Buyers Build the Shortlist Inside an AI Assistant. Learn to Sell to the Model

94% of business buyers used AI somewhere in their most recent purchase, up from 89% a year earlier. Around 90% of enterprise software deals still close with a vendor the buyer already had in mind on day one. Those two facts meet in the AI shortlist, which is the answer an assistant returns when a buyer asks who to consider. If your name is missing from it, you are not losing those deals. You were never in them.

What the AI shortlist actually is

The AI shortlist is the set of vendor names an assistant produces when a buyer describes a problem and asks who solves it. It forms in a private chat window, before any form fill, before any demo request, and before your sales team has a lead to work. It is a consideration set assembled by a model reading whatever it can find about your category.

That description sounds like search, and it is not. A search results page hands a buyer ten links and makes them choose. An assistant hands them three names and a reason for each. The filtering that a buyer used to do is now done upstream, by something that never visited your pricing page and never spoke to your team.

Forrester's John Buten put the shift plainly in January 2026. Buyers named generative AI or conversational search as a meaningful information source at twice the rate of any other option, including vendor websites, product experts and direct contact with sales. That is not a channel gaining share. That is a channel that has taken the top position from the assets you own.

The shortlist closes before your funnel opens

The uncomfortable part is not that AI is involved. It is that the consideration set has always closed early, and AI has moved the closing point earlier still.

Bain published research in June 2026 based on 750 responses fielded by NewtonX in February, covering human capital management and cloud data platform purchases. The finding was that around 90% of buyers purchase from their Day 1 list, meaning the vendors already in mind before formal evaluation starts. Bain's own summary is blunt about the consequence. Most seemingly lost deals were never winnable, because the vendor was not on the initial list.

Two datasets, three years apart, one finding

6sense reached the same conclusion from a different direction. Its Buyer Experience Report surveyed more than 900 B2B buyers in June and July 2023, restricted to purchases above $10,000 in annual value. It found that 84% of buyers said the first vendor they contacted ultimately won the business, and that buyers do not engage a seller until roughly 70% of the way through their journey.

Read those two studies together and the sales funnel looks different. By the time a buyer appears in your CRM, the outcome is largely determined. The visible sales cycle is mostly a confirmation ritual for a decision made in private.

Gartner's 45% and Forrester's 94% are not contradictory. They measure different things: one is generative AI specifically as an information source during a purchase, the other is any AI use anywhere in the process. Read the basis column before comparing any two rows.

Most of this research never touches your analytics

Here is the finding that should reach a revenue leader rather than a marketing team. Forrester reported that 61% of business buyers use private AI tools provided by their own organisation. More than half of ChatGPT and Microsoft Copilot users are working inside a private version deployed behind the corporate firewall.

A private deployment sends no referrer. It produces no session in your analytics, no UTM parameter, and no entry in any AI visibility dashboard sold on the market today. The buyer forms an opinion about your category and you receive no record that the conversation happened.

This is why the pattern many teams are reporting makes sense. Organic sessions fall, brand search holds or rises, and pipeline quality changes shape without changing volume. The research did not stop. It moved somewhere you cannot instrument. We looked at the measurable half of that shift in the analysis of why AI referral traffic converts at a multiple of organic. The private-deployment share means even those numbers understate what is happening.

My position on this is firm. Any vendor selling you a complete picture of your AI visibility is selling a partial picture with confidence, because the private tier is structurally unmeasurable from the outside. Treat every AI visibility number you buy as a sample of the public tier only.

What buyers say makes a vendor stand out in an AI answer

Semrush asked 622 US B2B professionals about this directly between March and April 2026. The answers are more useful than any ranking study, because they describe what a human notices once the model has already produced a list.

The complaints are a product brief

The same survey asked what frustrates buyers about AI vendor recommendations. 33% said the recommendations were too generic for their use case. 27% said the answers did not reflect real pricing or contract terms. 25% said relevant vendors were missing entirely.

Read that as a list of gaps you can fill rather than a critique of the tools. If a third of buyers find the answers too generic, the vendor whose public material is specific enough to survive a narrow prompt gets pulled into answers the generalists cannot reach. The pricing complaint is its own opportunity, and the accuracy problem behind it is examined in the test of how often assistants get software pricing right.

Four presence tactics that survive contact with the evidence

Most advice in this area is asserted rather than tested. What follows is limited to actions justified by the buyer-reported data above, which is a smaller set than the usual playbook.

1. Publish the use case, not the category

53% of buyers said the deciding signal was a match to their specific use case. A category page describes what you are. A use case page describes a situation a buyer recognises as their own, and it is the only asset that can match a prompt containing constraints.

The practical form is narrow. One page per situation, naming the industry, the team size, the system it replaces and the constraint that makes the situation hard. Write the sentence a buyer would type, then answer it in the first paragraph.

2. Put real numbers where a model can read them

27% of buyers complained that AI answers do not reflect real pricing or contract terms. If your pricing lives inside a PDF, a form-gated calculator or a sales conversation, no model can state it, and the answer a buyer receives about your pricing will be a guess.

Publish the numbers you can defend in crawlable text. Seat prices, usage tiers, minimum terms, overage rates. This is not a marketing decision, it is a decision about whether the machine describing you to your buyer has anything accurate to say.

3. Write your own comparison page, and concede something real

Comparison and best-of formats dominate the material assistants draw on for commercial questions. That is well established and it is also where most vendors write the least credible content, because they win every row of their own table.

A comparison page that concedes a genuine weakness is more likely to be quoted, and it is more likely to survive the buyer's validation step. I would rather be cited accurately with one loss on the table than not cited at all. The wider evidence on which formats earn citations is set out in the review of what GEO can actually demonstrate.

4. Be present on the third-party surfaces buyers check next

The AI answer is not the end of the journey. After a model names a vendor, 71% of buyers visit that vendor's website, 63% search the company name, and 38% check a review platform. Those are the steps where an unprepared vendor loses a place it just won.

The implication is unglamorous. Your site needs to confirm what the model said within one scroll, your review profiles need recent entries, and your G2 or Capterra category needs to exist. A buyer who arrives from an AI answer and finds a mismatch drops you at the validation stage, having been shortlisted.

The model shortlists. A human still closes.

Anyone concluding from the above that sales headcount is now optional should read Gartner's May 2026 research. It found that 69% of B2B buyers prefer to validate AI-generated insights with a sales rep, even though 67% say they would prefer an experience without one.

That contradiction is the whole job. Buyers want a self-directed process and they do not trust the output of it. Gartner also found 51% of buyers say they are more likely to meet misleading information from generative AI, against 49% who say the same about a sales rep. The two sources are now viewed as roughly equally unreliable, which is a strange compliment to both.

Robert Blaisdell, VP analyst at Gartner Sales, framed it as a role change rather than a removal. Buyers navigate the purchase themselves, and that does not eliminate the seller. Forrester's data points the same way: 36% of buyers said generative AI made them more confident in their decision, while 20% said it made them less confident because they hit unreliable output.

So the sequence has two stages with different owners. The model decides who is considered. The human decides who is chosen. Investing in one and not the other loses at whichever stage you skipped.

Where this argument is weakest

Every post on this subject sells a service at the end. This one has three problems it cannot resolve, and you should weigh them before reallocating a budget.

The core evidence is self-reported

Forrester, Gartner, Semrush and 6sense all asked buyers what they did. People are unreliable narrators of their own decision process, and they systematically over-report tool use that feels current. None of these studies observed a buying process end to end. The direction of the finding is corroborated across four independent samples, which is meaningful. The precise percentages are not.

The Semrush figures also come from a company that sells visibility tooling into this exact problem. The methodology is published, the sample is defined, and the incentive still exists. I use those numbers because the methodology is disclosed, and I would not use them if it were not.

The Day 1 finding is older than the AI shortlist

Bain's roughly 90% figure and 6sense's 84% describe a pattern that predates assistants entirely. Buyers had preferred vendors in mind long before ChatGPT existed. It is plausible that AI simply automates a shortlist that used to come from peers, analysts and memory, without changing who ends up on it.

If that is true, the correct response is much smaller than the one this article implies. Nobody has published a study connecting presence in AI answers to closed revenue, with a control group. Until someone does, treat the causal claim as unproven and the correlation as strong.

Bain's sample is two categories, not a market

The 750 respondents covered human capital management and cloud data platforms. Those are mature categories with established incumbents, which is exactly where a Day 1 list effect should be strongest. In a young category with no obvious leader, the same effect is probably weaker. Do not carry the 90% into a market that looks nothing like enterprise HR software.

What to measure instead of a visibility score

The instinct is to buy a tool that reports a number and watch it rise. That number is a sample of the public tier, and it tells you very little about revenue. Three measurements are more useful and all three are cheap.

The third row is the one that matters and it costs nothing but a habit. It is also the only one of the three that reaches the private tier. The buyer sitting in front of you can tell you what a tool you cannot instrument said about your company. That question is a free instrument pointed at the invisible part of the funnel, and almost nobody is asking it. Related evidence on how these answers overlap with conventional rankings is in the comparison of ChatGPT citations against Google rankings. The volume problem behind all of it is in the analysis of content saturation and visibility.

Frequently asked questions

Where to start this month

Three actions, in order of how quickly they pay back.

First, add one question to your discovery call script this week: did you use an AI assistant while researching this, and what did it tell you about us. Log the answers for 30 days. You will learn more from 20 of those answers than from any visibility subscription, and you will learn it about the tier nobody can measure.

Second, take your single most valuable use case and give it a page that names the industry, the team size, the system it replaces and the constraint that makes it hard. Publish the pricing that goes with it. That page exists to be quoted, not to be clicked.

Third, run your 20 unbranded prompts once and write down what the answers say about you. If the answers are wrong, that is a content problem you can fix. If your name is absent, that is a Day 1 list problem, and it is costing you deals that never reached your CRM in the first place.

Related reading