When the Buyer Is a Bot: AI Agents and the Coming Shortlist Economy

Somewhere right now, a buyer has typed "find me three consultants who can fix this, with evidence" into an assistant, and gone to make coffee. The research meeting happened. Nobody from marketing was invited.

What is actually asking the questions now?

Let us be precise, because this topic attracts fog. An AI agent, in the form that exists today, is a language model given tools: it can search, fetch pages, read documents, and assemble what it finds into a structured output. The deep-research modes in the major assistants are the most visible version. You give one a brief, it spends several minutes reading dozens of sources, and it returns a comparison, a recommendation, or a shortlist with citations.

That is not science fiction, and it is not a robot with a credit card. It is a tireless junior analyst who reads faster than any human and clicks on almost nothing. The behaviour already shows up in buyer data: a G2 buyer-behaviour survey from March 2026 found that 51% of B2B buyers start their research in an AI chatbot more often than in Google, and 71% use AI chatbots somewhere in their research. The human still signs the contract. Increasingly, though, the human signs it from a menu a machine wrote.

The infrastructure is visible if you know where to look. OpenAI operates three distinct crawlers: GPTBot gathers training data, OAI-SearchBot builds the search index, and ChatGPT-User fetches pages live when a user or an agent asks about something in the moment. When your site logs show ChatGPT-User, that is not an abstraction. That is a machine reading your pages because somebody, somewhere, asked a question you might be the answer to.

How does an agent research a purchase?

Roughly the way a diligent, sceptical, slightly literal-minded researcher would, with three important differences. First, it reads everything as text: your positioning is whatever your words and structure actually say, not whatever your design implies. Second, it cross-checks: a claim on your site that no third party corroborates carries little weight, which is why where AI looks before recommending someone matters more than what you say about yourself. Third, it summarises ruthlessly: from your twenty pages, perhaps three sentences survive into the output the buyer reads.

The difference between what a human buyer notices and what an agent extracts is worth staring at directly:

A worked example: shortlisting a fractional CFO

Make it concrete. A founder types a brief into a deep-research mode: "Find me three fractional CFOs with healthcare experience, UK-based, who have taken a company through Series B, and show your evidence." Watch what the machine does with each candidate in the market.

Candidate one has a handsome site that says "trusted finance partner for ambitious companies". The agent extracts no sector, no stage, no geography, no cases. She matches zero of the four constraints in text and is eliminated in the first pass, not because she lacks the experience, but because she never wrote it down where a machine could read it.

Candidate two has the experience buried in a PDF brochure and a LinkedIn profile that gives a different job title from his website. The agent finds partial matches, notices the contradiction, and either drops him or includes him with a hedge. Hedged candidates rarely survive when the brief asked for three names and five clean matches exist.

Candidate three has a plain page stating exactly who she serves and at what stage, two dated case studies with named companies, a podcast interview where she discusses a Series B raise, and consistent details everywhere the agent checks. She is not the most experienced of the three. She is the most verifiable, and she is the one whose name the founder reads over coffee. That, in miniature, is the entire shortlist economy: the market did not choose the best candidate, it chose the most legible one, and nobody involved ever knew the others existed.

What survives the agent's filter?

Specificity, consistency, and corroboration. That is nearly the whole list. An agent compiling a shortlist of, say, fractional CFOs for a healthcare startup is pattern-matching on evidence: who is described, by more than one source, as doing exactly this work, for exactly this kind of client, with results stated concretely? The generic contender ("strategic finance leader passionate about growth") gives the machine nothing to grip. The specific one ("fractional CFO for seed to Series B healthcare companies, cases published, quoted in two trade outlets") practically shortlists herself.

This is why the questions buyers ask matter so much. We have catalogued the prompts buyers type into AI, and agent-delegated briefs are even more constraint-rich than chat questions: budget bands, sector, geography, evidence requirements. Every constraint is a filter, and every filter eliminates whoever failed to state the matching fact anywhere a machine could read it. You are not competing on persuasion at this stage. You are competing on retrievability.

Where does the hype end?

Honesty compels a few deflating facts, because this field has a hype problem. AI referral traffic is still only about 1.08% of all website traffic across the ten industries in Conductor's 2026 benchmarks. Agents hallucinate, misread pages, and inherit whatever staleness sits in their training data. Most purchases still involve plenty of ordinary human googling, asking around, and gut feel. Anyone telling you the entire funnel is agentic already is selling something.

But look at what the small stream carries. Those same benchmarks show ChatGPT referrals converting at 14.2% to 15.9%, Perplexity around 10.5%, and Claude up to 16.8%, against roughly 1.76% for Google organic. A visitor who arrives from an AI answer converts at many times the ordinary rate, because the comparison shopping happened before the click, inside the machine. That is the shortlist economy in miniature: fewer visitors, drastically pre-qualified, because a bot did the elimination round. And elimination rounds are unforgiving in a way humans are not. A human buyer might stumble across you at a conference despite your thin website. The agent will not. It has no serendipity.

How do you get on a machine-made shortlist?

By doing deliberately what the agent's filter checks accidentally. The work is unglamorous and compounds quietly. A practical checklist:

• State the shortlistable facts in plain text. Who you serve, what you do, where, at what scale, with what outcomes. On your own site, in crawlable HTML, since your website matters more now, not less.
• Name things machines can verify. Named clients where permitted, real numbers, dated work. Vague social proof reads as noise.
• Keep every profile consistent. One name, one title, one story across your site, LinkedIn, directories, and bios. Contradictions make an agent hedge, and hedging machines drop names.
• Earn corroboration. Interviews, reviews, press, community mentions. An agent weighs what others say about you far above what you say about yourself.
• Let the readers in. Check your robots.txt is not blocking the retrieval crawlers you actually want, such as OAI-SearchBot and ChatGPT-User; the AI crawler landscape is worth twenty minutes of your attention.
• Test the brief yourself. Run a deep-research query a buyer would run, in your category, and see whether you appear. Nothing focuses the mind like reading your own absence.

And no, you cannot simply buy your way in. We have examined whether you can pay to appear in ChatGPT, and the short answer is that organic answers are not currently for sale, which makes evidence the only reliable currency. If you would rather run this as a structured programme than a hobby, our services exist for exactly that.

Myth: this only matters for big, expensive purchases

It's tempting to file all of this under enterprise procurement and move on, since "AI agent shortlisting a vendor" sounds like something happening in a large company's back office. But the same behavior shows up at much smaller scale, someone asking an assistant to find a wedding photographer, a divorce lawyer, a freelance designer, comparing a handful of options on stated experience, location and reviews before ever visiting a single website themselves. The brief is smaller and less formal, but the mechanism is identical, plain-text facts win, vague confidence loses. Anyone whose work gets chosen through comparison, which is most people, is already subject to this filter whether or not they've noticed it yet.

What does this do to marketing as we knew it?

It splits the funnel into two audiences with opposite tastes. The human downstream still wants story, warmth, and a reason to choose you over the other two names on the shortlist. The machine upstream wants facts, structure, and third-party proof, and it decides whether the human ever hears of you. Most marketing budgets are still weighted overwhelmingly toward the first audience, which made sense right up until the second one started holding the door.

The shortlist economy is not the end of persuasion. It is a new admission fee charged before persuasion begins. The professionals who will do well in it are not necessarily the loudest or the best funded. They are the ones a diligent machine can read, verify, and confidently repeat. Being that person is a choice, and it is available now, while most of your competitors are still writing for readers who blink.