• AI visibility

Guide · AI search

How do machine learning companies get found when enterprise buyers ask AI who to hire?

By earning the independent, checkable proof that AI assistants draw on when a buyer asks who can build or supply a machine learning solution, then turning that first mention into a paid pilot that reaches production. Enterprises now buy most of their AI rather than build it, and they look hard for vendors who can show results. Whether an assistant names you depends far more on what others publish about your work than on what your own site claims.

By Underneath Team Guide Updated

The short version

  1. Enterprises spent $37 billion on generative AI in 2025, up from $11.5 billion in 2024, and 76% of AI use cases are now purchased rather than built internally, against 53% purchased a year earlier, according to Menlo Ventures (opens in a new tab).
  2. Buying from outside works better for buyers: in MIT research reported by Fortune (opens in a new tab), purchasing AI tools from specialized vendors and building partnerships succeeded about 67% of the time, while internal builds succeeded only one-third as often.
  3. Once an enterprise commits to exploring an AI solution, 47% of AI deals reach production, against 25% for traditional software, Menlo Ventures found.
  4. Buyers have been burned: in IBM’s survey (opens in a new tab) of 2,000 CEOs, only 25% of AI initiatives had delivered the expected return and only 16% had scaled enterprise-wide.
  5. Being known is not being found: in a test of 112 Product Hunt startups (opens in a new tab), ChatGPT recognized 99.4% when asked by name but surfaced only 3.32% for discovery questions such as “What are the best AI tools launched this year?”

Who pays for machine learning work, and what is a client worth?

Enterprise teams buy it, mostly from outside vendors now, starting with a pilot and growing into production contracts.

Spending on AI is huge and still climbing. Gartner (opens in a new tab) forecasts worldwide AI spending of nearly $1.5 trillion in 2025 and more than $2 trillion in 2026, though much of that is hardware and AI built into existing products. Business adoption is broad: the Stanford AI Index (opens in a new tab) reports that 78% of organizations used AI in 2024, up from 55% the year before.

Two kinds of machine learning company sell into that demand, and they monetize differently:

  • Machine learning consultancies and service firms sell expertise: a discovery phase, a paid proof of concept, then a build, deployment and support engagement. The buyer is often a chief data or technology officer, or a business-unit leader with a specific problem such as fraud, forecasting or document processing.
  • Machine learning product companies sell models, platforms or applications on subscription or usage pricing. By Menlo Ventures’ count, 27% of AI application spend comes through product-led motions, where individual users adopt first, nearly four times the traditional software rate. Platform vendors are covered in how data science platforms win enterprise buyers.

The shift toward buying is what makes discovery matter. In 2024, Menlo found 47% of AI solutions were built internally; in 2025, 76% of use cases were purchased. Startups captured 63% of the AI application market, up from 36% the year before, so new vendors can win against incumbents.

We found no reliable public benchmark for the size of a machine learning consulting engagement, so we do not quote one. What the evidence does show is the shape of the value: Menlo counts at least 10 AI products earning over $1 billion in annual recurring revenue and 50 earning over $100 million, and a pilot that reaches production usually becomes a multi-year relationship. We infer that the first pilot is where most of a machine learning vendor’s lifetime revenue with a client is decided.

When do enterprise buyers ask assistants about machine learning vendors?

At the research and shortlist stages, where buyers now ask assistants to explain options and name vendors.

No public survey isolates machine learning buyers, so the best evidence covers software buyers in general. In G2’s 2026 buyer behavior report (opens in a new tab), more than 80% of buyers had sourced software recommendations from an AI chatbot in the last two years, and half of those said AI had the greatest influence during shortlisting and evaluation. G2 runs a review marketplace and has an interest in that finding.

Google’s own search is just as present. Of the 1,248 US searches in our AI Overview study, B2B software and technology keywords carried an AI Overview 96.0% of the time, more than any of the other seven industries. Google’s announcement of AI Mode describes a “query fan-out” technique (opens in a new tab) that splits a question into multiple related searches across subtopics and then combines what comes back. A single question about a fraud model vendor can therefore pull in case studies, reviews, pricing and compliance pages at once.

A reasonable expectation is that machine learning buyers use these tools more than most. They are technical, they work with AI daily, and the category changes so fast that last year’s comparison articles are already out of date.

Which questions do machine learning buyers ask AI assistants?

Problem-first questions: who can solve this, build or buy, how vendors compare, and what it costs.

Buyers rarely start by naming a vendor. They start with a business problem and narrow from there. We wrote the example prompts below ourselves to show how a CTO or business-unit head might frame a machine learning need; none is a recorded query:

  • Problem to vendor: “Which companies build demand forecasting models for mid-size grocery chains?”
  • Build or buy: “Should a regional bank build its own fraud detection model or buy one?”
  • Alternatives: “Alternatives to building on SageMaker for a team without platform engineers.”
  • Comparison: “Compare machine learning consultancies with experience in medical imaging and FDA submissions.”
  • Cost and timeline: “How long and how much does a computer vision proof of concept usually take?”
  • Risk: “Which vendors can fine-tune a model on our data without it leaving our cloud account?”

Problem and build-or-buy questions decide who is considered. Comparison, cost and risk questions decide who gets the pilot. The answers to the cost questions are often generic, which is a reason to publish your own ranges and timelines where assistants can find them.

How does an AI mention turn into machine learning revenue?

Through the pilot: the assistant names you, the buyer checks your proof, a paid pilot runs, and production follows.

Shortlist. A buyer who asks an assistant for vendors usually leaves with a short list of names. If you are not on it, the buyer may never visit your site or meet your team.

Due diligence. Technical buyers check what the assistant said. They read case studies, papers, code and reviews, and they talk to peers. For machine learning, the proof they want is specific: the problem, the data, the result and how it was measured. Proof matters even more when every rival claims AI, as how AI software firms cut through the hype explains.

Pilot. This is where machine learning deals are most fragile. Gartner predicted (opens in a new tab) that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs or unclear business value. It also predicts that through 2026, organizations will abandon 60% of AI projects (opens in a new tab) unsupported by AI-ready data.

Production and expansion. Pilots that survive tend to convert. Menlo’s 47% production rate for AI deals, against 25% for traditional software, suggests that the hard part is getting into consideration, not closing once there. That puts unusual weight on the first step, where AI assistants now sit.

Why does an assistant name one machine learning firm for a problem and not another?

Assistants don’t explain their picks; research points to third-party coverage, links from other sites and practitioner discussion.

Documented by the platform. Beyond Google’s description of query fan-out, no assistant publishes how it chooses which vendors to name for a question like “who builds fraud models?” Anything more specific is inference.

What researchers found. In the US categories Chen and colleagues (opens in a new tab) tested, 72.7% of AI search sources were independent “earned” sites; for Google the share was 45.4%. In the Product Hunt study, the signals that predicted visibility in Perplexity were the number of referring domains and community presence on Reddit; the products’ own optimization scores showed no correlation with discovery. For consultancies, one common shortcut does not help much: in our study of AI-cited “best X” lists, 24.2% of numbered lists with an identifiable publisher ranked their own publisher first, and such lists made up just 1.1% of all citations.

Trust factors specific to machine learning. We infer that what persuades a skeptical technical buyer also gives an assistant something to cite:

  • Case studies that state the problem, the data, the metric and the measured result, ideally confirmed by the client.
  • Peer-reviewed papers, conference talks and open-source code that show real technical depth. The AI Index notes that nearly 90% of notable AI models in 2024 came from industry, so buyers expect vendors to publish, not just claim.
  • Benchmarks reported honestly, with dates. The AI Index reports scores on one coding benchmark, SWE-bench, rising 67.3 percentage points in a single year; a benchmark claim from last year may already be stale.
  • Clear statements on data handling, security and compliance, since the risk questions above are often where vendors are cut.

What does a machine learning vendor lose by not being named?

The pilot, and with it most of the client relationship, because vendors rarely get added once a shortlist forms.

The Product Hunt study shows the size of the gap for young AI companies: near-perfect recognition by name, but a 3.32% chance of appearing when a buyer asks an open question. A consultancy that is well known to its existing clients can be equally invisible to a new buyer who starts with a problem rather than a name. Our article on why well-known brands miss AI recommendations covers the mechanism.

There is also an opportunity in the other direction. Menlo’s data shows startups taking most of the AI application market from incumbents in a single year. Buyers are open to new names; we infer that the vendors who make their proof easy to find are best placed to benefit. Companies selling AI agents face the same opening, covered in how AI agent companies reach buyers’ shortlists.

How does GEO work for a machine learning company?

It makes your expertise and results easy for assistants to find, verify and cite; it cannot promise a mention.

For an ML consultancy or product company, generative engine optimization (GEO) tends to span six pieces of work:

  1. A clear entity. Describe what you do, for whom and in which domains the same way on your site, LinkedIn, partner directories, cloud marketplaces and conference bios, so assistants have one consistent picture.
  2. Proof others publish. Earn client-approved case studies, coverage in trade and technical publications, podcast and conference appearances, and mentions in independent vendor comparisons. How an ML firm can earn that kind of third-party attention is the subject of our authority-building guide.
  3. Technical depth in public. Publish papers, open-source code, model cards and honest benchmark write-ups with dates and methods. These serve technical buyers and give assistants specific facts to cite.
  4. Answers to the buyer’s real questions. Publish ungated pages on build-or-buy decisions, typical pilot timelines and cost ranges, data requirements and security practices, written for the industries you serve.
  5. Community presence. Contribute where practitioners talk, such as GitHub, technical forums and relevant Reddit communities, since community presence predicted visibility in the Product Hunt study.
  6. Measurement tied to pipeline. Track a fixed set of problem-first buyer questions across ChatGPT, Gemini, Perplexity, Copilot and Google’s AI features, repeated over time, and compare with discovery calls and pilots. For new companies, our article on why ChatGPT misses new products explains the timing problem.

Which gaps remain in the evidence on how ML buyers use AI assistants?

Nobody has measured how ML buyers use assistants, or whether a mention leads to a pilot.

The buyer data comes from general software surveys, several run by companies with an interest in the answer. Menlo Ventures invests in AI companies, and its market figures are its own estimates. The MIT findings come from interviews, a survey and public deployments, reported by Fortune. The Product Hunt study tested one cohort of startups with two assistants. No published study yet follows machine learning buyers from an AI answer to a signed pilot or a production contract, so any claim that AI visibility drives revenue in this industry should be tested against your own pipeline.

How can a machine learning firm learn whether assistants put it in line for pilots?

Ask assistants about the business problems you solve, not your company name, and see whether you are named.

A useful first step is an audit of the problem-first, build-or-buy, comparison and risk questions your buyers ask, across the main assistants, checked against where your discovery calls and pilots actually come from. That shows which proof is missing and which gaps cost the most paid pilots. To run that audit with us and plan a route onto more enterprise shortlists, ask us for a review of your machine learning visibility. How case studies, technical publishing and problem-first pages are then built and tracked against pilots is explained on our generative engine optimization service page.

Frequently asked questions

Do AI assistants recommend machine learning consulting firms?

They can name vendors when asked, but no platform documents how it chooses them. Studies of AI search find it leans on independent sources more than Google does, so third-party coverage and client-approved case studies matter more than self-description.

Do published papers and open-source code help an ML company appear in AI answers?

No study has measured that directly for machine learning firms. Papers and code are public, specific and often discussed by others, which makes them useful evidence for buyers and plausible material for assistants to draw on.

Should a machine learning consultancy publish its own “top ML firms” list?

It is a weak bet. In our study, about a quarter of AI-cited numbered lists ranked their own publisher first, but such lists were a tiny share of citations. Independent comparisons carry more weight.

How do we measure whether AI assistants send us pipeline?

Track how often assistants name you for a fixed set of buyer questions, ask every new lead how they found you, and follow those leads through to pilots and production contracts.

Sources

Free strategy call

Some questions are easier to answer about your own business.

Bring the one that matters most. On a free 30-minute call we’ll take a first look at it and send you a short written read afterward.

Ask about my businessHow we work