AI-era DevRel metrics | DevRel Almanac

Page views, signups, activation, retention, and revenue remain relevant, but agent-mediated discovery and integration can begin before conventional analytics record a visit. Teams supporting that route may also need to measure assistant referrals and product descriptions, MCP adoption and errors, agent-assisted activation, and the relationship between those measures and later product behaviour.

The three measurement surfaces

A useful mental model:

Human-funnel metrics. The classical DevRel measurement stack: traffic, signups, activation, retention, community engagement, revenue influence.
Agent-surface metrics. How AI assistants represent your product, MCP server adoption, tool-call success rate, and agent-mediated activation.
Cross-surface metrics. How the two interact: what fraction of new signups arrived via AI mediation, how AI presence drives or fails to drive downstream activation.

Most DevRel teams in 2026 are still learning to instrument the agent-surface layer. The human-funnel layer is mature; the cross-surface layer is still the unsettled part.

Human-funnel metrics that still matter (but now mean different things)

Web traffic

Direct, organic, and referral traffic still matter. But “organic search” now includes traffic from AI assistants where they cite a link. Some segments:

Classical organic. Google-search-mediated traffic to your docs and blog.
AI-assistant-mediated. Visitors who arrived from a ChatGPT or Perplexity citation. Increasingly tracked separately. Many sources tag their click-throughs (?utm_source=chatgpt.com and similar appear in 2025 to 2026 referrer data).
Unattributed AI influence. Visitors may learn about the product from an assistant and then type the URL directly; conventional referral data cannot identify them.

Traffic data with AI-source tagging is starting to be widely available; most analytics platforms (GA4, Plausible, Fathom, PostHog) now break this out.

Signups

Still useful. But the source attribution has shifted. One common pattern from 2025 to 2026 is an onboarding question such as “How did you hear about us?” with explicit AI-assistant options.

Developers may not remember which source they encountered first. Reports from AI-first DevRel teams place the share of new signups citing AI mediation between 10% and 30% or more, depending on category; compare the question and sampling method before using those figures.

Activation rate

Segment activation by whether an agent participated in the integration:

Agent-mediated activation. Some developers activate having barely touched your docs because Cursor or Claude Code wrote the integration. Their path to activation is different from the classical “read docs, write code” path. Track them separately.
TTFHW per source. Activation rate from AI-mediated signups often differs from activation rate from search or referral. The AI-mediated developers sometimes activate faster (the agent did the work) or slower (the agent got confused and abandoned).

Retention

Track retention by activation source. This can test whether agent-generated integrations differ in production reliability rather than assuming that they do.

Agent-surface metrics

For a defined set of category-relevant prompts (e.g., “What are the best vector databases?”), what percentage of AI-assistant responses mention your product?

How to measure:

Choose 20 to 50 high-value prompts.
Run them across ChatGPT, Claude, Perplexity, Gemini monthly.
Track: mention rate, position (first / mid / late), sentiment, accuracy.
Stratify by assistant: your visibility on Perplexity may differ wildly from your visibility on Gemini.

Tools that automate this:

Profound (tryprofound.com), AthenaHQ, Otterly.AI, Peec AI, Sight AI, Pure Visibility.
Methodology and accuracy vary; use them for relative tracking rather than absolute claims.

Citation accuracy

When an AI assistant describes your product, how often is the description correct?

How to measure:

For each cited mention, classify: fully accurate, partially accurate, inaccurate, hallucinated.
Track over time and across assistants.
Use repeated mistakes to prioritise corrections in documentation and llms.txt.

MCP server adoption

If you publish an MCP server (see MCP for DevRel):

Active connected clients per day / week.
Tool-call volume by tool.
Tool-call success rate (was the agent able to use the tool without error?).
Tool-call error rate by type.
Average latency per tool.
Top-failing tools (these are the candidates for redesign).

Agent task completion rate

Inferred from server-side telemetry. When an agent invokes a sequence of your tools to complete a workflow, what fraction of sequences end in a successful business outcome (the message gets sent, the resource gets created, the query returns useful data)?

Hard to measure perfectly without telemetry from the agent client, but inferable from sequences of related tool calls.

Documentation completeness for agents

A health metric for the agent-readable surface:

Percentage of API endpoints documented in OpenAPI spec.
Percentage of API endpoints exposed via MCP tools.
llms.txt freshness (does it match current docs structure?).
Code-sample coverage (every important API has at least one runnable Python and one runnable TypeScript example).
Doc-update lag (how stale are pages compared to the API they describe?).

AI-assistant sentiment

When AI assistants describe your product, what is the sentiment of the description? “Stripe is the developer-friendly payments platform” is positive; “Twilio has historically had developer-friendly APIs but recent pricing changes have frustrated some developers” is mixed.

Tools that score this exist but are early. Manual coding of a sample is often the most reliable approach.

Cross-surface metrics

Attribution complexity

A developer in 2026 might:

Hear about your product from a podcast (channel: human).
Ask Claude about it (channel: AI-mediated).
Visit your docs (channel: classical).
Have Cursor write the integration (channel: agent-execution).
Sign up.
Activate.

Classical attribution models (first-touch, last-touch) don’t capture this well. Some emerging approaches:

Multi-source attribution with explicit AI-mediated touchpoint. Onboarding questions that ask developers about all the surfaces they encountered.
Cohort analysis by stated AI use. “Among developers who said they consult ChatGPT regularly, our activation rate is X. Among those who don’t, it’s Y.”
Influence-attribution panels. Survey-based research into the path-to-purchase across surfaces.

Presence in independent sources

Possible inputs to AI visibility include:

Number of mentions on Reddit / Hacker News / dev.to / Hashnode / authoritative newsletters / podcasts per quarter.
Top-quartile YouTube tutorials mentioning your product.
Sample-app repositories from third parties on GitHub.
Stack Overflow archive presence (still cited even as the platform declines).

Retrieval systems may use these sources, but their relationship with later citation rates needs to be measured for the product and prompt set.

Relationship to product outcomes

Are improvements in agent-surface metrics translating to improvements in human-funnel outcomes?

Specifically: does improving Share of Voice in AI assistants correlate with growth in signups? Does improving MCP server quality correlate with retention? These are open questions that DevRel teams in 2026 are still working out empirically.

Metrics that mean less than they used to

A few things to deprioritise:

Raw blog page views. They omit answers produced from the material without a referral and do not establish product use.
Documentation page views. Retrieval and agent use may occur without a human page view, while a page view alone does not establish task completion.
Social media follower counts. Continued vanity-metric concerns; less correlated with AI-mediated discovery.
Stack Overflow tag-watching activity. As the platform declines, this is a noisy signal.

The sentiment-telemetry gap

Research from Faros AI, JetBrains, and Anthropic reports cases where developer sentiment about AI tools diverges from behavioural telemetry such as code churn, revision, or task performance.

Surveys therefore cannot establish productivity or adoption on their own. Combine them with usage, retention, support-volume, and error-rate data where those measures are available.

The sources differ in method and scope. Taken together, they justify reporting sentiment and behavioural measures separately rather than treating one as a proxy for the other.

A practical scorecard

A representative 2026 DevRel monthly review might include:

Metric	Type	Direction or target
New signups	Human funnel	Growth
TTFHW (median)	Human funnel	Decline
Activation rate, by source	Human funnel + cross-surface	Growth; convergence across sources
90-day retention, by source	Human funnel + cross-surface	Growth; convergence
AI-mediated signup share	Cross-surface	Growth
Share of Model (top-20 prompts)	Agent surface	Growth
Citation accuracy (sampled)	Agent surface	High
MCP server tool-call success rate	Agent surface	High
MCP server adoption (active clients)	Agent surface	Growth
Documentation completeness score	Agent surface	High
Authority-domain mentions	Cross-surface	Growth
Community NPS	Human funnel	Steady or growth
DevRel-influenced ARR	Human funnel	Growth

No team in 2026 perfectly tracks all of these. Instrument what you can and state clearly what remains unmeasured.

What’s still missing from the field

Two measurement gaps remain as of mid-2026:

No standard for cross-assistant AI visibility. Profound, AthenaHQ, Otterly.AI, and others all report different numbers using different methodologies. The field needs a stable benchmark.
No established model for agent-mediated business impact. Available data cannot separate the contribution of product documentation, the coding agent, and the developer’s own decisions to a completed integration.

These gaps may close as the measurement methods mature. For now, teams should report what their instruments can show and label the remaining attribution as unresolved.