top of page
Search

Teaching AI visibility in fall 2026: a field guide for marketing professors

  • Writer: Clark Boyd
    Clark Boyd
  • Jun 24
  • 7 min read

Two pieces of recent research say very different things about AI visibility, and both are correct. Similarweb's downstream-impact analysis shows that referral traffic from AI assistants, although still small in volume, is among the highest-quality traffic a website now receives. iPullRank's Personal Intelligence experiment, by contrast, shows that the measurement of AI visibility is leakier than the GEO industry has been claiming. Both findings have direct implications for what marketing professors should put on the autumn syllabus.


This piece sets them out, with the data behind each.

The downstream impact is real


Adoption is no longer the question. Pew Research's June 2026 survey of 5,119 US adults found that 49% have now used an AI chatbot, up from 33% in 2024 and 23% in 2023; 24% use one daily. ChatGPT adoption alone reached 44% of US adults, up from 18% in 2023, a 2.4x rise in three years. Among under-30s, 58% have used ChatGPT. Among under-50s, adoption runs at roughly twice the rate of the 50-plus group (57% vs. 28%). On the supply side, HubSpot's 2026 State of Marketing reports that 86.4% of marketing teams now use AI somewhere in their workflow, up from 41% in 2024 and 67% in 2025.


Volume context matters. SparkToro and Datos found that Google still received roughly 373 times more searches per day than ChatGPT in 2024 (around 14 billion against 37.5 million). Even at maximum overlap, the combined AI tools (ChatGPT, Perplexity, Claude, Copilot, Gemini) account for less than 2% of the search market. AI search is not, yet, replacing classic search by volume.


What it is doing is producing exceptionally well-qualified visitors. Similarweb's clickstream data shows ChatGPT referral traffic converting at 7.1%, second only to paid search (7.8%) and well above direct (3.6%), organic (2.4%), and social (0.9%). The same dataset shows that users arriving from ChatGPT spend around 15 minutes on site and view 12 pages per visit, against roughly 8 minutes and 9 pages for visitors arriving from Google. Semrush's analysis of more than one billion lines of US clickstream data reports LLM-referred visitors converting at 4.4 times the rate of organic search visitors, with some sites seeing 18% conversion on LLM-referred users.


The traffic is also growing quickly. Similarweb data shows ChatGPT web visits grew 84% between September 2024 and March 2026; Gemini visits grew roughly ninefold in the same window. After ChatGPT's 7 May 2026 change adding in-response brand links, referrals from ChatGPT rose 157.7% week-over-week, and homepage referrals rose 354.7%.

The conclusion is plain. Being cited by an AI assistant is not currently a high-volume game. It is a high-intent game. Each citation is rare and disproportionately valuable in a way classic SEO traffic mostly is not.


The measurement is broken


The harder finding landed on 21 May 2026, when Garrett Sussman, Michael Tandoh, and Cate Dombrowski at iPullRank published the results of their Personal Intelligence experiment. They set up three Google accounts. One blank control. One blank account opted into Google's Personal Intelligence mode and seeded with brand signals via Gmail and Google Photos. One mature personal account. They ran 1,922 AI Mode responses across 8 categories (coffee machines, streaming, banks, SEO agencies, hoodies, running shoes, smartphones, productivity tools) and 6 prompt types each, producing 22,064 brand-level rows between 30 March and 15 April 2026.


The seeded brands were 46 percentage points more likely to appear in the Personal Intelligence-connected account than in the control. Brand appearance rose from 23.9% in the control account to 66.8% in the PI-enabled one. Top-three placement rose by 23.1 points (from 4.5% to 24.9%). Top-ten placement rose by 42.8 points (from 17.7% to 54.6%). Gmail seeding outperformed Photo seeding by an order of magnitude: 53.6% appearance rate for Gmail-seeded brands against 10.5% for Photo-seeded brands; 11.35% citation rate for Gmail against 0.54% for Photos.


The most striking result. The team seeded entirely fabricated brands with no public web presence at all. Those fabricated brands still appeared in 35.7% of relevant Personal Intelligence responses. They earned a 0% citation rate, because there was nothing on the web to cite, but they appeared in the answer all the same. In other words: Google AI Mode will surface a brand to a logged-in user purely because that user has interacted with it by email, with no other supporting signal.


This breaks the central premise of the current generation of GEO measurement tools, which run prompts through anonymous servers and report which brands appear in the responses. If logged-in user state changes which brands appear by 46 points, then no anonymous measurement can produce a universal "share of model" number. There is no universal AI visibility. There is your AI visibility to a specific logged-in user, with their specific inbox, their specific photo library, their specific search history.


What Jim Lecinski calls this


Jim Lecinski, Professor of Marketing at Northwestern Kellogg and previously a senior leader at Google, has been making this point publicly since the iPullRank study landed. In his LinkedIn summary, he flagged two specific consequences. First:

GEO firms running simulated queries from anonymized servers can't fully measure what logged-in users actually see, making a universal "AI share of model" metric pretty much a myth.

Second:

Email marketing, loyalty programs, owning the Inbox, are now the newest frontier for LLM discovery, not just Reddit posts.

The implication for marketing teaching is the one to sit with. AI visibility is not a single number to be measured and optimised. It is a layered phenomenon that depends on the user's logged-in state, their inbox, their search history, and the public web all at once. The same brand will appear differently to different users on the same day.


Jim has spoken with us before for our Office Hours series, where he made a related point about problem framing being the new literacy for marketing students.


Three things to teach about AI visibility this autumn


The two pieces of evidence above are not impossible to reconcile. They produce three concrete teaching moves for the coming academic year. Each one is also a useful corrective to the simpler version of AI visibility that has dominated marketing trade press for the past eighteen months.


1. Teach AI visibility as per-user, not per-brand


The single biggest update is to retire the idea that there is one number describing how visible a brand is in AI search. There isn't. The Princeton GEO paper from 2024, still the most-cited piece of academic work on the topic, tested nine optimisation tactics across 10,000 queries on a Bing Chat-style system and found that statistics boosted AI citation visibility by around 41%, quotation addition by 28%, and citing sources by up to 115% for lower-ranked content. That research is useful. It describes one layer, the web-content layer. The iPullRank study describes another, the user-state layer, that the Princeton work does not address. Students should leave your module with both layers in their heads, and with a clear sense that no single number captures both.


A useful in-class exercise: have students run the same prompt while logged in to their own accounts and while incognito, and compare the answers. The differences will not be subtle.


2. Teach the per-engine citation diet


Different AI engines source from different places. Profound's analysis of 27 million prompts and responses found ChatGPT citing Wikipedia in roughly 47.9% of cases and Reddit in 11.3%; Google AI Overviews citing Reddit in 21.0% and YouTube in 18.8%; Perplexity citing Reddit in 46.7% and YouTube in 13.9%. A separate Ahrefs analysis of 75,000 brands found YouTube mentions in video titles and transcripts correlating with AI Overview visibility at a coefficient of 0.737, the highest correlation of any signal they measured, well above backlinks. A separate Ahrefs cut found brand mentions across the web correlating roughly three times more strongly with AI visibility than backlinks do.


A student who leaves your module with "you should publish on Reddit" has a quarter of the picture. A student who leaves with "the answer is different on each engine, and here is how to find out which sources each engine prefers" has all of it.


3. Teach the inbox and the CRM as visibility infrastructure


Lecinski's second point is the one that most directly changes what a marketing module should include. Email and loyalty programmes have been treated for the past decade as legacy channels with predictable returns, often relegated to the last week of the term. The iPullRank finding upgrades them. Owning the inbox is now an input to AI visibility, which is an input to discovery, which is an input to demand. The 35.7% appearance rate for entirely fabricated brands seeded only by email is the number to put on the slide.


That is a structural change in the value of a channel. The class on email marketing in 2026 is a different class from the one in 2022. The class on loyalty programmes is a different class. The class on CRM strategy is a different class. Most syllabi have not caught up.


The syllabus question


The simplest way to put what marketing professors should take from all of this. AI visibility is real, the impact is concentrated and high-intent, the measurement is honest only when bounded by the user's logged-in state, and the channels that produce signal are now more varied than the GEO industry's tooling suggests. A teaching module built on that frame produces graduates who can think about AI search as a layered, dynamic, personalised system. A teaching module built on "optimise for share of model" produces graduates who are wrong about something an interviewer will notice within five minutes.


If you would like a structured place for students to practise the layered version, our AI Search simulation is one. It lets students run the kind of queries the Princeton paper and the iPullRank team are running, see how outputs differ across engines and across users, and treat AI visibility as the layered problem it actually is. For more on what marketing students should be learning more broadly this year, our piece on the AI marketing skills students need to learn in 2026 is the companion read.


Sources

 
 
 

Comments

bottom of page