Product-Led Growth Meets Agents: Your Activation Metric Now Measures Machines

Agents now complete signup, onboarding and evaluation inside products that were instrumented to measure humans. Cloudflare reported in July 2026 that more than 50% of internet traffic is non-human. Product-led growth metrics were defined against one human actor per account, and that single assumption is now the weakest joint in the entire growth stack.

What "user" used to mean in product-led growth

The short answer. In product-led growth the user is now better defined as the principal, meaning the human or the organisation that an action was taken on behalf of. An agent is an actor, not a principal. Activation, product qualified leads and expansion should each be counted once per principal, because the principal is the party that decides to pay.

Every metric in the standard product-led growth stack rests on three assumptions that nobody wrote down, because until recently they were free. One account maps to one person. That person's behaviour reveals their intent. And the effort a task costs them is roughly proportional to how much they wanted it done.

Activation works because a human choosing to complete four setup steps has revealed something expensive about their intent. Time to value works because the clock is measuring a person learning. Product qualified lead scoring works because event volume is a costly signal when a human has to generate each event.

An agent invalidates all three at once. It maps many principals to one actor or one principal to many actors. Its behaviour reveals the intent of whoever wrote its instructions, which you cannot see. And its effort is close to free, so volume stops carrying information about desire.

This is not a new category of problem. It is the same problem that email open rates hit when mail clients began prefetching images. The metric kept reporting a number, the number kept moving, and it quietly stopped meaning what the dashboard label said it meant.

Your analytics already made this decision for you

Before you redesign anything, check what your measurement layer is already doing on your behalf. Google Analytics excludes traffic from known bots and spiders automatically, using a combination of Google research and the International Spiders and Bots List maintained by the Interactive Advertising Bureau.

The part that matters for growth reporting is the sentence immediately after. Google's own documentation states that you cannot disable known bot traffic exclusion or see how much known bot traffic was excluded. That is documented in Google's bot filtering page.

So your funnel already has a hole of unknown size, applied by a list you do not maintain, and you cannot audit either side of it. Agents that declare themselves get dropped silently. Agents that present a normal browser signature get counted as people. Both errors land inside the same activation percentage.

The composition of that machine traffic is also shifting under you, which means last year's calibration does not hold. Cloudflare found that AI training requests made up 22% of crawler activity in spring 2025, and 52% by June 2026. Pure search crawling is now a small and declining slice.

The practical consequence is narrow and worth acting on. Any activation number computed from client-side page events is now measuring a population your vendor defines. Move the definition to server-side events keyed on an account identifier, where you control the filter and can report what you dropped.

Four product-led growth metrics that break when the actor is an agent

These break in different directions, which is the worst available outcome. A metric that is uniformly wrong gets caught in a quarterly review. A set of metrics that are wrong in opposing directions produces a stable aggregate and a false sense that nothing has changed.

Activation inflates, and it inflates fastest at the top

Activation is usually defined as completing a set of key actions within a window. An agent completes those actions reliably, quickly, and without any of the hesitation that made the metric informative in the first place.

Your activation rate rises. Your week-four retention does not follow it, because the human on the account never formed a habit and in many cases never saw the interface. The correlation between activation and revenue, which is the only reason anyone tracks activation, weakens exactly when the number looks best.

Product qualified leads misroute your most expensive resource

Product qualified lead scoring weights usage intensity, because for a human, intensity is costly and therefore meaningful. An agent generating 400 API calls in an afternoon is not expressing 400 units of purchase intent. It is executing a loop.

This one costs real money rather than reporting accuracy. Sales capacity is finite, and a scoring model that ranks agent-driven accounts at the top routes your best people toward the accounts least likely to close. I would treat any unexplained jump in PQL volume during 2026 as a measurement fault until proven otherwise.

Time to value improves for reasons that predict worse outcomes

Agents compress setup. Integration steps that took a customer two weeks now complete in an afternoon, and the time-to-value chart improves without anyone changing the product.

What lengthens instead is the interval before a human understands what was built. That interval is invisible in every dashboard I have seen, and it is the one that predicts whether the account survives a renewal conversation. The same displacement is happening to the interface layer, which is the argument in the case that the dashboard is dead.

Expansion and net revenue retention lose their referent

Per-seat expansion assumes that more work means more people. When the additional work moves to agents running under one seat, revenue stops tracking value delivered, and the customer notices that arithmetic before you do. That mechanism is worked through in the autopsy of per-seat pricing.

Consumption pricing fails differently and less obviously. Agent retries, failed tool calls and redundant polling all bill as usage, so waste arrives on your revenue line looking exactly like expansion. It reverses at renewal, when the customer's own finance team runs the same query.

What each metric assumed, and what replaces the assumption

Metric	Hidden assumption	Why the agent breaks it
Activation rate	Completing setup is costly, so completion reveals intent	Setup is now close to free, so completion reveals capability instead
Product qualified lead score	Event volume is proportional to desire	Volume is proportional to loop count and retry policy
Time to value	The clock measures a person learning the product	The clock now measures an agent executing a plan
Net revenue retention	More value delivered means more seats or more considered usage	Seats flatten, and retries bill identically to real work
Free tier cost to serve	Signup volume is bounded by human patience	Signup volume is bounded by whatever the calling script decides

The last row is the one finance notices first. Free tier compute economics under machine-scale signup volume are covered separately in the piece on metering the free tier.

The traceable evidence says the surge is mostly humans

Here is where the popular version of this story and the checkable version part company. The gap is large enough to change what you should do on Monday.

The loud claim in growth writing this year is that agents are becoming your dominant user base. The best-sourced figure I could trace points the other way. It comes from a16z's published summary of its 30 January 2026 episode with Netlify chief executive Matt Biilmann. Netlify was running 16,000 daily signups, five times the prior year, with 96% of them not coming from AI coding tools.

Those signups were people who had built something in a chat window and then needed somewhere to put it. The same summary notes that 25% of users immediately paste error messages into a language model rather than debugging by hand. That is a human user with a changed working method, not a machine user.

That reframing changes the prescription completely. If most of your new volume is inexperienced humans delivered by an assistant, building a separate agent funnel is premature, and rewriting your onboarding for people with no prior product knowledge is urgent. The same distinction drives how buyers now assemble a shortlist, which is examined in the piece on the AI-assembled B2B shortlist.

Netlify itself is explicit that the machine population is coming. Its own blog says autonomous agents are "going to be a whole new user base, and quite possibly their most important user base", and the company has published an agent-facing entry point to serve them. A future user base and a current one call for different budgets.

A replacement metric set you can ship this quarter

The goal is not a new dashboard. It is to restore the property that made the old metrics useful, which was that each one counted a decision taken by a party who could pay you.

Two definitions to change

Redefine activation as principal activation. Count an account as activated when a human principal returns and takes one decision-grade action within 14 days. A decision-grade action is one that carries a cost to reverse, such as inviting a colleague, connecting production data, setting a spend limit or changing a plan.

Then redefine qualification by capping agent-originated events at one per session of record. Score on distinct principal actions rather than raw event volume. Both changes are query-level work against data you already collect, assuming your events carry an actor identifier.

Two numbers to add

Add agent share of served work, reported next to every funnel metric so that no aggregate is ever read without its mix. A 62% activation rate means two different things at 3% agent share and at 40%, and the blended figure hides which one you are looking at.

Add cost to serve per unverified account, because free tier compute is now your exposure rather than a rounding error. This is the number that turns an abstract measurement debate into a line item your finance team will defend.

The replacement set, and what each one needs from your data model

Replace this	With this	What you need first
Activation rate	Principal activation: a human principal takes one decision-grade action within 14 days	An actor identifier on every event, and a written list of decision-grade actions
Product qualified lead score	Principal-weighted qualification, capping agent events at one per session of record	Session-of-record grouping in your event pipeline
Time to value	Time to first human confirmation after the agent's work exists	A timestamp on the first authenticated human action following setup
Net revenue retention	Work-unit retention: units of the job actually done, per account, ignoring seats	One agreed work unit per product, such as deploys, resolutions or reviews
Nothing, this is new	Agent share of served work, published beside every funnel metric	The identity signals in the next section
Nothing, this is new	Cost to serve per unverified account	Per-account compute attribution on the free tier

Work-unit retention is the row that takes longest, because the argument about what your work unit is will outlast the engineering. Have that argument once, in writing, and do not revisit it quarterly.

Telling agents from humans without waiting for a standard

All of the above depends on attributing an action to a principal, and the honest position is that the infrastructure for doing so is unfinished. Web Bot Auth is the leading attempt, and it lets agents sign requests cryptographically so a site can verify who is calling.

It builds on HTTP Message Signatures, published as RFC 9421. Operators publish a public key at a well-known path and sign outbound requests, which Cloudflare describes in its work on cryptographically recognising agent traffic. Cloudflare, AWS, Akamai and Vercel already verify these signatures in production.

It is also, as of August 2026, an individual draft rather than an adopted standard. Anyone who tells you agent identity is solved has not read the working group status. The broader governance problem this creates is covered in the piece on agent identity and non-human access.

So instrument three signals and accept that none of them is complete. First, presence of a valid signature, which gives you a small population of honestly declared agents. Second, session shape, because inter-event timing variance separates a person from a loop with useful reliability. Third, and most important, a declared principal.

The third is the only signal you fully control. Add an optional "acting on behalf of" field to your API tokens and your programmatic signup path, then give accounts that populate it something worth having, such as higher rate limits or a cheaper tier. You are not detecting agents at that point. You are making honesty the path of least resistance, which works better and costs one sprint. The same design question shows up in where value accrues in agent marketplaces.

Where this argument is weakest

Three places, and the third is the one most likely to apply to you.

There is no benchmark set for any of this

Every replacement metric proposed above is a definition, not a measured norm. I found no published benchmark for agent-adjusted activation, principal-weighted qualification or work-unit retention across a real panel of companies. The circulating conversion and activation benchmarks for 2026 trace to vendor blogs and aggregator roundups, and several restate each other without a common underlying dataset.

Treat any figure presented as an agent activation benchmark this year with suspicion, including by asking who ran the study, over what sample and in which window. That test disqualifies most of what is currently in circulation.

The evidence base here is traffic data, not funnel data

Cloudflare measures HTTP requests across its network. It does not measure signups, activation or conversion in your product, and the two populations are not interchangeable. A web that is majority non-human tells you nothing precise about what share of your trial starts are machine-driven.

The Netlify figures are one company, in one category, sourced from an episode summary rather than an audited disclosure. They are the best-sourced counter-evidence I could find, and they remain a single observation. Related traffic composition effects on the demand side are examined in the analysis of AI referral traffic and conversion.

The strongest case is often to do nothing yet

If agents account for 2% of your served work, rewriting metric definitions costs more in engineering and organisational confusion than the accuracy is worth. Measure the mix first, publish it, and set a trigger threshold before you touch a definition.

I would set that trigger at 5% of served work, reviewed monthly. That number is judgement rather than measurement, and I am stating it so that it can be argued with rather than adopted quietly. The value of a pre-committed threshold is that it stops the definitions moving every time the chart does.

Frequently asked questions

Where to start this week

Two things, and the first one takes an afternoon.

Pull your activation and product qualified lead numbers for the last four quarters, then split each one by whether the account has any programmatic authentication. You do not need agent detection for this, only a token type. If the two populations have diverged, you now have the size of your problem rather than an opinion about it.

Then write down your decision-grade action list and circulate it. Four to six actions, each costly to reverse, each already logged server-side. That list is the entire prerequisite for principal activation, and disagreement about what belongs on it is the most useful argument your growth team will have this quarter.

Related reading