten-claude-managed-agents-enterprise-deployment	1	0s

Two patterns in that list. First, ChatGPT appears to cite us disproportionately for model comparisons — grok-vs-claude, opus-4-7-vs-gpt-5-4-vs-gemini-3-1-pro, and the cross-model claude-pro-vs-chatgpt-plus page. Second, and stranger, ChatGPT sent visits to two hyperlocal Pacific Northwest pages: an Everett waterfront guide and a Hood Canal shellfish season page. That is ChatGPT using our site as a reference source for geographic queries, which is not a pattern any other model shows.

The hidden gem: /claude-computer-use-tutorial received one ChatGPT referral and that referral stayed for 2 minutes 7 seconds. ChatGPT appears willing to cite long-form technical tutorials in a way Claude does not.

Perplexity treats us like a research database

Perplexity sent 12 sessions across 10 users to 9 pages — the most evenly distributed of the three and the only model that cites people, founders, and company-history content.

Page	Sessions	Avg Time
/anthropic-founders-2	2	17s
/claude-code-on-vertex-ai-gcp	2	54s
/claude-student-discount	2	0s
/claude-desktop	1	4s
/claude-team-plan	1	0s
/how-to-install-claude-code	1	0s
/restoration-team-training-claude-cowork	1	0s

Perplexity is the only model that pulled visits on /anthropic-founders-2, which implies Perplexity is fielding a different query shape — something closer to “who founded Anthropic” than “how do I use Claude.” Perplexity is also the only model that surfaced the very niche B2B page /restoration-team-training-claude-cowork. That is a long-tail, vertical-specific query and Perplexity cited us as the source. That is exactly the behavior you would hope for from a research-flavored assistant.

The three models have completely different citation personalities

Once you lay the three patterns side by side, the strategy falls out of the page.

• Claude.ai favors short, factual, access-related pages. Product info, pricing, how-to-access. If you want more Claude citations, write more narrow “how do I do this one specific thing” pages.
• ChatGPT favors comparisons and long-tail references. X vs Y, alternatives, and — unexpectedly — some geographic content. If you want more ChatGPT citations, write more “X vs Y” posts with tight comparison tables.
• Perplexity favors people, history, and niche research. Founders, company background, domain-specific tutorials. If you want more Perplexity citations, write more research-flavored background pieces.

This is the single most practical insight in the data set. Most people talk about “AI SEO” as if it is one thing. It is three things, at minimum, and the content shape that wins one model will not automatically win the other two.

The crown jewel: one page, 17% of all AI-referred traffic

The clearest cross-model winner on the site is /claude-student-discount. Claude sent 22 sessions. ChatGPT sent 3. Perplexity sent 2. Combined that is 27 sessions — roughly 17% of all AI-referred traffic we received in 29 days, from a single URL. No other page on the site is cited by all three major LLMs in meaningful volume.

There is a playbook inside that one data point. The page works because the query “how do I get Claude for free as a student” is an extremely high-frequency question across every chat surface, and the page happens to be structured the way LLMs like to cite: a short, direct answer near the top, specific eligibility rules in a scannable block, and no wall of context before the reader gets to the fact. That structural recipe — front-load the answer, make the facts liftable, keep the page narrow — is repeatable.

The bigger finding: 90% of our Claude content is invisible to AI

tygartmedia.com has more than 250 Claude-related articles. Exactly 25 of them show up in the AI-referral data set at all. The 90% that do not get cited are not low-quality — several of them have strong engagement from regular search traffic:

• /claude-managed-agents-complete-pricing-guide-2026 — 17 sessions at ~1 minute from search, zero AI citations
• /notion-knowledge-base-for-claude — 10 sessions at 1m 23s, uncited
• /claude-rate-limits — classic FAQ shape, 6 sessions, not cited
• /claude-md-playbook — 1 session at 2m 33s, zero AI pickup
• The full /claude-cowork-* family of 12+ pages, almost entirely invisible to every model

The difference between an AI-cited page and an AI-invisible page is rarely the quality of the content. It is the shape. Pages that get cited have an early summary, short headings, bulleted facts, and a quotable direct-answer sentence. Pages that do not get cited tend to open with context, build up to the answer, and bury the quotable line in paragraph 9.

The content-cluster scorecard

Cluster	Approx. Pages	Approx. Sessions	Engagement	AI Citations
Claude pricing & access	~10	~160	Mixed	High
Claude managed agents	~12	~130	Strong (25s–1m)	Low
Claude Code	~8	~60	High (18s–3m)	Moderate
Model comparisons (X vs Y)	~10	~45	Very high (1–7 min)	Moderate
Anthropic people/company	~8	~30	Medium	Moderate
Claude how-to / tutorials	~20	~50	Medium	Low
Claude Cowork family	~15	~40	Very low (0–10s)	Almost none

Two clusters deserve action. The Claude Cowork family is a content swamp — 15 pages, low traffic, no AI citations, and 0–10 second engagement on the traffic that does land. That cluster should be consolidated into two or three flagship posts and the rest redirected. The model comparisons cluster is the opposite: low volume but 1–7 minutes of engagement and cross-model citations. One well-researched comparison post outperforms ten mediocre explainers on every metric that matters here.

The playbook, in one list

• Write more narrow single-answer pages. Candidates I would ship next: /claude-web-search, /claude-api-keys, /claude-max-plan-vs-pro, /how-to-cancel-claude, /claude-mobile-app, /claude-desktop-vs-web, /claude-subscription-refund. Each is ~600 words, answer-first, scannable. That is the shape Claude cites.
• Add a Quick Answer block to the top of every long-form piece. Two or three sentences. Quotable. That alone moves a real share of our invisible content into AI-citation range.
• Invest in comparison posts for ChatGPT pickup. We already know ChatGPT cites our existing X-vs-Y content. Ship more of them, with tight tables.
• Write more founder/history/background pieces for Perplexity pickup. Research-flavored. Dates, names, primary sources.
• Consolidate the Cowork cluster. Two or three flagship pages, everything else redirected.
• Ship a permanent AI-Referral dashboard in GA4. Segment on all seven assistant domains. Watch it weekly. This is now a first-class channel.

Frequently asked questions

What kinds of pages does Claude.ai cite most often?

Based on the tygartmedia.com data, Claude.ai disproportionately cites short, factual, access-related pages — product info, pricing, how-to-access, and eligibility details. On our site, two pages (/claude-student-discount and /anthropic-console) accounted for 54.5% of all Claude-referred traffic in a 29-day window.

What kinds of pages does ChatGPT cite most often?

ChatGPT’s citation pattern favors comparison and long-tail reference pages — “X vs Y” posts like Grok vs Claude, model-to-model comparisons, and, surprisingly, some geographic and local content. ChatGPT tends to cite many pages once each rather than concentrating on a small set.

What kinds of pages does Perplexity cite most often?

Perplexity cites research-flavored content — founders and company history, domain-specific tutorials, and niche B2B pages. It is the only major AI assistant that sent traffic to our Anthropic founders page and to a vertical-specific training page in our data set.

Why does the same page get different citation volume from different AI models?

Because each assistant is answering a slightly different distribution of queries. Claude is most often used for “how do I use this product” questions and favors narrow how-to pages. ChatGPT receives more comparison and alternative-seeking queries. Perplexity skews toward research and background questions. A page that is the best answer for one query type will not automatically be the best answer for another.

How do I structure a page to get cited by AI assistants?

Lead with a direct, quotable answer in the first paragraph. Use short scannable headings. Keep facts in bulleted or tabular form. Include an explicit FAQ block with question-shaped subheadings. Keep the page narrow — one topic, one canonical answer — rather than a sprawling multi-topic explainer.

The bigger picture

The meta-insight worth sitting with: we are currently being cited inside Claude’s internal answer graph for “Claude student discount” because a human sat down and wrote a clear, narrow page about it. That is almost the entire game for publishers for the next three years. Most of the web has not noticed yet. We noticed, and now we have a measurement stack to act on what we noticed.

If you are a publisher, the thing to do this week is boring and powerful: segment your GA4 on the seven AI-assistant domains from Part 1, sort your landing pages by AI-referral volume, and look at the pages that are winning. They will have a shape. Copy it.

— If you missed it, Part 1 is here.

Direct test

Second attempt