The answer
Agent-led growth is a new go-to-market (GTM) motion. In agent-led growth, artificial intelligence (AI) agents research, test and pick vendors for the buyer. 94% of business buyers already use large language models in research. Marketers and GTM leaders start by making the product findable, evaluable and actionable for agents, then run a one-quarter plan.
What is agent-led growth?
Agent-led growth is a new go-to-market (GTM) motion in which artificial intelligence (AI) agents act for the buyer. The agent finds vendors, reads documentation and tests the product. It compares options and can complete the purchase. The buyer then reviews a short list.
A large language model (LLM) is software that reads text and answers questions. An agent is an LLM that can also act. An agent can search, open pages, run tests and start a purchase.
Your homepage still exists, but its first reader changed. The agent reads your documentation, your pricing page and your interface for developers. Then it hands your buyer a short list of two or three vendors. You cannot win a short list that you never appear on.
What are supply-side and demand-side agent-led growth?
Supply-side agent-led growth is your company using agents to sell, and demand-side agent-led growth is your buyer using agents to buy. Ask ChatGPT to define agent-led growth, and you get only the supply side. Both ideas use the same name.
| Idea | Who uses the agent | What the agents do | Effect |
|---|---|---|---|
| Supply-side agent-led growth | Your company, to sell | Write outbound email, draft content, clean the pipeline | Makes your current funnel cheaper |
| Demand-side agent-led growth | Your buyer, to buy | Compare vendors, build a feature table, test your product, recommend a winner | Decides who owns the funnel |
This reading of the distinction is mine, and it follows the distinction that Insight Partners draws. Most marketing teams budget for the supply side. The demand side moves market share.
Why is agent-led growth possible now?
Agent-led growth is possible now because its infrastructure arrived in February 2026. Each growth motion waited for a layer of software first. Sales-led growth scaled because Salesforce gave sellers one system of record. Product-led growth needed product analytics, and account-based work needed intent data.
| Motion | Status | Infrastructure | What changed | Early movers | Marketing's job |
|---|---|---|---|---|---|
| Sales-led growth | Proven | Customer relationship management systems, such as Salesforce | Sellers could hold thousands of relationships in one system of record. | The enterprise software field as a whole | Build the list. Feed the sellers. Measure meetings booked. |
| Product-led growth | Proven | Product analytics tools, such as Mixpanel and Amplitude | Teams could measure a self-serve experience and repeat what worked. | Slack, Dropbox and Zoom | Make the free tier easy to find and easy to start. |
| Account-based experience | Proven | Intent data platforms, such as 6sense | Teams could target a named account instead of a single lead. | Snowflake and DocuSign | Pick the accounts. Coordinate the touches across every channel. |
| Agent-led growth | Forming now | Model Context Protocol, Agent-to-Agent, payment protocols, and WebMCP in Chrome 146 | An agent can now read a site, test a product and finish a purchase for the buyer. | Supabase and Resend, in the developer market | Make the product readable, testable and buyable by an agent. |
The motions, infrastructure and early movers follow the Insight Partners analysis. The "Marketing's job" column is my addition.
What is WebMCP?
Google and Microsoft published WebMCP on February 12, 2026. It shipped in Chrome 146. WebMCP is short for Web Model Context Protocol. It lets a website describe its own forms to an agent.
You add a few HyperText Markup Language (HTML) attributes to a form. You give the form a name and a short description. Chrome then turns that form into a tool the agent can call. You do not have to change your backend.
Before WebMCP, agents worked the way a person works. They took a screenshot, guessed where to click and tried again. Research cited by Insight Partners found that two-thirds of the computing spent on AI-to-web interaction went to that guesswork.
WebMCP joins three other layers:
- Model Context Protocol (MCP) from Anthropic connects agents to tools and data.
- Agent-to-Agent (A2A) from Google lets agents coordinate with each other.
- Payment protocols from Google and Stripe let an agent complete a purchase.
My view: WebMCP is the first protocol decision that belongs to marketing. Your pricing page, your trial signup and your demo request are now conversion paths that an agent can operate. Somebody must decide which ones to open. That decision belongs to demand generation.
Why do defaults compound?
Agents build trust the way people build trust. An agent that finds your product reliable picks it again. Then it recommends the product to the next buyer. That default is a moat, and it grows on its own.
The developer market picked its defaults in under 18 months. Customer relationship management (CRM), human resources, marketing automation and vertical software have not picked theirs yet. That is the open window. The open question is which category goes first, and whether its leader earns the default or loses it.
What does the data show?
The data shows that agents already pick vendors, and buyers follow those picks. All data here comes from "Agent-led growth: The next GTM motion is already here" by Neal Behrend at Insight Partners, published March 13, 2026. The 94%, 77% and two-thirds figures are outside research cited in that article. The reading of the data, the plan and the opinions are mine. Insight discloses investments in OpenAI, Anthropic, StackBlitz and 6sense.
How many buyers use AI in research?
94% of business buyers already use large language models in their research. 77% of buyers purchase the option their AI research favored first. That single pair of numbers should change your media plan.
How much is agent selection worth?
Resend sells transactional email. Resend went from zero to 400,000 users after a 2023 launch. Very little of that growth came from a sales team. When a developer asks Claude Code to add email to a project, Claude Code picks Resend in 63% of tests. SendGrid is a much larger competitor, and Claude Code picks SendGrid in 7% of tests.
| Email provider | Share of Claude Code tests that picked it |
|---|---|
| Resend | 63% |
| SendGrid | 7% |
Insight Partners did not publish the rest of the split.
Supabase tells the same story at a larger scale. Supabase became the default backend for the tools that write code from a prompt.
| Supabase measure | Before | After | Time |
|---|---|---|---|
| Developers | 1 million | 4.5 million | Under 12 months |
| Valuation | $765 million (September 2024) | $5 billion (October 2025) | Roughly 13 months |
"Our sign-up rate doubled in three months because of Bolt, Lovable, Cursor."
Supabase chief executive officer (CEO), quoted by Insight Partners
What do Resend and Supabase share?
None of what they share is a campaign:
- Deep documentation that software can read without help.
- A free tier, so no buyer needs budget approval to start.
- A clean and predictable interface, so the agent makes fewer guesses.
- Heavy presence in model training data, through open-source work and public writing.
Three of those four items sit with engineering and technical writing. In this motion, part of your growth budget buys engineering time.
Which funnel stages move?
The whole funnel moves. Discovery is the smallest part of this shift.
| Stage | What the agent changes |
|---|---|
| Discovery | The agent decides which vendors the buyer sees at all. |
| Evaluation | The agent reads your documentation and tests your product, often before any sales call. |
| Decision | Your seller no longer creates the preference. Your seller confirms it or fights it. |
| Purchase | Agent-started transactions are already live for consumers. Business use is close behind. |
Speed is the part people miss. A developer in Claude Code can move from a problem to a working integration in under five minutes. Insight Partners recorded one run through the full funnel in under four minutes. The funnel got faster, and it moved earlier.
How many of your deals arrive with a favorite already picked?
Apply the two published rates to your own pipeline. Multiply your deals by the share of buyers who research with an agent. Then multiply that result by 77%. The example below uses 100 deals and the 94% rate.
| Step | Deals |
|---|---|
| Deals in the period | 100 |
| Deals researched by an agent (94%) | 94 |
| Deals decided before your first call (77% of those) | 72 |
| Deals still open to persuasion | 28 |
This is a simple model. It assumes those rates hold in your category. If the agent's first choice is you, your agent standing is worth about 72 of 100 deals before a seller says a word. If the first choice is a competitor, your sellers must reverse about 72 formed preferences. That recovery work costs more per deal than it used to.
How do you start an agent-led growth plan this quarter?
Start small, and start where the gap between your marketing, your documentation and your sellers is widest. Agents reward consistency between those three. Agent-led growth sits on top of good marketers and sellers. Each move below is small enough to finish in one quarter.
- Move 10% of paid budget to documentation. Fund a technical writer. Documentation is now demand generation.
- Remove the form in front of your key content. An agent cannot fill in a gate. Every gate is a wall.
- Publish your prices. A hidden price forces the agent to guess or to skip you.
- Add one field to your win and loss form. Ask which vendor the buyer's AI research named first.
- Track token-to-value like you track page speed. Test it monthly. Report it to the board.
- Rewrite the handoff to sales. Calls now arrive later and better informed. Coach sellers to confirm the buyer's preference.
- Ask an agent to buy your product. Record the session. Show it to your executive team.
What are the three tests?
Y Combinator joked about a new motto: build something agents want. Selling asks for more. Insight Partners frames the work as three properties: findable, evaluable and actionable. I agree with that frame, so I use it here.
Findable. Generative Engine Optimization (GEO) is work that makes an AI answer name your company. Answer Engine Optimization (AEO) does the same for answer engines. Both are table stakes now. Review the split between classic search work and this new work every quarter.
Evaluable. Agents read documentation, and they skip pitch decks. Supabase and Resend both spent more on developer documentation than on marketing copy. This week, ask your sellers which questions come up in every demo that your public pages do not answer. Every one of those gaps costs an agent extra work.
Actionable. Work with engineering to describe your key conversion paths for WebMCP. Start with the demo request, the trial signup and the pricing page. A free tier or usage-based pricing removes the budget approval step. Most agent-started deals stop at that step.
How do you score your agent readiness?
Count the items below that are true of your company today. The three properties come from the Insight Partners framework. The nine checks are mine.
| Property | Check |
|---|---|
| Findable | An AI assistant names us when asked to list vendors in our category. |
| Findable | We track how often AI answers cite us, and we review it monthly. |
| Findable | Our category pages answer the buyer question, as well as the product feature. |
| Evaluable | Our full documentation is public and needs no login. |
| Evaluable | Our prices are on the website in plain text. |
| Evaluable | An agent can answer every common demo question from our public pages. |
| Actionable | A buyer can start a trial with no sales approval and no credit card. |
| Actionable | We describe our key forms for WebMCP, so an agent can use them. |
| Actionable | We have watched an agent try to buy our product end to end. |
| Score | What it means | What to do |
|---|---|---|
| 0 to 3 of 9 | An agent cannot yet build a case for you. | Start with public documentation and public prices. |
| 4 to 6 of 9 | An agent can find you and judge you. It still cannot act. | Open one conversion path this quarter. |
| 7 to 8 of 9 | You are ahead of most of your category. | Close the last gap before a competitor does. |
| 9 of 9 | You are set up to earn the default. | Measure token-to-value and drive it down. |
What mistakes should you avoid?
The biggest mistakes are moving too fast on the funnel and too slow on documentation. Buyers arrive further down the funnel now, and they value their time more. That raises the stakes of every human conversation.
| Mistake | Fix |
|---|---|
| Rebuild the whole funnel around agents this year. | Make the seven small moves above. Keep the current funnel running. |
| Cut the sales team. | Keep the sellers. Coach them to confirm or reverse a formed preference. |
| Budget only for supply-side agent work. | Fund the demand side too: documentation, public prices and open conversion paths. |
| Put a form in front of key content. | Remove the gate. An agent cannot fill it in. |
| Audit documentation only against version numbers. | Audit it against token-to-value. Lower is better. |
| Hand WebMCP decisions to engineering alone. | Let demand generation decide which conversion paths to open. |
Common questions
What is token-to-value?
Token-to-value is the number of tokens an AI agent needs to go from a stated problem to a working setup with your product. Lower is better. Audit your documentation against this target, and test it monthly.
How do AI agents choose which vendor to recommend?
Agents favor vendors that are cheap to evaluate. Four things drive the choice: public documentation that software can read, a free tier, a clean and predictable interface, and a strong presence in model training data. Claude Code picks Resend for email in 63% of tests and SendGrid in 7%.
Who owns WebMCP work, marketing or engineering?
Marketing decides, and engineering builds. WebMCP turns your pricing page, trial signup and demo request into paths that an agent can operate. The choice of which paths to open is a demand generation decision.
What are GEO and AEO?
Generative Engine Optimization (GEO) is work that makes an AI answer name your company. Answer Engine Optimization (AEO) does the same for answer engines. Both are table stakes now.
What should a chief marketing officer (CMO) do first about agent-led growth?
Start with three moves that take one quarter. Publish your prices in plain text. Remove the form in front of the content an agent must read. Then point an advanced agent at your own product, ask it to buy, and record where it stops.
Does agent-led growth replace sales teams?
No. Agent-led growth changes the seller's job. Buyers arrive later in the funnel with a preference already formed. The seller confirms that preference or works to reverse it, so each human conversation matters more.
Next step
This week, point an advanced agent at your own product and ask it to buy. Record the session, and note each place where the agent stops. Show the recording to your executive team. The gaps show up fast, and each gap tells you which move in the quarterly plan comes first.
Sources
- Neal Behrend, "Agent-led growth: The next GTM motion is already here," Insight Partners, March 13, 2026. Source for all data in this article. The 94%, 77% and two-thirds figures are outside research cited in that article. Insight discloses investments in OpenAI, Anthropic, StackBlitz and 6sense.























