top of page

How to split a search budget across SEO, SEM, GEO and GEM

Writer: Clark Boyd
Clark Boyd
Nov 6
8 min read

Marketing professors ask me one question about search more than any other: how should a student split a search budget once AI assistants are one of the places their buyers go? The organic-and-paid version of that exercise has served courses well for years, and the AI column is new to everyone, so it is a fair thing to be curious about. The honest answer is that it depends on how much the brand already ranks and gets mentioned, which is true and not much use on its own, so here is the longer version. Pew Research Center published the clearest figure I have seen on what AI answers do to search behaviour in July 2025. It tracked roughly 69,000 Google searches made by 900 American adults in March of that year and found that when the page carried no AI Overview, users clicked a traditional result 15% of the time. When the page did carry one, they clicked a traditional result 8% of the time, and they clicked a link inside the overview itself about 1% of the time.


Pew Research 2025: Google users click a result 15% of the time without an AI Overview and 8% with one.
Pew tracked 900 US adults through 68,879 Google searches in March 2025. An AI Overview appeared on 18% of them.

That is what makes the budget question new. The same organic position buys fewer visits than it did, because roughly half as many people click through when an answer sits above the results, and the brands named in that answer are not always the ones that rank first in Google. A student who wants their brand in the answer has to spend on different things to get there, and I want to work through what that looks like in practice, with one example and the choices it forces.


What are the four cells?


The framework comes from the textbook I wrote with Dr. Michael Pettiette, Integrated Search Marketing, with a foreword by Jim Lecinski. It is the companion to our AI Search simulation, a classroom simulation we make at Novela for business school marketing courses and launched this year, and it covers the same ground the simulation puts students through: the two kinds of search a buyer now does, and the earned and paid ways of showing up in each.The book splits search spend into four cells, and the budget example in the next section uses all four, so here they are.

SEO is earning a position in Google's organic results. SEM is paying for a slot on the same results page. GEO, generative engine optimisation, is earning a mention or a citation when an AI assistant answers a question. GEM, generative engine marketing, is paying for an ad shown alongside that answer.


The four cells of a search budget: SEO and GEO are earned, SEM and GEM are bought; two reach buyers on Google, two reach buyers asking an AI assistant.
The earned cells keep working after the budget stops. The bought cells stop with it.

Buyers now search in both places, Google and the AI assistants, so a search budget that only funds the Google cells is leaving the AI cells to competitors.


Where SEO, SEM, GEO and GEM each appear: a Google results page and an AI assistant answer, with the four cells labelled.
The query, brand and answer are invented for illustration. The placements are how Google and ChatGPT show results and ads as of late 2026.

How the budget problem is set up


Here is the situation students face when they play the AI Search simulation in a marketing course. Beckwith & Co. is a fictional eyewear brand, created for the simulation, that is launching a direct online channel. In the first quarter the marketer has $100,000 and can only spend it on organic work. That means choosing which topics to compete on, where a topic is a cluster of searches the brand's buyers run, and producing content an AI assistant might cite. There are no ads yet. In the second quarter the budget doubles to $200,000 and two paid channels open, Google Ads and ChatGPT Ads. Nothing rolls over, and the half-year is judged on orders and on share of voice in AI answers, meaning how often the brand is named when an assistant answers the questions its buyers ask, measured by putting those questions to the assistants and counting the brand's appearances.


The simulation runs this way, an organic quarter first and then a bigger budget with paid channels added, because the paid cells work best on positions the organic cells have already earned. Students have to look at the organic topics they chose in the first quarter and decide which ones deserve paid support, which is a harder and more useful decision than spreading a fresh budget across everything.


The habit we ask students to build is a simple one: decide the organic spend first, and add paid only to topics that already rank or get mentioned, so the ads have something to build on. More spend is not always more orders, and students find that out quickly.


What are the trade-offs a student has to weigh?


The obvious move in the second quarter is to put the new paid budget on new topics, and I understand the pull. Paid is the fast option, the ads can be live within days, and spending on a new topic feels like expansion, while spending more on a topic you already have feels like standing still.


The trouble is what each kind of spend buys. Organic is earned: you build authority over time, there is no charge when the buyer clicks, and the position keeps bringing clicks after the quarter ends. Paid is rented: you bid for an ad slot on the same results page, you pay for each click (or, on the newer AI platforms, for each time the ad is shown), and the ads stop the moment the budget runs out. So paid spend lasts exactly as long as the budget, and organic spend is what remains when it stops. That is why we teach paid as something to add to topics where the brand already has some organic position or some AI mentions, because then the ads reinforce a page that is already earning. On a topic with neither, every visit has to be bought and nothing is left behind when the ads are switched off.


The second trade-off did not exist five years ago. Similarweb's clickstream data shows visitors arriving from ChatGPT converting at 7.1%, against 2.4% for visitors from organic search, and spending around 15 minutes and 12 pages on a site against 8 minutes and 9 pages from Google. Semrush's analysis of more than a billion lines of US clickstream data puts the conversion rate of visitors referred by an AI assistant at 4.4 times that of organic search visitors. The volume is still small next to Google, but the buyers are further along. I covered both studies in the fall field guide. A student now has to decide how much of the budget goes to reaching those buyers inside the assistants and how much stays on Google. On Google, the ad auction still ranks on relevance as well as bid, so copy written for the search and a landing page that matches it remain the cheapest things to improve.


Then there is the question the Pew number is really about: clicks or mentions. SEO and SEM buy visits. GEO buys something different: the brand gets named in the AI's answer, and the buyer may read that answer and never visit any website at all.


That means GEO cannot be judged by website visits. It has to be judged by how often the brand is named in AI answers. A student who judges it by visits will see a low number and conclude it is not working, when it may be doing exactly what it is meant to do. GEM is the newest cell and the least settled. Ads alongside AI answers exist, ChatGPT has carried them since early 2026 and Google shows them with AI Overviews, but the benchmarks are thin and the formats keep changing. For now, the sensible approach is a small test: pick one topic the brand already ranks for, put a modest amount on AI ads for it, and see the results before spending more.


So what does a defensible split look like?


The right split depends on how much the brand already ranks and gets mentioned, so there is no percentage I can give that would hold for every brand. What I can describe is what a good defence of any split looks like, because it is the same test I would apply to a marketer at work asking for sign-off on a search budget.


Before they commit, students write down four things:

  1. Which topics they are extending and which they are dropping.

  2. Whether they still have an informational topic funded (content that answers a buyer's question rather than selling to them).

  3. Where paid ads will add orders a topic would not get on its own, and where they would only pay for clicks the brand was already getting.

  4. What this brand can say about itself that its competitors cannot.


A weak answer to the first is a list of topics with no reason beside any of them. A weak answer to the third is ads on everything that sold last quarter. If a student can answer all four in plain words, and can say what they chose not to fund and why, the split is defensible whatever the percentages turn out to be.


That is the exercise students run in the AI Search simulation, with a market that responds to their decisions and a debrief afterwards in which the split gets argued over. We build each of our simulations with practitioners at Jellyfish (part of the BrandTech Group), so students face the market as it works in reality.


Two things to teach this spring


Neither of these needs a new module. Give students a fixed budget, the four cells and a brand they know, ask for a split and the four-question defence as homework, and spend the class session comparing the splits and asking each group to justify theirs. 


  1. Teach the budget across all four cells. If your search module still teaches organic and paid as the two choices, add the two AI cells and give students one allocation problem where the budget is shared across all four. The organic-first, paid-second sequence described above works in a classroom as well as it does in a simulation.

  2. Teach how the AI cell is measured. If you already teach GEO, tell students how that cell is judged, so that they report mentions and citations alongside clicks and do not treat it as failing because it brought few visits. A student who measures GEO by traffic will reach the wrong conclusion every time.


To try the Beckwith & Co. version with your students, instructors can request free access to the AI Search simulation on the product page, or book a demo and I will walk you through the two quarters. The four-cell framework, with the reading to go with it, is in Integrated Search Marketing.


There may be several defensible splits for Beckwith & Co. in the AI Search simulation, and I have read good plans that split the budget differently. The only plan I would fail is one that puts nothing at all into the two AI cells, because that is a plan for the search market of five years ago.


Three questions I get asked


How much should I spend on GEO? Enough to produce content an assistant would want to cite, and then measure the cell on mentions and citations rather than on the clicks it was never going to bring.


Is GEM worth it yet? It is worth a small test on a topic the brand already ranks for, and it is too early to give it a budget the size of a search ads budget.


Does SEO still matter? Yes, because organic position is what both paid cells are built on and it is still where most clicks go. The longer answer is in Is SEO still worth teaching?


Sources

1 Comment


kimeta fezol
5 days ago

Per mesi ho scelto i tavoli di blackjack guardando solo i limiti di puntata. Poi ho notato che due tavoli identici pagavano il blackjack in modo diverso: uno 3:2, l’altro 6:5. Tra i tavoli di Spin Sahara Italy la regola è indicata nella descrizione, senza doverla cercare aprendo il gioco. Sembra un dettaglio, ma quella differenza moltiplica per tre o quattro il vantaggio del banco rispetto a un tavolo 3:2. Da allora controllo quel numero prima dei limiti, perché è il dato che costa davvero.

Like
bottom of page