A company appears in 98.4% of AI-generated responses across five surfaces. It lands on the shortlist 93.2% of the time. By every GEO scorecard in circulation, that's a win. Champagne, screenshots, Slack celebrations.
The pipeline was shrinking anyway.
The Gap Between Being Named and Being Chosen
Bill Hunt, President at Bisan Digital, published the numbers in a November 2026 Search Engine Journal analysis that should make any demand gen leader pause. His team built 38 criteria-based prompts (not the generic "best vendor" type, but the kind actual procurement managers use) and ran them across ChatGPT, Gemini, Claude, Perplexity, and Google. The brand showed up almost everywhere. But 73.7% of those responses had weak or missing evidence supporting the claims AI made about the company. In 15.8% of cases, a critical evidence gap existed around an important decision criterion.
AI knew the company belonged in the conversation. It couldn't consistently explain why once a buyer started applying real eligibility criteria: integration specifics, workflow fit, regulatory compliance, measurable outcomes.
Why This Hits Pipeline, Not Just Vanity Metrics
Forrester's January 2026 Buyers' Journey Survey reported 94% of business buyers used AI in the buying process, up from 89% the prior year. A G2 survey found 51% of B2B software buyers started research with an AI chatbot more often than with Google, compared with 29% in its 2025 report. And 69% said an AI chatbot led them to pick a different vendor than initially planned.
Those numbers come from different surveys with different populations and different definitions of "used AI," so collapsing them into one headline metric would be sloppy. But the directional signal is consistent: buyers are forming shortlists inside AI surfaces before they ever visit your site. If your evidence doesn't travel with your brand name into those answers, you're getting mentioned and then filtered out in the same session.
Forrester's framing of "zero-click buying" makes this sharper. If buyers gather enough information to narrow a field without clicking through to vendor sites, your proof has to exist wherever AI systems can retrieve it. The click-dependent attribution model most teams run will miss this entirely.
The Measurement Problem You're Probably Not Solving
Most GEO dashboards track the wrong layer. They report mention rate, share of recommendations, citation frequency. Those are consideration-set metrics. They tell you whether you got on the list, not whether you survived the next three questions.
Hunt's analysis found that the company's existing visibility scorecard would have flagged zero problems. The evidence gaps only surfaced when prompts mimicked real buyer criteria, the kind of multi-part, constraint-heavy queries procurement teams actually type. His team had to build a custom evidence scorecard because no off-the-shelf tool measured what they needed: whether AI had enough substantiated information to keep the brand competitive as criteria got specific.
The operational implication: your measurement stack needs to separate AI visibility (mentions, recommendations, citations) from evidence quality (criteria coverage, claim substantiation, proof availability). Reporting the first without the second can overstate progress in a way that looks great in a deck and explains nothing about pipeline decline.
What "Evidence" Actually Means Here
Weak evidence didn't mean the company lacked capability. It meant the available information was fragmented, vague, or disconnected. A marketing page claimed "end-to-end" without showing the actual process. Efficiency claims had no supporting data. Integration details required connecting three separate pages before an AI model could infer how things worked. Each inference an AI system has to make is a point where a competitor with clearer documentation becomes the safer recommendation.
That's not a content-volume problem. Product teams know how integrations work. Sales knows which objections lose deals. Customer success knows which outcomes are real. The content team can't manufacture that knowledge; it can only organize what other teams surface. Hunt also noted that first-party evidence isn't inherently weak (authoritative product docs can be strong proof that a feature exists), but claims about superiority or measurable outcomes benefit from independent corroboration: customer validation, analyst coverage, third-party testing.
The Danger of "Evidence-Shaped Content"
Once the market internalizes that AI needs evidence, the predictable response is an explosion of content that looks like evidence but isn't: longer pages, superficial diagrams, weak surveys, synthetic case studies built to influence an LLM rather than inform a buyer. Hunt flagged this risk explicitly. If the content wouldn't help a knowledgeable customer make a better decision, manufacturing it for a GEO score is just the next version of keyword stuffing.
The hypothesis worth testing: if you audit your top 10 decision criteria (the ones sales hears in deal cycles), score each for evidence quality in AI responses, and close the gaps with real substantiation, pipeline velocity at the evaluation stage should improve. The trade-off: this is slow, cross-functional work that won't move your mention rate. It might not show up in any dashboard you currently run.
AI visibility metrics are leading indicators for consideration. No survey data available today establishes causality between GEO/AEO visibility and revenue outcomes. Treating mention rate as a revenue proxy is the same mistake as treating last-click as proof of incrementality. To get closer to truth, you need self-reported source-of-discovery in intake forms, win/loss analysis that asks about AI research, shortlist capture from sales conversations. None of these are clean. All of them are better than a Gartner survey showing 45% of B2B buyers used generative AI in a recent purchase extrapolated into a pipeline forecast.
A brand can win the AI visibility game and still lose deals because the evidence didn't survive the second question. The scorecard that matters isn't the one counting mentions. It's the one asking whether your proof travels as far as your name does.