Agent-led growth (ALG) is a GTM motion where AI agents do growth work inside human-set guardrails. On the demand side, buyer agents research the company, judge the product and take a next step, often before the first call. Agent-ready means findable, evaluable and actionable.
- Findable: the company is named on buyer questions from the message map, with the mapped claim.
- Evaluable: documentation, visible price, integration examples, and the answers that usually wait for a demo, in a form the agent can read.
- Actionable: a trial, a documented API, a booking, or a request completes as a usable record.
This page checks the demand side: a buyer’s agent on public pages, docs and tools. Company-run agents for drafting, routing and support are useful. That is not what this page checks.
ALG is not AEO
Answer engine optimisation (AEO) is the work of getting named in an AI answer. It is also called generative engine optimisation (GEO). That is supply-side work. It sits inside findable. It does not make the company evaluable. It does not make the next step actionable.
The company optimising for AEO/GEO does not control the buyer’s agent. It does not control the last step.
Neal Behrend, writing for Insight Partners on 13 March 2026, names three properties: findable, evaluable and actionable.1 That framework is the contract I hold the public site to. But Insight’s piece focuses on what companies can do, the supply side. This guide checks what a buyer’s agent can verify on public pages, the demand side.
I keep a fixed list of buyer questions from the message map. On each question the company lands once: recommended, cited, mentioned, or absent. A mention is not a recommendation. A citation I cannot open is not proof.
The page has to carry the sentence I want quoted, in the first short block. That sentence needs a source a person can check. If the page leaves out the price, the limit, or who it is not for, the agent fills the gap from another page. That is how a rival gets named on the question.
Findable, evaluable, actionable
These three states run in order. An agent that cannot find the company cannot judge it. An agent that cannot judge the product cannot complete a next step worth having.
Findable
Someone has a problem. The agent is searching on their behalf. The company has to come up in that answer.
Findable means the company is named on buyer questions from the message map, with the mapped claim. Not a vague positioning statement. The specific claim that answers the buyer’s question.
Hugo Santana, writing on 13 August 2026, puts the citation after the product, the distribution and the pricing.2 A vague claim stays vague even when it is repeated. On a miss I write down what happened. The company may be absent. The answer may name a competitor instead. The mapped claim may be missing, or it may be too vague to use.
Evaluable
Judgement has to come from what the agent can read.
Evaluable means documentation, a visible price, an integration example, and the answers that usually wait for a demo exist in a form the agent can read. Behrend calls the spend token-to-value: how many tokens go into deciding the product fits, and how many more go into putting the product to use.1 That is time-to-value counted in tokens.
A hidden price raises that cost. Proof that only lives on a slide does the same. Once the cost is high enough, the agent picks a product that is cheaper to judge.
hcd.ai, on 1 June 2026, says a quoted line has to match the source before a generated page is safe to ship.3 They point to a Columbia Journalism Review test of eight AI search engines, in which more than half of the responses had incorrect attributions.
I have not re-run that test. If the rate is even roughly right, the summary can still be someone else’s wording.
Actionable
The next step has to land as a usable record.
Actionable means a trial, a documented API, a booking, or a request completes as a usable record. Not a form post that sits unseen. Not a booking link the agent cannot follow. A record a person can use.
Behrend points at trials and usage-based pricing, because an agent cannot get budget approval. He reports that Google and Microsoft published WebMCP on 12 February 2026, in Chrome 146, so a form can be labelled and used as a tool.1
It is too early to treat that as a channel. A form with no stable meaning should stay off a machine.
Evidence table
| Part of the check | What I write down |
|---|---|
| Findable | Which buyer questions from the message map the agent is searching with, and whether the company is named with the mapped claim. |
| Evaluable | Whether documentation, a visible price, an integration example and the answers that usually wait for a demo exist in a form the agent can read. |
| Actionable | Which next step the agent may complete (a trial, a documented API, a booking, or a request that arrives intact), and what proves it landed. |
| Proof a person can check | The same claim in a form a machine can cite, with a source a person can open. |
| Failure handling | Call-only claims, an action the agent cannot finish, a quote that does not match the source, and who reviews those misses. |
| What I would not treat as done | A citation count, a successful form post nobody uses, or a generated interface sold as the fix. |
I write those rows for the next evaluator: the buyer’s agent, then the person who has to use the result.
Who evaluates first
Under sales-led growth, a person still does the early judging. The CRM is how that relationship runs at scale. The seller is the one who compares, qualifies and books the next conversation.
Under product-led growth, a lot of that work moved into the product. A person could reach a result without anyone booking the first meeting.
Behrend maps the tools the same way.1 Sales-led work runs on the CRM. Product-led work runs on product analytics. He maps account-based work to intent data. After that come agents that can read pages and call tools. Sometimes they start a trial or a booking.
| Sales-led | Product-led | Agent-led, demand side | |
|---|---|---|---|
| Who evaluates first | A seller, in a conversation and a demo | The user, inside the product | A buyer agent, from pages, docs and tools |
| What has to be true | A person can get a meeting | A person can reach a result alone | An agent can find the product, judge it and complete a trial, a documented API, a booking, or a request |
| Where the proof lives | The call, the case study, the security review | The product, and the time to a result | The same proof, in a form a machine can cite and a person can check |
| What breaks it | A handoff the seller never receives | A product that needs a person to explain it | A claim that only exists on a call, or an action the agent cannot complete |
| What I would not claim | That the seller is gone | That product-led growth already covers this | That a B2B category has already picked its defaults |
The cases Behrend cites are developer tools. Supabase grew from 1 million to 4.5 million developers in under 12 months. Its chief executive attributes a doubled sign-up rate over three months to Bolt, Lovable and Cursor. Resend, launched in 2023, is reported at 400,000 users. In one test, Claude Code chose Resend for email 63% of the time and SendGrid 7%. Santana adds a June 2026 figure: Supabase said more than 60% of new databases on the platform were created by AI coding tools.2
Behrend puts his own limit next to those numbers. The developer ecosystem tipped in under 18 months. Category-specific B2B software has not picked its defaults.1
A developer can often accept an agent’s choice. A free tier takes the budget conversation away. That is a developer-tool sample. I would not treat it as mid-market SaaS, or turn it into a forecast.
I would still run the public-page check: what can the agent prove without a person, and where does it have to wait.
A company can run more than one of these motions. A sales-led product can still be agent-ready. The claims have to sit on pages a crawler can read. So does the proof, and the path to a person. They have to be specific.
What ALG is not
Agent-led growth is not agentic marketing. Salesforce and Adobe use that term to describe company-run agents that draft content, route leads, or handle support. Those are useful. They are supply-side automation. This page checks the buyer’s agent, a demand-side readiness.
Agent-led growth is not a generative-UI rebuild. I track generative UI.4 I do not sell a generated website as the fix. An agent may still answer without the customer loading the designed page. I want the claims, the proof and the next step on the version the site means an agent to read.3
Agent-led growth is not a citation count. A citation I cannot open is not proof. A mention is not a recommendation. Findable is one state. Evaluable and actionable come after, or the agent stops.
Agent-led growth is not only company-side drafting agents. Company-run agents for drafting, routing and support are useful. That is not what this page checks. The check on this page is the buyer’s agent on public pages, docs and tools.
Where this sits in the engine
I do not run this as a separate programme. An agent may now be the one reading the same five parts.
The voice-of-customer positioning guide is where interviews, reviews and calls become a message map: trigger, alternative, outcome, objection, proof. The agent will repeat the page. The language on that page is what gets passed on.
The demand generation guide is where a campaign is told what a conversion is. An agent that starts trials is a cheap start. The stage in that instruction still has to be one the business believes.
The lead scoring and lifecycle guide is where fit stays apart from engagement. Traffic on the docs can be one developer and one agent. A buying group is bigger than that.
Some of these choices never appear as a referral. The person types the URL. I use the analytics and attribution guide to see what the dashboard is actually counting.
The page, the form and the CRM event still need one name for one action. That is marketing engineering. I still want a source and a date on a comparison. I want an owner on a trial. I want a way to tell a synthetic run from a buyer. A saved submission still has to become a record somebody can use. The handoff guide is that check.
What I measure
Santana’s long-term number is revenue from purchases an agent initiated. A person still signs most of the contracts I work on. I cannot tie most of them to an agent, so I do not report a precise “agentic revenue” figure.
Four checks fit on one page.
- Findable. Keep a fixed list of buyer questions from the message map, and check it on a schedule. Note whether the company is named. Check the mapped claim too. See if a competitor is named instead.
- Evaluable. The agent should return the price, who the product is for, what it does not do, and what proof exists. None of that should be invented. A wrong price is a defect.
- Actionable. The action has to show up as a record a person can use. A form sitting unseen has not landed.
- Human override. Count how often a person corrects a comparison or switches a path off. If that rate rises, the library or the limits are wrong.
Pipeline can still fall while the product gets easier for an agent to choose. A competitor may already be the familiar default.
The page shows readiness. That does not, by itself, promise pipeline.
Who sets the limits
A person writes the intent: the segment, which claims can be made, which cannot, which actions an agent may take, and which always come back to a person.
Price stays on that last list. So does a legal commitment, a security promise, and anything that spends money. Until someone moves them on purpose.
The agent works inside those limits. On the demand side it usually belongs to the buyer. The company controls the readable material and the actions its own site offers. The agent stays the buyer’s.
If the library has a blank, I leave the claim out.
A person checks on a schedule, against the source. That person can switch the path off. Any sentence a machine might quote should have a source the company will stand behind. A customer quote is fine. A document is fine. So is a dated figure.
A sheet and a named owner are enough to begin. I would still open the source. Skip that check, and a wrong sentence gets repeated. A model can treat it as the right one.
Behrend says marketers and sellers still have a job. A buyer who still takes a meeting has often already formed a preference. They have less patience for an evaluation that starts again from nothing.
Where I start
I would not start by buying an agent platform.
The first checks use pages, claims and records that already exist.
- Open the live pages against the calls. Read the homepage and the last handful of sales-call notes. By day 30, write the claims down. Keep the wording, the source, the date and the segment. Ask whether a competitor could honestly paste the sentence onto their own site. Drop anything without a source. Publish what remains where a crawler can read it in the first response. Beside the library, list the questions from real sales calls that the site still leaves unanswered. The buyer spent the call and still has no decision.
- Compare the three surfaces that already exist. Take the answer an agent gives now, the page a person lands on, and the record a seller already receives. By day 60 those three should say the same thing. Put up a short machine-readable index. Write a product description that includes the price, the limits and the integrations. Make the form fields match the CRM. Run a handful of real questions by hand. If the agent invents an answer, fix the source. If it names a competitor, fix the source.
- Follow one live request until a person can use it. Pick a trial, a booking, or a documented integration that already exists. By day 90, one of those is live and a person reviews it. Write down who reads what arrived. Note how an agent-made request is labelled, and what gets switched off if the quality is wrong.
Then decide whether to continue, narrow the claims, or stop.
If a company only sells through a named buying group, and nobody maintains the library, do not pretend a file will run itself.
Questions
What is agent-led growth?
Agent-led growth is a GTM motion where AI agents do growth work inside human-set guardrails. On the demand side, buyer agents research the company, judge the product and take a next step, often before the first call.
Agent-ready means findable, evaluable and actionable. Findable: the company is named on buyer questions from the message map, with the mapped claim. Evaluable: documentation, visible price, integration examples, and the answers that usually wait for a demo, in a form the agent can read. Actionable: a trial, a documented API, a booking, or a request completes as a usable record.
Company-run agents for drafting, routing and support are useful. That is not what this page checks. This page checks the buyer’s agent on public pages, docs and tools.
How is agent-led growth different from product-led growth?
Product-led growth asks whether a person can reach a result without a salesperson.
Agent-led growth asks whether a buyer’s agent can find the product, judge it before that first meeting, and complete a trial, a documented API, a booking, or a request. Token-to-value is the tokens spent before the agent can decide and act.
Sales-led growth is where a person still does the early evaluation. A company can run more than one of these. A sales-led product can still be agent-ready when the claims, the proof and the path to a person are public and precise.
How do you measure agent-led growth?
I keep a fixed list of buyer questions. I check whether the mapped claim appears. I check whether price, fit, limits and proof come back with nothing invented. I check whether the next step becomes a usable record. I also count how often a person overrides the result.
Santana’s longer-term measure is revenue an agent initiated. A citation count does not tell me that. Use the revenue figure once you can tie it to an agent.
Are buyer agents just chatbots?
A chatbot is the support widget on your site. A copilot helps someone draft. The person still sends the email, books the meeting, or completes the purchase.
A buyer agent can compare vendors. Where it has the authority, it can start a trial or write a comparison. The person then does less of that work. If the accurate description only lives on the call, the agent uses whatever else it can find.
What is agent-ready GTM?
Agent-ready GTM is the work of making a company findable, evaluable and actionable for buyer agents. Findable: the company is named on buyer questions with the mapped claim. Evaluable: price, fit, limits and proof in a form the agent can read. Actionable: a trial, a documented API, a booking, or a request that completes as a usable record.
The Agent-Ready GTM Start kit is the fixed project. It delivers a claim and proof library, answer surfaces a machine can read, offer modules for a trial, an entry and an upgrade, and a one-page readiness score. Your team implements it.
The fixed project
This page explains the framework. The fixed project is the Agent-Ready GTM Start kit. The kit has a claim and proof library, and answer surfaces a machine can read. The modules cover a trial, an entry and an upgrade. The score is one page, across findable, evaluable and actionable. Your team implements it.
I do not launch an agent inside a buyer’s tools as part of that kit. I do not rebuild the site as a generative interface. If you are unsure whether it fits, Start a project is the front door.
I am engaged with client work just now, so none of this is a pitch. If you are building growth-stage B2B SaaS and want this kind of work when capacity opens, tell me what you are building. I read every note.
Sources
-
Neal Behrend, “Agent-led growth: The next GTM motion is already here,” Insight Partners, 13 March 2026. insightpartners.com (opens in new tab) ↩ ↩2 ↩3 ↩4 ↩5
-
Hugo Santana, “The Agent-led Growth Framework,” 13 August 2026. agentledco.substack.com (opens in new tab) ↩ ↩2
-
hcd.ai, “Get Ready for the GenUI Web,” 1 June 2026. The Columbia Journalism Review attribution finding is reported there, not re-tested in this guide. The same piece notes that an agent may answer without the customer loading the designed page, and that a file such as
llms.txtcan point it at the version the site intends. hcd.ai (opens in new tab) ↩ ↩2 -
Tamara Martinez, “Generative UI (GenUI) Strategy,” Capicua, posted 4 August 2026, updated 5 August 2026. She describes an interface a model assembles at runtime from a user’s intent, and reports Gartner’s expectation that by 2028 a third of user experiences move from native applications to agentic front ends. The Gartner expectation is as Capicua reports it. capicua.com (opens in new tab) ↩