Leadership1 publisher3 min readPublished
GEO vendor CEO says much of the industry is simulating its data
Andrew Higgins, who runs the AI search marketing firm Parsnipp, wrote that generative engine optimization is about two years old and that marketers cannot observe how people query the models. Ten experts, he said, give ten metrics.
The Board Room · Leadership desk
What happened
- Andrew Higgins, chief executive of the AI search marketing firm Parsnipp, wrote in a Forbes Tech Council column that his worry is not generative engine optimization itself but the certainty with which its playbook is being sold.
- He put the category at maybe two years old and said it changes every few weeks.
- Because measurement is unsettled, he wrote, the industry cannot credibly claim fully baked solutions for how brands should reposition their content, marketing strategy or investment.
- Asking ten GEO experts what to measure and how to measure it may produce ten different answers, according to Higgins.
- He framed GEO as the organic slice of marketing to AI, one piece of a job that will also cover advertising, commerce and brand as agents take over more of the customer journey.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- constraint Without an agreed definition of what a GEO metric counts, a buyer cannot line this quarter's vendor report up against last quarter's. That takes away the usual renewal test of whether the number moved.
- exposure The named voice cautioning on measurement sells inside the category, so a buyer weighing the warning is also weighing one competitor's characterisation of the others.
- precedent Hidden text and Like-gating paid until the platforms closed them, so a tactic that lifts AI citations now should be budgeted as something that depreciates.
- decision The live choice is contract shape: short tests a marketing team can stop, or a multi-year content rebuild priced against a number the seller both defines and produces.
Higgins runs Parsnipp, an AI search marketing company [1], so a column saying the playbook is unfinished is also a column against the people selling a finished one. That interest does not dispose of the claim, though it does tell a buyer which part to check first. The measurement part is cheap to check: ask each shortlisted vendor to write down what it counts and how it observes what it counts, then read the definitions side by side.
The distinction underneath the complaint is between a log and a model. Marketers grew up on site analytics, Google Search Console and keyword trackers [7], each of which records an event that happened. With AI search, Higgins wrote, "much of the industry is effectively simulating the data because we cannot see exactly how people are interacting with large language models" [6]. A simulated figure can be precise and stable, and reported every month, without having touched a single user session. If two vendors run the same simulation method, their agreement falls short of independent confirmation.
The column does not include spending figures [16]. So the proposition that marketing budgets are already being reallocated against unverifiable dashboards is not established here. What is on the record is one practitioner saying that consumer behaviour is still forming, that the labs keep shipping changes and that much of what marketers are doing remains untested [5]. His prescription is a sentence long: "Learn, test, measure what you can and stay nimble." [18]
Two timescales are running together in the piece, and they call for different decisions. The long one is about the surface: people are asking an agent instead of going to a retailer, a brand website or a social feed, and a middle layer now does the researching and the comparing, and increasingly the deciding [15]. Higgins dates the era of getting customers to somewhere you control and converting them there at 30 years [14]. The category being sold as a discipline is roughly a fifteenth of that age [17]. Waiting costs something as well, because the text an agent reads is being written this quarter. Spending now is defensible, and a multi-year content rebuild priced against a metric the seller defines is a different commitment altogether.
For the tactics themselves, Higgins offers a test in two questions: "Does the tactic help an AI system better understand what you sell, who you serve and why you are relevant? Or are you simply exploiting the fact that the system is not very good yet?" [11] He has the precedents to hand. Marketers hid keyword text against matching backgrounds, stuffed keywords into pages, built early Facebook apps that spammed users' friends and gated content behind a Like; those tactics worked, some extremely well, and then stopped working, because they exploited shortcomings in the system instead of helping the platform do what it wanted for users [10]. "It is perfectly fine to use a tactic that works today," Higgins wrote. "What is dangerous is mistaking that tactic for a permanent marketing strategy." [12]
What to watch
- Whether any model provider publishes observable query or referral data, which would let a GEO metric be checked against a log instead of a simulation.
- Whether GEO vendors converge on a shared definition of a citation or share-of-answer metric, or keep shipping proprietary ones.
- Whether practitioners selling a settled GEO playbook revise it in public as the labs keep shipping changes.