The plan

Each pillar carries the same four fields: current state, intervention, how I’d know it worked, and what I don’t know yet.

One exception to the badges, stated once. Every number about OpenRouter on this page carries a provenance grade. The ones about my own prior work don’t — the vocabulary grades what I could measure from outside, and none of it reaches my own record. Where those appear I’ve said whose they are and left them unbadged.

Self-serve growth engine

Current. [QUOTED] 10M users, 250k+ apps, 200T+ tokens/month, “nearly all of it starting in self-serve.” Signup→first-purchase conversion, ARPU and gross margin per routed dollar govern every target below, and I have none of them.

Intervention. Instrument end to end before buying against it: signup, key creation, first API call, first credit purchase, second purchase, 30-day retained spend. Buy against first credit purchase, never signup. A signup with no key isn’t a user.

How I’d know. Blended CAC per credit-purchasing user, payback in months of gross margin, reported against a forecast set at quarter start — not against last quarter’s actual.

What I don’t know. Whether activation is instrumented at that granularity. And if ARPU is bimodal — a long tail of small accounts against a short head of production apps — blended CAC is the wrong metric, and this section needs rewriting around segments.

Answer-engine and search

Current. What you publish for machines is advanced. The gap is that per-model usage data renders and isn’t crawlable.

Intervention, in ascending effort.

1. Expose what already renders. [MEASURED] Per-model /apps, /activity, /providers, /uptime return 200 and appear in zero sitemap entries. Days of work, not a content programme. Highest value-per-hour item in this document.
2. Harvest what the community has already written into canonical public answers. [MEASURED] The #models census alone is 1,027 threads of people comparing models on real workloads — the experiential signal a comparison template can’t carry.
3. Turn the integration directory into an outcome library. [MEASURED] 54 tools are named; no page says what any of them achieved.
4. Test the summarisation hypothesis before acting on it, on the panel below.

How I’d know. Citation share, not ranking: how often OpenRouter is named and cited by answer engines across a fixed panel of high-intent routing questions, re-run monthly against a baseline.

What I don’t know. Whether Disallow: /seo/ conceals an existing programmatic effort. And I have no baseline panel yet — building it is week-one work, not a claim I can make today.

I’d bring an SEO partner for the technical half of this pillar. My depth is the measurement and AI side; thirteen years of enterprise technical SEO is not something I’d claim.

Conversion rate optimisation

Current. [MEASURED] The anonymous model page offers Try this model, Playground, Compare — and no account or credit CTA; authenticated, Create API Key appears. Three of those are the right three: a developer who has just landed wants to run the thing, not read about it, and the page lets them. What it never does is ask for the one action that turns a run into an account. [MEASURED] openrouter.ai/signup returns HTTP 200 titled “Author Not Found” (the path collides with the /{author} namespace; the working path is /sign-up). It carries noindex, so the search impact is contained — the conversion impact isn’t.

Intervention. Standing experimentation cadence on the credit-purchase flow first, because it’s the only step where money changes hands and therefore the only step where a lift is unambiguously revenue. Start with the coupon cohort as a natural experiment: what distinguishes the people who came back from the ones who didn’t. Nothing visible from outside splits that cohort.

How I’d know. Hypothesis → test design → readout → decision, logged per test, primary metric and minimum detectable effect pre-registered. Tests that can’t be powered ship as judgment calls and get labelled as such rather than dressed up as experiments.

What I don’t know. Per-step traffic volume, which decides whether these pages can be A/B tested at all or whether the honest answer is sequential testing with geo splits.

Prior work. I reduced registration friction 35% through a data-driven intake redesign for first-generation and adult learners, and improved program retention 22% through friction analysis on an 80+ participant annual pipeline. Different building, same problem class.

Paid media, and the $5M

Current. [QUOTED] The req names a “paid budget of $5M+ for 2027” and a director-level “player/coach” to own it.

I have not run a paid-media P&L. I have owned a $500K capital budget end to end: my own studio buildout, where I set the allocation, chose the vendors and carried the variance. Inside a campus construction and relocation programme my authority was variance tracking and executive reporting, not approval over allocation. Governance, vendor discipline and executive spend reporting transfer, and the posted paid line explicitly includes [QUOTED] “sponsorships” — bought, negotiated and governed like vendors. The continuous reallocation loop does not, because a capital budget has fixed scope, and the table below is how I would learn it in public, with a kill rule written down before the first dollar.

Here is the reasoning the credential would stand in for.

What those numbers are priced against. An allocation is a judgment and stays ungraded above. The rates it is priced against are not, so here they are with their sources.

[QUOTED] EthicalAds publishes a developer-audience CPM card — “Q3 2026 Prices are per thousand impressions (CPM) $1,000 minimum ad buy” — listing AI / machine learning at $4.40 run-of-network, $7.25 US, $5.75 blended. AdvertiseCast publishes host-read podcast benchmarks of “60 second spots: $24-26 CPM” and “30 second spots: $18-22 CPM”. Blip prices self-serve digital boards at “$0.01 per 8-second play”, with no minimum spend. All three fetched 2026-08-27.

[INFERRED] So the programmatic line buys 55 to 91M impressions depending on geo-targeting, the OOH experiment is on the order of 40M eight-second plays, and the podcast and newsletter line is bounded above at 35 to 50M impressions if every dollar of it went to podcasts, which it will not. These are order-of-magnitude checks on the allocation, not a media plan, and I have not blended them into a single cost-per-developer: an impression, a host-read spot and a billboard play are not the same event.

Five of the eight lines have no published rate. They are priced at the negotiating table, which is a fact about this market rather than a gap in the plan.

Clipping is the cheapest line on this list, and the one I would start first.

[MEASURED] Klap publishes clip-generation tiers at $14, $39 and $94 per month for 100, 300 and 1,000 clips a month; OpusClip’s starter tier is $15 per month; Memelord lists a solo tier at $19.99 per month and an agent tier at $199 per month. All fetched 2026-08-27.

The meme agent takes assignments and returns creative rather than posting as a person; the constraint on meme-shaped work is brand judgment, not budget.

Governance I’d set on day one.

• Buy against first credit purchase, never signup.
• Every line gets a CAC ceiling before it gets a dollar. The ceiling needs ARPU and gross margin, which I don’t have — so the rule is stated here, not the number.
• Kill rule, pre-committed. Two quarters and $100K minimum spend to prove out. Above $100K cumulative, CAC persistently >1.5× blended target ⇒ cut, reallocate to reserve. Written down before launch so it isn’t relitigated when the failing channel is someone’s favourite.
• Incrementality. Podcast/creator and OOH are not click-attributable and must not be judged on last touch. Geo holdout on both, designed before spend, matched controls, readout date fixed in advance.
• Reserve stays unallocated through Q1 on purpose. A plan with no slack can’t respond to its own results.

[INFERRED] Every number in that table is a starting hypothesis with a kill rule attached, not a forecast. The forecast is the first quarter’s work.

Measurement and attribution

Intervention. Sequence matters: event schema first, then attribution, then incrementality. Attribution built on an inconsistent schema produces confident wrong answers, which are worse than no answer because they get acted on.

How I’d know. One number for “customers acquired this month,” agreed across marketing, product and finance. The stack is working when nobody keeps a private spreadsheet because they don’t trust the dashboard.

What I don’t know. Whether the warehouse can join marketing touch to usage revenue at account level. If it can’t, that’s the actual first project, and it’s an engineering dependency with a different owner.

Prior work. Built automated reporting infrastructure that removed 40+ hours of manual aggregation per month, and executive dashboard reporting on milestones, budget variance and operational continuity.

Developer relations

Current. Fast first response on one channel, structural silence on the other — the two failure modes documented in the evidence section.

Intervention.

1. Per-channel response measurement, right-censored. A two-hour-old thread with no reply is new, not abandoned, and any metric that conflates them is wrong in the flattering direction.
2. Post-incentive retention as a named metric. Same coupon spend, cohort tracked to 30-day retained usage.
3. Harvest to publish — route recurring answers into public canonical pages.
4. Staff the moment, not the notice. Treat #events and lu.ma/openrouter as one calendar with a bridge.

How I’d know. Aged no-reply rate per channel trending down. Coupon-cohort 30-day retention trending up. Second-event attendance by first-event attendees — the only event metric that distinguishes a programme from a party.

What I don’t know. Whether the post-launch no-reply rate on the product-feedback forums is capacity or deliberate triage. Those call for opposite interventions, and I’d want the team’s read before assuming the first. Nothing here is a judgment about anyone’s discipline — the blind spot is in the measurement, not in the answering.