Developer marketing is measurable - but not with a lead-gen dashboard. Much of its value is leading-indicator (activation, retention, sentiment) and influence (deals it shaped without a clean attribution line). The job is to measure honestly: instrument what you can, be explicit about what you can only influence, and resist forcing everything into last-touch attribution.

The funnel metrics worth instrumenting

Adapt to your product, but most developer businesses should be able to see:

  • Time-to-first-success and time-to-value - the DX metrics from developer experience. Trend them; they gate everything downstream.
  • Activation rate - share of signups/installs that reach first value. The single most important growth metric for a developer product.
  • Signup → integration → production conversion - where developers drop between trying and shipping.
  • Retention / active integrations - are they still calling the API next month? For dev tools, retained usage beats signups every time.
  • Expansion - usage and surface-area growth within accounts.
  • Adoption-phase cohorts, with the passive seat counted - for seat- or license-based tools, sort users by how far they’ve progressed (first use → agentic/advanced use → orchestration) and break out the licensed-but-unengaged segment explicitly instead of folding it into an “active” denominator. GitHub built exactly this into its Copilot admin dashboard (2026-07-22): three engaged phases plus a named “Passive” cohort. Naming shelfware gives a renewal buyer an honest number to trust - and a skeptical one a number you’ve already answered. The buyer needs that number more than the pitch admits: 88% of engineering organisations say they measure AI impact or ROI, but only 39% of those do it through any formal or automated process (SlashData, September 2026; 2,341 leadership respondents in its 31st Developer Nation wave), so the vendor that ships the adoption dashboard is usually supplying the buyer’s own reporting, not competing with it. That reporting now reaches past the vendor’s own product: since October 2026 GitHub’s Copilot usage metrics API names each MCP server, skill and custom agent an organisation configured (they were lumped under other or custom before; up to 25 entries per list with the rest pooled under other, and only from Copilot’s own, SDK-based and ACP clients - GitHub’s changelog, 2026-10-22) and counts traffic by client tool, Claude Code and OpenCode included. If you ship an MCP server, your enterprise buyer can now see from GitHub’s report whether their developers use it - under the name their admin gave it, from client telemetry GitHub itself says can undercount - so expect that number in the renewal conversation and know your own first.
  • Docs metrics - top landing pages, zero-result searches, quickstart drop-off step.

Measuring DevRel without breaking it

Don’t put a lead quota on DevRel (community and DevRel). Measure it on the layer it actually moves:

  • Reach & awareness - developers reached through talks, content, streams; qualified traffic to docs and repos.
  • Activation & enablement - workshop attendees who activate; tutorial readers who reach first success.
  • Community health - contributors, time-to-first-response, sentiment, returning participants.
  • Product feedback delivered - issues surfaced, features shaped. Real value, usually uncounted.
  • Influenced pipeline - deals where DevRel touchpoints appear in the story, reported as influence, not attribution. Sourced-lead credit misrepresents how developer trust actually converts.

Product usage is the signal nobody else can buy

The same instrumentation that measures activation doubles as go-to-market data. Third-party intent signals - website visits, keyword tracking - are commoditized; every competitor can buy the same feed. First-party product usage is proprietary by definition, and operators who run outbound on it weight it heavily: scoring models built on roughly 80% behavioral signals (usage volume, new users added, in-app engagement) to 20% firmographics, with the play list deliberately capped at a dozen or fewer so sales can actually run it. Get raw usage events into a warehouse you control before you rent anyone’s signal layer - the vendors that package these signals keep getting acquired into big GTM suites, and their roadmaps follow their new owners.

Leading vs lagging

Revenue is lagging and heavily mediated for developer products - the path from a great tutorial to a signed contract is long and multi-touch. Manage to leading indicators you can move this quarter (activation, time-to-value, community responsiveness, content that ranks) and hold them accountable to the lagging ones over quarters, not weeks.

Attribution honesty

  • Peer recommendation and organic discovery are usually your biggest channels and the hardest to attribute. A model that credits only trackable touches will systematically undervalue exactly what’s working and push spend toward what’s merely measurable.
  • Prefer a blended view: self-reported attribution (“how did you hear about us”), cohort analysis, and holdout/geo tests over a false-precision last-touch dashboard.
  • Report ranges and influence, not invented certainty. For a developer audience, an honest “we influenced this, here’s the evidence” is more defensible - and more useful - than a made-up attributed number.
  • An AI-referral count is a number the citing engine controls. The segment rests on a tag or a referrer the engine chooses to send. For about a week from 2026-10-08 most ChatGPT citation links lost the utm_source=chatgpt.com parameter: Profound’s crawl put the tagged share as low as about one answer in twenty at the worst hour, and LLM Pulse saw ChatGPT sessions across its monitored sites run about 46% below the same days of the prior two weeks over 9-11 October (both sell the tracker; neither publishes a sample). An Adobe director’s own analysis found the missing visits arriving as direct traffic from links that withheld the referrer, though LLM Pulse found that rise too small to spot in its own site-level data - which is the point: once the tag goes, nobody can say for certain where the visit went. OpenAI called it a bug, in a relayed account to one Adobe director, and the tag was back by 10-16. Pair the segment with the self-reported field, annotate the dates an engine changed how it tags, and report the two as a range rather than the tagged count as the rate.

The point of measurement is better decisions, not a prettier board. Instrument the path to value, watch it honestly, and let it tell you where the next unit of effort goes.