Why GEO scores aren’t reliable for B2B yet

Inside the GEO gold rush in 2026

Many of the marketing leaders we support describe a gold rush around generative engine optimization (GEO). Earlier this year we wrote about findability and Share of Answers, the practice of tracking how often a brand appears when buyers ask AI for a shortlist. That post recommended scoring a fixed set of prompts each month. Since then, the question we hear most has moved from whether to measure AI visibility to which platform to trust. A growing set of tools runs prompts against the major AI assistants on a schedule. They report back a visibility score, a ranking, or a share of voice. With that much money now going into tracking, it is worth checking how reliable the score is before it informs budget decisions.

What the research found

In a study published in January, Fishkin worked with Patrick O’Donnell of Gumshoe, a company that sells AI visibility tracking. They asked 600 volunteers to run 12 recommendation prompts through ChatGPT, Claude, and Google’s AI. Across 2,961 runs, ChatGPT and Google’s AI had less than a 1-in-100 chance of returning the same list of brands twice for the same question. Claude was only slightly more consistent. The chance of the same list in the same order was closer to 1 in 1,000.¹ The answers were badly fragmented. Consistency was higher in categories led by a few big names. When volunteers wrote their own prompts about choosing headphones, the 142 prompts that came back barely resembled each other. Bose, Sony, Sennheiser, and Apple still appeared in 55 to 77% of the answers. Fishkin credits the AI with reading the buyer’s intent behind those prompts. In a crowded category with many options, SparkToro saw far less consistency.¹ Many B2B categories are crowded.

In May, four researchers at Unusual, a firm that sells research and tools on how AI assistants recommend brands, published a paper that isolates the effect of wording. They ran roughly 12,000 tests on OpenAI and Anthropic models with web search turned on. Asking an identical question twice produced recommendation sets that shared only 50 to 61% of the brands named across both lists. Two natural rewordings of the same question, such as “best CRM” and “top CRM,” shared only about 29%. Adding a detail that narrows the buyer, such as “for a SaaS startup” or a specific country, cut that to about 14%. Turning up the models’ reasoning effort did not close the gap. The authors concluded that the wording of the prompt, more than the buyer’s underlying intent, decides which brands appear.²

Here is where the two studies converge. A visibility score is only as reliable as the prompts behind it. A brand’s position in any single answer means almost nothing. A small, fixed set of prompts cannot give a stable read. Fishkin believes enough prompts, run enough times, can make visibility a reasonable metric. The Unusual authors doubt any practically sized prompt set can.

Why B2B is likely to feel this more

A B2B purchase runs through a buying committee. Each member brings a different question to an AI assistant. An IT leader evaluating a security platform asks about integration and compliance. The finance lead on the same committee asks about cost and payback. Those questions shift again by industry and company size. They also shift with how far along the buyer is, from someone exploring a category for the first time to someone comparing finalists before a contract. If a few changed words can cut the shared brands between two answers in half, every one of those variables can change which brands come back. SparkToro saw lower consistency in crowded categories. When volunteers asked for brand design agencies for a friend’s coffee shop, the most visible names appeared in only 30 to 40% of answers.¹ Fishkin links that kind of inconsistency to categories with many options to choose from. Many B2B categories fit that description.

Where many platforms fall short

Two things have to be right for an AI visibility score to mean anything. The prompts have to sound like real buyers. There also have to be enough of them. In the platforms we have evaluated, brands are typically tracked on 20 to 50 prompts. Those prompts tend to read like consumer searches. That is not how a business decision maker or an IT leader talks when evaluating enterprise software. A score built on a small set of generic prompts tests questions the brand’s buyers do not ask. And because the same prompt returns a different list on almost every run, the score can move from week to week without anything changing in the market.

In our experience, these platforms also tend to stop at the diagnosis. They report where a brand is missing from AI answers without showing which buyer questions or content gaps caused it. That leaves the marketing team with no confident next step. Some platforms now add content tools or AI agents to close the gap. However, those tools are add-ons to a measurement product. The execution shows it. The content they produce tends to read as generic AI writing, heavy on the stock phrases B2B buyers have learned to skim past.

Teams haven’t agreed on the unit of measure

In the large marketing organizations we work with, several teams often run their own GEO analysis, each with different tools and methods. When the reports disagree, no one has a number they can defend at budget time. In our view, the organization first has to agree on who its buyers are, what the baseline is, which measures count, and what it is optimizing against. The measures matter because a brand can be in very different states in an AI answer. It can be mentioned or recommended. It can also be named and ruled out. A team has to decide which of those it is counting. The fix is one agreed set of measures tied to a shared map of the buying process. In our experience, teams should track those measures quarterly and read them alongside pipeline.

How we approach it at Intercept

We built our own AI visibility tracking around the two requirements above. It starts with learning how buyers ask. Our qualitative and quantitative studies put buyers through realistic buying scenarios. Those studies show us how they would prompt an AI assistant and how the conversation unfolds. Review sites show how people describe the solutions they evaluate. Watchtower, our AI-powered audience research platform, reads what roughly 20 million professionals say to each other in public on X, Reddit, Quora, Bluesky, and other platforms. That research is where our prompts come from.

Then we test at scale. An average run is about 1,500 prompts per AI engine. We spread them across different personas, industries, and levels of buying intent. The data shows which brands an IT buyer and a business buyer each see recommended. It also shows how that changes between a buyer who is exploring and one who is ready to choose. We use the same research to decide which content a brand should publish so AI assistants can find and cite it.

Even at that scale, no prompt set captures every way a buyer might ask. We treat our results as a directional read. They show us which buyer questions and content gaps to work on first.

Four signals to watch

1. Rankings lose credibility

SparkToro found that the order of brands in AI answers changed so much from run to run that ranking position was close to useless.¹ Watch for marketing teams to shift from asking where a brand ranks to asking how often it is mentioned and recommended across a large and realistic set of questions.

2. Prompt sources become a buying criterion

Watch for marketing teams to ask any platform how many prompts it runs and how it builds them. A platform grounded in real buyer research can explain its method without handing over its prompt list.

3. B2B measurement moves to the segment level

A single category score blends together buyers who ask very different questions. Watch for measurement that breaks visibility out by role and buying stage.

4. GEO spend gets judged on pipeline

Today, many teams judge GEO work by how often their brand gets mentioned. The Unusual researchers point to business results such as qualified pipeline and revenue as better measures to explore.² Watch for GEO budgets to be held to the same pipeline standard as the rest of the marketing mix.

Sources

1. Rand Fishkin, SparkToro, “New Research: AIs Are Highly Inconsistent When Recommending Brands or Products; Marketers Should Take Care When Tracking AI Visibility,” January 27, 2026. sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/

2. Will Jack, Noah Lehman, Keller Maloney, and Sarah Xu, Unusual, “Paraphrase Brittleness in Production Retrieval-Augmented Commercial Recommendation: Reproducibility Below the Rerun-Stability Baseline,” arXiv:2605.27440, May 2026. arxiv.org/abs/2605.27440