In most B2B deals, the winner is on the shortlist before sales ever hears about it. 6sense's buyer research, covering about 4,000 buyers, found the winning vendor was on the buyer's Day One shortlist 95% of the time, and about four in five deals went to the favorite the buyer already had in mind. The survey is vendor-run, and the pattern is consistent with how buying works.
The open question is where that shortlist forms. Increasingly, part of the answer is an AI assistant. Forrester's 2026 research says generative AI search is now a starting point for vendor research, and that buyers verify what it says with peers and trials. Gartner's 2026 survey of 645 buyers reports that 69% prefer to validate AI-generated insights with sales reps. AI opens the research and people close the decision.
That puts a premium on two things: being in the consideration set before first contact, and being described accurately when a buyer checks a vendor with AI.
I have spent my career building the software that decides what gets shown to whom. The judge has gained a new member. Alongside analyst reports and peer recommendations, a model now reads your website, your customers' language, your structured data and the rest of the internet, and it does not care how good your sales deck is. The model responds to evidence, and evidence can be built.
What the data does and does not show
Precision matters here, because this topic attracts inflated claims.
- AI is mid-journey more than first touch. 6sense found large language model use peaks mid-journey for comparing and synthesizing, more than for discovering vendors.
- The traffic is small and the influence is larger. Octane11, analyzing 400 million B2B sessions (vendor-reported), found AI search under 2% of B2B search referrals, and about 9% of closed-won deals had an AI search touch.
- No study isolates AI as the cause of a shortlist spot. The shortlist forms early with or without AI. Treat AI as a growing influence, not a proven gatekeeper.
The answer depends on who is asking, and on the run
Most companies check how they appear in AI with one prompt from one account, and read the result as the truth. Two things distort that reading.
Who asks matters. Profound analyzed 71,147 AI responses (vendor-run, three consumer categories) and found different personas asking the same question shared about 25% of brands, against about 40% when the same persona asked again. I have not found a published test of B2B roles, such as a CFO against a COO. That gap is why we run buyer-persona panels: the same questions asked from different buyer profiles, compared against a control with no profile. The difference between the control and each persona is the signal.
The run matters. SparkToro's testing of nearly 3,000 prompts found the same brand list in fewer than 1 in 100 repeat runs, though the strongest brands kept recurring. Measure how often you appear across runs. A single screenshot proves very little.
The shortlist audit
Open ChatGPT, Gemini and Perplexity in separate sessions. For each, ask the questions a real buyer would:
- "Who are the leading [category] vendors for a company our size?"
- "I run [finance / operations / marketing] at a [industry] company. How should I evaluate [category] options?"
- "Compare [your company] with [your top competitor]."
- "What is [your company] known for?"
Run each question five times, and repeat questions one and two with the role changed: a CFO, a COO, a CMO, a CTO. Record how often you appear and grade what comes back on four points.
- Are you named, and how often? If you rarely appear across ten to fifteen variations, you are close to invisible in this channel.
- Is it accurate? Models sometimes assign offerings you do not sell or omit ones you do. A wrong description is worse than silence, because the buyer arrives with the wrong expectations.
- Is it specific? "A well-regarded provider" is a non-answer. "A firm that builds renewal systems for telecom accounts" is a real answer.
- Does it change by persona? If you appear for the CMO question and vanish for the CFO question, you have a gap that generic content will not close.
Ask a follow-up on any answer: "Where did that information come from?" Perplexity often cites sources. Check whether your site, profiles and third-party mentions tell the same story in the same words.
Four signals to strengthen
These are the signals practitioners work on. No controlled study has isolated their individual effect, so treat them as sound hygiene with a plausible payoff.
- Entity consistency. Your company name, offerings, leadership names and descriptions should match, word for word, across your site, LinkedIn, Crunchbase, directory listings and any press or bios.
- Topical depth over breadth. One page that mentions twelve capabilities gives a model nothing to cite. A page per capability that covers the problem, your method, the evidence and what an engagement looks like gives it something to quote.
- Structured data that describes what you actually do. Schema markup that lists your services as distinct entities, with clear relationships between the company, its people and its offerings.
- Third-party language that matches your claims. Case studies with specifics, customer quotes that name the problem and the result, press mentions, and profiles that describe you consistently.
What to fix first
If you do one thing this week, run the audit and read what the machines say about you. You cannot fix a signal you have not measured.
Fix entity consistency first, because it is usually a few hours of cleanup. Then go deep on the three capabilities where you most want to be the unambiguous answer. The shortlist is still formed by reputation and familiarity, and AI is becoming one of the places that reputation gets read back to your buyers.
The Demand and AI Visibility section of the Leverage Diagnostic shows how mature this is in your company, and what it is worth to fix.