---
title: "How AI agent companies get onto buyers’ AI search shortlists"
description: "By making their agent easy to verify: clear use cases, public security and governance evidence, honest pricing and independent proof that assistants can find."
canonical: "https://underneath.agency/resources/ai-agent-companies-customers-from-ai-search"
published: 2026-10-09
updated: 2026-10-10
publisher: "Underneath (https://underneath.agency/agent)"
entity: "https://underneath.agency/.well-known/entity.json"
---
Guide · AI search

# How do AI agent companies get onto buyers’ shortlists in AI search?

By being easy to tell apart from the hype: a clearly named use case, public evidence of what the agent does and how it is governed, and independent proof that AI assistants can find. Buyers of AI agents are curious but wary, and a growing share of their research now runs through ChatGPT, Gemini, Perplexity and Google’s AI features. The agent companies that get named are, we expect, the ones that answer the buyer’s safety questions before they are asked.

## The short version

1. [Gartner](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) expects 33% of enterprise software applications to include agentic AI by 2028, up from less than 1% in 2024, but estimates that only about 130 of the thousands of agentic AI vendors are real.
2. The same firm predicts over 40% of agentic AI projects will be canceled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls.
3. Trust is falling as adoption rises: in [Capgemini’s survey](https://www.capgemini.com/news/press-releases/trust-and-human-ai-collaboration-set-to-define-the-next-era-of-agentic-ai-unlocking-450-billion-opportunity-by-2028/) of 1,500 large-company executives, confidence in fully autonomous AI agents dropped from 43% to 27% in a year.
4. Security is the sharpest worry: 80% of organizations told [SailPoint](https://nhimg.org/wp-content/uploads/2025/09/SailPoint-The-Rising-Risk-of-AI-Agents-Expanding-the-Attack-Surface-report-SP2648-.pdf) their AI agents had performed unintended actions, and only 44% had governance policies for them.
5. Spending is still early: [Menlo Ventures](https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/) puts agent platforms at $750 million in 2025, 10% of horizontal AI application spend, with copilots taking 86%.

## Which companies are buying AI agents, and how much can one account grow?

Mostly large companies moving from pilots to production; a customer’s value grows with every task the agent takes on.

Most buyers are early. Capgemini found that nearly a quarter of large organizations had launched agent pilots and 14% had begun implementation, while only 2% had fully scaled deployment. A [Gartner poll of 147 CIOs](https://www.gartner.com/en/newsroom/press-releases/2025-06-11-gartner-predicts-that-guardian-agents-will-capture-10-15-percent-of-the-agentic-ai-market-by-2030) found 24% had deployed a few agents and 50% were researching and experimenting. Intent is strong: in [Microsoft’s 2025 Work Trend Index](https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born), 81% of leaders expected agents to be moderately or extensively integrated into their AI strategy within 12–18 months.

The buyer depends on the job. Asked about use cases, a majority of 125 leaders in the same Gartner poll said their agents focus on internal administration such as IT, HR and accounting, and 23% on customer-facing work. Developers are a buying group of their own: among developers who use agents at work, 84% use them for software development, per [Stack Overflow’s 2025 survey](https://survey.stackoverflow.co/2025/ai). Smaller companies move faster. In [LangChain’s survey](https://www.langchain.com/stateofaiagents), 51% of respondents had agents in production, rising to 63% at mid-sized companies. Customer-facing voice agents for contact centers follow their own path, covered in [our guide to AI voice companies](https://underneath.agency/resources/ai-voice-software-revenue-from-ai-search).

Pricing makes the customer’s value elastic. Many agent vendors charge per task rather than per seat. Intercom prices its Fin customer service agent at [$0.99 per outcome](https://fin.ai/pricing), and Salesforce sells [Agentforce Flex Credits](https://www.salesforce.com/agentforce/pricing/) at $500 per 100,000 credits, paid per action. Revenue therefore depends less on the signature than on how far the agent spreads once it works. Capgemini projects that organizations with scaled agent deployments will generate about $382 million each over three years, against about $76 million for others, which is the size of prize buyers are weighing.

## How do buyers evaluate an AI agent before they buy?

Through pilots and proofs of concept, judged on reliability, cost and control.

Agents are judged harder than ordinary software because they act. LangChain found performance quality was the top concern for small companies, cited by 45.8%, against 22.4% for cost. Many teams still have humans checking agent output by hand. Gartner warns that most agentic projects today are “early stage experiments or proof of concepts that are mostly driven by hype,” which is why so many will be canceled.

Buyers are also learning to discount labels. Gartner describes “agent washing,” the rebranding of chatbots, assistants and robotic process automation as agents without real agentic capabilities. [Menlo Ventures’ data](https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/) suggests the skepticism is warranted: only 16% of enterprise deployments it studied qualified as true agents that plan, act, observe and adapt.

Software buyers in general now start that evaluation with AI. In [G2’s July 2026 survey](https://sell.g2.com/2026-buyer-behavior-report) of 1,038 software decision-makers, more than 80% had sourced software recommendations from an AI chatbot in the last two years, and 39% named IT security review as the biggest delay between choosing a vendor and buying. That survey covers all software, not agents alone, and comes from a review platform with a stake in the topic.

## Which questions do agent buyers ask AI assistants?

Mostly use-case, comparison and safety questions, because the category is new and crowded.

We wrote the example prompts below to show how an operations, IT or support leader might shop for an agent; none was observed in real use:

- Use case: “Which AI agents can resolve tier-one support tickets in Zendesk without a human?”
- Build or buy: “Should we build an internal IT help desk agent or buy one?”
- Comparison: “Salesforce Agentforce vs Sierra vs Decagon for retail customer service.”
- Pricing: “How much does an AI SDR agent cost per meeting booked?”
- Governance: “Which AI agent platforms support audit logs, role-based permissions and human approval steps?”
- Legitimacy: “Is [vendor] a real agent or a chatbot with workflows?”

Governance and legitimacy questions carry unusual weight here. In [our reputation study](https://underneath.agency/research/is-it-legit-ai-reputation-study), assistants asked whether a brand was legitimate said yes in every complete answer, but 99.7% raised at least one problem. A buyer asking about an agent vendor will hear whatever concerns the public record contains.

## How does a shortlist mention become usage revenue for an agent vendor?

Through a pilot that reaches production, then usage that grows as the agent takes on more work.

From the pricing and adoption data above, we infer that most agent deals move through four steps:

1. **Shortlist.** A buyer asks an assistant for agents that handle a specific task in their stack. Three to five names come back.
2. **Pilot.** The buyer tests one or two on real data. This is where most deals are lost, because the agent must perform on the buyer’s own workflows.
3. **Production.** Security, legal and finance approve. Here the buyer checks the vendor’s public evidence again, often with an assistant.
4. **Expansion.** With per-outcome or per-action pricing, revenue grows as volumes rise and new use cases are added.

Because revenue scales after deployment, a vendor named early in a category gets more than one deal from each mention: it gets the chance to grow inside the account. The cost of being absent is the reverse. A buyer who pilots two competitors rarely adds a third, and an AI answer that leaves you out produces no visit to measure. For how lost answers show up in pipeline, see [our article on AI answers and pipeline](https://underneath.agency/resources/ai-answers-pipeline-revenue).

## Why does an assistant name one agent vendor over another that looks the same?

The assistant makers explain little; research points to independent coverage, consistent facts and visible proof of what an agent does.

**What OpenAI has published.** Its [ChatGPT search help page](https://help.openai.com/en/articles/9237897-chatgpt-search) says a question is typically rewritten into “one or more targeted queries” for search providers, sometimes followed by more specific ones. Nothing there explains how an agent vendor gets picked. We counted the searches ourselves: in [our hidden searches study](https://underneath.agency/research/ai-hidden-searches-study), ChatGPT averaged 3.7 per buyer question, enough for a single governance question to pull in an agent vendor’s reviews, pricing and security documentation together.

**Research findings.** When [Chen and colleagues](https://arxiv.org/abs/2509.08919) looked at US software questions, 72.7% of the sources AI search used were independent “earned” sites; for Google, the share was 45.4%. [Our brand entity study](https://underneath.agency/research/brand-entity-ai-recommendations-study) found the same pull toward outside sources: each tenfold increase in independent sites naming a brand was associated with 4.7 times the odds of a recommendation. Prices are a weak spot: of the software plan prices assistants quoted in [our pricing study](https://underneath.agency/research/ai-pricing-accuracy-study), only 61.9% fully matched the vendor’s page. Usage-based agent pricing, with credits and outcomes, is harder still to quote correctly.

**Trust factors specific to agents.** Security and governance come first. SailPoint found 23% of organizations reported AI agents being coaxed into revealing access credentials, and 92% said governing AI agents is paramount to enterprise security. Capgemini reports that organizations are prioritizing transparency about how agents make decisions. Gartner predicts that “guardian agents,” which monitor other agents, will account for at least 10 to 15% of agentic AI markets by 2030. We infer that agent vendors with public documentation of permissions, audit trails, human approval steps, evaluation results and certifications give both buyers and assistants something concrete to cite. For a close parallel in a security-led market, see [our cybersecurity article](https://underneath.agency/resources/cybersecurity-software-revenue-from-ai-search).

## Why are agent companies so hard for assistants to tell apart?

Because the category is crowded, new and full of look-alike claims.

Supply is exploding. In Y Combinator’s Spring 2025 batch, 67 of 144 startups were [building AI agents or tools for creating them](https://getcoai.com/news/nearly-50-of-y-combinators-spring-2025-batch-builds-ai-agents/). New companies also start with little for assistants to draw on: in a test of 112 Product Hunt startups, [Sharma](https://arxiv.org/abs/2601.00912) found ChatGPT surfaced them in only 3.32% of discovery questions, even though it recognized almost all of them by name. The causes are set out in [why ChatGPT misses new products](https://underneath.agency/resources/why-chatgpt-misses-new-products), and they apply to any agent startup launched recently.

When many vendors describe themselves with the same words (“autonomous,” “agentic,” “AI workforce”), we infer an assistant has little to separate them except what independent sources say. That is the opening for a vendor that names a narrow job, publishes results for it and earns coverage that repeats them. Comparison pages are one way to make that separation explicit; [our article on comparison pages](https://underneath.agency/resources/do-comparison-pages-help-b2b-ai-citations) reviews the evidence.

## Which GEO work helps an agent vendor stand out from look-alikes?

Making your agent’s use case, proof and safeguards findable and consistent, with no promise of a mention.

For an agent vendor, generative engine optimization (GEO) breaks into six tasks:

1. **A precise entity.** Say what the agent does, for whom, in which systems and with what level of autonomy, the same way on your site, marketplaces (Salesforce AppExchange, Microsoft, AWS), review profiles, GitHub and LinkedIn.
2. **Use-case pages with evidence.** Publish one page per job the agent does, with measured results, limits and the human steps that remain. Vague “autonomous AI workforce” pages give an assistant nothing to quote.
3. **Public trust documentation.** A trust center covering permissions, data handling, audit logs, human approval, evaluation methods and certifications answers the questions buyers and security teams ask first.
4. **Clear pricing.** State your unit (per outcome, per action, per seat) and a worked example, and keep one current page so assistants do not quote stale credits or plans.
5. **Independent proof.** Earn analyst mentions, practitioner reviews, case studies told by customers, integration partner listings and community discussion. Earning that standing step by step is covered in [building authority for AI search](https://underneath.agency/resources/how-brands-build-authority-for-ai-search).
6. **Stage-by-stage tracking.** Re-ask use-case, comparison, pricing and governance questions in ChatGPT, Gemini, Perplexity, Claude, Copilot and Google, and link what you find to incoming pilot requests.

## Where is the evidence on AI search for agent vendors still missing?

Mostly at the money end: no study ties AI visibility to pilots, production deployments or usage revenue for agent companies.

The adoption figures come from consultancies, analyst firms, investors and vendors, and they disagree: 51% of LangChain’s respondents, many from technology firms, had agents in production, while Capgemini found 2% of large organizations fully scaled. Definitions of “agent” differ between surveys, which is part of the agent washing problem. The security figures come from SailPoint, which sells identity security. Gartner’s cancellation and adoption figures are predictions, not measurements. And no published study shows how assistants weigh security documentation or certifications when naming an agent vendor; our trust-factor points above are inference.

## How can an agent vendor find out whether AI answers are costing it pilots?

Check whether assistants name you, describe your agent accurately and answer safety questions from your own evidence.

Put your use-case, comparison, pricing and governance questions to the main assistants, then set the answers beside the pilots and production deployments you won and lost, so you can see which gaps cost you shortlists and, later, per-outcome usage. We can run that comparison with you and plan the trust documentation and independent proof that would close the gaps: [talk to us about an agent visibility audit](https://underneath.agency/contact). Our [generative engine optimization service](https://underneath.agency/services/generative-engine-optimization) page sets out how that work runs, from mapping the governance questions buyers ask to measuring whether new evidence changes the answers.

## Frequently asked questions

### Do AI assistants recommend AI agents by name?

Yes, when asked category questions, they typically name products. How they choose is not documented by the platforms. Studies of software questions find assistants lean on independent sites, so reviews, analyst coverage and case studies matter.

### What do enterprise buyers check before buying an AI agent?

Reliability on their own data, cost and control. Surveys point to security and governance as the main brakes: SailPoint found 80% of organizations had seen unintended agent actions, and Capgemini found trust in fully autonomous agents falling.

### Should an agent company publish its security and governance details?

Yes. Buyers ask about permissions, audit logs and human approval early, and an assistant can only repeat what is public. A gated or missing trust page leaves the answer to others.

### Does “agent washing” hurt real agent companies in AI search?

It makes them harder to tell apart. When many vendors use the same words, specific, verifiable claims and independent proof are what separate a real agent from a rebranded chatbot.

## Sources

- Gartner (2025), [Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027)
- Gartner (2025), [Gartner Predicts that Guardian Agents will Capture 10-15% of the Agentic AI Market by 2030](https://www.gartner.com/en/newsroom/press-releases/2025-06-11-gartner-predicts-that-guardian-agents-will-capture-10-15-percent-of-the-agentic-ai-market-by-2030)
- Capgemini (2025), [Trust and human-AI collaboration set to define the next era of agentic AI](https://www.capgemini.com/news/press-releases/trust-and-human-ai-collaboration-set-to-define-the-next-era-of-agentic-ai-unlocking-450-billion-opportunity-by-2028/)
- SailPoint (2025), [AI agents: The new attack surface](https://nhimg.org/wp-content/uploads/2025/09/SailPoint-The-Rising-Risk-of-AI-Agents-Expanding-the-Attack-Surface-report-SP2648-.pdf)
- Menlo Ventures (2025), [2025: The State of Generative AI in the Enterprise](https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/)
- Microsoft (2025), [2025: The Year the Frontier Firm Is Born](https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born)
- LangChain (2024), [State of AI Agents](https://www.langchain.com/stateofaiagents)
- Stack Overflow (2025), [2025 Developer Survey: AI](https://survey.stackoverflow.co/2025/ai)
- G2 (2026), [2026 Buyer Behavior Report: The Evaluation Maze](https://sell.g2.com/2026-buyer-behavior-report)
- Intercom (2026), [Fin pricing](https://fin.ai/pricing)
- Salesforce (2026), [Agentforce pricing](https://www.salesforce.com/agentforce/pricing/)
- CO/AI (2025), [Nearly 50% of Y Combinator’s spring 2025 batch builds AI agents](https://getcoai.com/news/nearly-50-of-y-combinators-spring-2025-batch-builds-ai-agents/)
- OpenAI (2025), [ChatGPT search](https://help.openai.com/en/articles/9237897-chatgpt-search)
- Chen and colleagues (2025), [Generative Engine Optimization: How to Dominate AI Search](https://arxiv.org/abs/2509.08919)
- Sharma (2025), [The Discovery Gap: How Product Hunt Startups Vanish in LLM Organic Discovery Queries](https://arxiv.org/abs/2601.00912)
- Underneath (2026), [hidden searches](https://underneath.agency/research/ai-hidden-searches-study), [brand entity](https://underneath.agency/research/brand-entity-ai-recommendations-study), [software pricing accuracy](https://underneath.agency/research/ai-pricing-accuracy-study) and [“is it legit” reputation](https://underneath.agency/research/is-it-legit-ai-reputation-study) studies

---

This is the Markdown twin of https://underneath.agency/resources/ai-agent-companies-customers-from-ai-search. The HTML page is canonical. Publisher: Underneath, https://underneath.agency/agent. Site index: https://underneath.agency/llms.txt.