You are sitting through the deck.
Slide 4 is the one that lands. Someone typed a question into ChatGPT, the kind of question your buyers ask, and your company was not in the answer. Three competitors were. There is a screenshot, and the screenshot is doing most of the persuading.
Slide 9 has the budget line. It might be called AEO, or GEO, or AI search visibility. The number next to it is not enormous, which is part of how it gets approved.
Then someone asks what it returns, and the room goes slightly quiet, and the answer that comes back is about what happens if you do nothing.
That is insurance being sold against a fear, and the fear is doing the closing.
So let me do the thing nobody in that room did, which is check the numbers the field is quoting at you.
The Forecast and the Survey Were Published by the Same Company
The urgency mostly traces back to one statistic. Gartner has forecast that within three years, 90% of B2B purchases will be handled by AI agents, moving more than $15 trillion through automated exchanges. It gets quoted everywhere, usually in a tone suggesting it has already happened.
It is a forecast. It is a bet an analyst placed on a stage in October.
Here is what the same firm found when it stopped predicting and asked actual buyers. In a survey of 645 B2B buyers, 45% said they had used generative AI at all. And 69% said they prefer to validate whatever the AI tells them with a human sales rep (Gartner, 2026).
Read those two sentences next to each other again. One firm forecasts near-total machine intermediation by 2028. The same firm’s own fieldwork finds that fewer than half of buyers have used the tools, and that most of the ones who do go and ask a person whether the machine was telling the truth.
The survey was in the field one month before the prediction was announced. This is not old data being unfairly compared to fresh thinking. They were contemporaries.
My first career was in trade journalism, where the job was to check a claim before printing it. The habit never left.
When a company’s forecast and a company’s evidence point in opposite directions, the forecast is the product.
What Your Buyers Are Actually Doing With the Machine
The entire AEO pitch rests on a premise about discovery. The story is that buyers now start with the machine, that the machine builds their shortlist, and that if you are absent from the shortlist you were eliminated before you knew a deal existed.
The best available evidence says the tool is doing a different job.
6sense surveyed around 4,000 B2B buyers and found that 94% use LLMs somewhere in the buying journey. That is the number the AEO decks quote, and often they attribute it to the wrong research firm and describe it as shortlist-building.
What 6sense actually found is that LLM use peaks in the middle of the journey, not at the start. Buyers use it to build comparison tables, to draft RFPs, to model costs, and to plan implementation. And 6sense’s own explanation for the mid-journey peak is the sentence the AEO industry never quotes: buyers are overwhelmingly likely to have prior experience with the brands they are about to evaluate.
They are asking the machine to help them think about companies they already knew about.
That is a comparison tool. It sits where the buyer is already holding a set of names, and the question that decides whether you are in that set was answered long before anyone opened a chat window.
None of which should surprise you if you have been reading along. In Your Buyer Researched With ChatGPT. Then They Walked the Trade Show Floor, I argued that the machine handles the research and humans still close the trust gap. This is the same buyer, one step earlier, doing the same thing. They use the tool where it is useful and they decide elsewhere.
I should be straight about the evidence quality here. 6sense sells software that benefits from B2B buying looking sophisticated and AI-saturated, and they do not publish their survey dates. Their 94% and Gartner’s 45% cannot both be measuring the same thing, and neither firm publishes enough for us to know exactly where the definitions diverge. I am not going to pretend that gap away. But notice that the vendor-friendly number and the conservative number tell the same story about what the tool is for.
The Tactics Do Not Work As Sold, and Nobody Went Back to the Table
Now to the part of the deck with the tactics on it. Add statistics. Add quotations. Cite your sources.
The numbers attached to those tactics come from a real, peer-reviewed paper. It is called „GEO: Generative Engine Optimization,” it came out of Princeton, and it was published at KDD in 2024. It is good work, and the agencies are right to point at it.
They are reading it carelessly.
The paper’s central table is captioned, in the authors’ own words, „Absolute impression metrics.” That caption then says the best methods „improve upon baseline by 41% and 28% on Position-Adjusted Word Count and Subjective Impression respectively.”
Read that again slowly, because the field did not. It is one set of methods, scored on two different yardsticks. Forty-one percent on the first yardstick. Twenty-eight percent on the second.
The field read that sentence and split it into two separate tactics.
Go looking for what the paper found and more than one AI-search consultancy will tell you that statistics deliver a 41% lift and quotations deliver 28%. SeenRank calls statistics „the largest lever” at „+41% citation rate,” with quotes „second largest” at „+28%.” Omnibound lists „Statistics addition (+41%)” and „Quotation addition (+28%)” as two separate tactics. Others report the same study as 42.6% and 32.8%, or 37% and 30%, or flip it around and hand quotations the 41%.
Not one of those is a tactic score. Both numbers describe the same method, Quotation Addition, measured on two different yardsticks. And the split gets the ranking backwards, because quotations outscored statistics in every column of that table. The paper says so in plain text: used alone, statistics run „8% lower than Quotation Addition.”
In fairness to the researchers, the caption invites the error. „The best methods,” plural, is sloppy phrasing in a paper that is otherwise careful. And statistics are genuinely the strongest tactic in a few specific domains, law and government among them, which the paper reports and almost nobody quotes.
But the numbers being sold to you do not appear in that paper. Not in any column, under any reading.
Nobody went back to the table.
I am not making a large point about two numbers. I am making a point about a field that will resell you a finding it has not read, with a confidence that scales inversely with how closely anyone has looked. The tell is not that the numbers are wrong. The tell is that they disagree with each other.
And when you do look, the paper is more interesting than the sales pitch. Keyword stuffing, for instance, scored below doing nothing at all on the paper’s headline metric. Some things are constant across every era of this industry.
Then there is the question of whether these techniques work in a live setting at all. A 2025 benchmark study called C-SEO Bench tested them directly and concluded that most current methods are „not only largely ineffective but also frequently have a negative impact on document ranking.” Plain SEO, the boring kind you already do, outperformed the specialized AI-optimization techniques.
What Happens When Everyone Buys the Same Ticket
Set aside whether the tactics work. Assume for a moment that they do. There is a harder problem underneath, and it is sitting in section 5.2 of the Princeton paper, which I suspect very few people quoting the paper have read.
The researchers tested what happens when multiple sites optimize at once. The site ranked fifth gained 115% in visibility. The top-ranked site lost 30.3%.
Visibility moved. It did not appear. There is a fixed amount of room in a generated answer, and optimization shuffles who occupies it.
A June 2026 study measured the decay directly. Its authors ran the adoption curve and found that the payoff to the first mover collapses to roughly nothing once everyone adopts, with a half-life of about 1.4 brands. Your advantage is cut in half by the time one and a half competitors copy you.
This is the answer to the question the room went quiet on. What does the AEO line buy you if your four competitors buy it too?
It buys you the right to stay where you are. Everyone spends to hold position. Nobody moves.
The only party whose position reliably improved is the one selling the tickets.
An AEO sprint is a random act of marketing with a new acronym on it. It does not compound. Next year the tactics will have changed, the models will have been retrained, and you will be asked for the budget again.
The Case Against What I Just Told You
Here is where I have to be careful, because the strongest argument against this piece is one search away, and you deserve to hear it from me rather than find it later.
The study I just cited on payoff decay also found something more dramatic: in a controlled test, the language models recommended the brand they already knew in 100% of 670 trials when the products were otherwise identical. Zero recommendations for the unknown challenger. That finding is the reason this article exists.
It is also a study about skincare. Moisturizer, sunscreen, and cleanser, with invented competitor brands. It has robustness checks on USB cables and AA batteries. It contains no B2B software, no professional services, and no considered purchase of any kind. It is a preprint from June, four weeks old, and it has not been peer-reviewed. Every sentence anyone writes transferring that finding to how a CIO picks a software development partner is an argument, not a result. Including mine.
And there is direct counter-evidence. Another June 2026 study tested real brands across five real industries and found recommendation concentration to be moderate rather than winner-take-all. The authors’ own words are that their results „sit in tension with a strong winner-takes-all narrative.” Niche challengers displaced big names. The three models they tested agreed on the top brand only 41.6% of the time.
And the least concentrated category they measured, the one where the field was most open and challengers did best, was SaaS.
That study is also a preprint, and its author is affiliated with a commercial AI-visibility tool, which is a conflict worth knowing about. I am not going to use that to wave the finding away, because the finding is inconvenient for me and I would rather tell you about it.
So the honest position is this. In low-information consumer categories, the machine appears to hand the win to whoever it already knew. In B2B software, the evidence points to a field that is still genuinely open.
That is good news. It is just not good news for the thing on slide 9.
What Actually Survives
Strip out everything that saturates, everything that gets competed away, everything the next model update erases. Ask what is left. What is the input to an AI recommendation that a competitor cannot neutralize by buying the same service you bought?
There is only one. Whether the model, and the buyer holding the model, already knew who you were.
That is the variable underneath all of this. It is what the incumbent advantage measures. It is why buyers reach for the machine mid-journey with a set of names already in hand. And it is the one thing an optimization sprint cannot manufacture.
Being known is an accumulation. It builds the slow way or it does not build.
You get it from being visibly, consistently the same company saying the same defensible thing for years, in enough places, until the market has a slot for you and the training data has your name in it. The unglamorous work. The work that has no launch date and cannot be finished in a quarter.
Which is exactly why it keeps losing to the thing on slide 9. Slide 9 has a number, a deadline, and a vendor. Being known has none of those, so it never wins the argument in the room, and it gets deferred again, and a year later you are looking at another screenshot of an answer you are not in.
The companies that will be in that answer in three years are the ones spending those three years being consistently worth knowing, which requires the marketing function to hold a position long enough for the market to learn it. That takes structure, and structure is the thing that survives when the people and the channels change. It is also the least purchasable thing in this entire conversation.
The Question to Put to Whoever Is Selling You Slide 9
Ask them one thing.
What does this buy us if all four of our competitors buy it too?
If the answer is honest, it will be some version of „then we all stay where we are.” That is a tax, and there may be quarters where paying a tax is the correct call. Pay it with your eyes open, as the cost of not being disadvantaged, and never as a growth plan.
Then ask the harder question, the one that has no vendor attached and no line on slide 9.
In three years, when a buyer asks a machine who they should be talking to, what reason will it have to know your name?
