Dave AndersonField note · Marketing operations · 2026
The marketers ship the code. All five of them.
Not a metaphor for moving fast. Every person on this marketing team has GitHub and Vercel and can open a pull request against the production website — two of us have merged to it so far, and the one who has done it most is not the CMO. No ticket, no queue, no waiting on a sprint that belongs to someone else.
Underneath it is one idea. We did not build a content pipeline, we built an evidence layer — thirteen systems of record wired into the working session — and every output format is a renderer hanging off it. Nothing is generated from a prompt and a memory. The numbers get fetched, the dataset gets frozen, and only then does anything get built.
Which is why the range is what it is. The same spine produces a landing page, a press release, a competitor audit, a weekly report nobody assembles, and a ninety-second explainer with narration-timed scenes and a generated voice track. Text, web, video, audio. Add a format and you are adding a renderer, not starting a project.
The counts below come from git — a shared record a stranger can audit, not a self-report. What follows is the method behind them, the working log, and the things it does not solve.
- merged pull requests to production
- 108merged pull requests to production
- commits, from marketing not engineering
- 275commits, from marketing not engineering
- live pages, whole site rebuilt
- 249live pages, whole site rebuilt
- systems wired in
- 13systems wired in
- people
- 5people
The argument
What this record actually means: you have six months.
01 · The compression
It was never only about time. It was about people.
Every one of these used to need a room of specialists — brand, product marketing, engineering, animation, SEO, copywriting — moving in sequence, each waiting on the last. Two axes below: how long it took, and how many people it took. Multiply them and you get the real number.
Added up, it is roughly the output of a team six times the size. That is the consequence, not the argument — the argument is the method, and the method is on the rows below.
Rebuild the company website
Not a reskin. Reimagined, rebranded, and roughly three times the size of the site it replaced — design tokens, navigation, page migration, demo redesign.
Elapsed time
People
660 person-days → 16 person-days
Repository opens with Initial commit on 1 Jun 2026. 205 commits in the first eight days, run as phased agent tasks.
Turn a research paper into a launch program
Landing page, gated guide, sources page, blog, press release, two videos, podcast, PR pack, social sequence.
Elapsed time
People
270 person-days → 14 person-days
Token Efficiency Study. Every asset produced by one person — the CMO — not a launch team.
Build a campaign from concept to playable
A working app, a horse race, AI-generated video, landing pages and a brand system around it. Not a deck describing a campaign — the campaign.
Elapsed time
People
240 person-days → 1 person-day
Cloud casino. Conventionally this is an agency brief, a production schedule and a branding round.
Build a webinar campaign, end to end
Story arc, the three reasons for PointFive on Google Cloud, the technical opportunities that differentiate us, landing page, registration form, email copy, ad copy, social posts, branding.
Elapsed time
People
20 person-days → 1 person-hour
A week of coordinated work across four functions, executed inside a single session.
Produce a 90-second product explainer
Conventionally: brand, product marketing, an animation specialist, storyboarding and approval rounds. Now: narration-timed scenes written as code.
Elapsed time
People
48 person-days → 1 person-hour
Re-rendered on a copy change rather than re-edited, so revisions cost minutes instead of another approval cycle.
Build and maintain competitor comparison pages
Positioning from internal IP, claim sourcing with capture dates, page build, browser QA, pull request.
Elapsed time
People
75 person-days → 4 person-hours
Seventeen pages live. An agent now reviews all of them weekly and recommends updates as competitors ship and our own product moves.
Report the week across every channel
Analytics, search, social, CRM, community and AI visibility, read in one pass and reconciled where the sources disagree.
Every week
People
4 person-days a week → none
Runs on its own. The CMO reads the output; nobody assembles it.
Conventional headcount and duration are my estimates from twenty years of running these functions, and are labelled as estimates. Measured figures are counted — commit dates, session records, merged pull requests.
02 · What got built
One square per page.
A number in a sentence is easy to skim past. This is the same number, drawn. Two hundred and forty-nine live pages, grouped by the sections the site itself uses, so any of it can be checked against the sitemap rather than taken on trust.
Worth saying plainly: this was a replacement, not an extension. Fourteen product pages were rebuilt alongside everything else. Counted from the live sitemap on 18 August 2026.
249 pages live, from a site that did not exist on 1 June.
03 · What it did to the numbers
Output is not the point. This is the point.
I took the CMO seat in January. These are the moves since, as percentage change — the underlying counts stay in the building. Reach is what arrived. Conversion is what happened next, which is the half that actually matters.
Reach
January against July
Unique visitors
People reaching the site.
Search impressions
How often the site surfaces in search at all.
Website sessions
Traffic more than doubled.
Pages viewed
Depth held up as volume climbed.
Organic search clicks
Peaked at +157% in June, the month the new site landed.
Conversion
Jan–Mar monthly average vs Apr–Jul monthly average
Event registrations
A repeatable registration path where there was none.
All form submissions
Every conversion path on the site, combined.
Demo requests
The bottom of the funnel moved, not just the top.
04 · The demand engine
This is not a team that did not exist. It is an engine that was not running.
The distinction matters. There were good people here in January, and there were forms — inherited ones, mostly whitepaper downloads, belonging to no program and coded to no campaign. What there was not was a machine: a webinar running every month, an event path that repeated, and an attribution trail from a click to a routed record.
The percentages above come from the CRM, which is the only system that held the answer for the whole period. Analytics did not measure conversions until August, so it cannot be compared against January — the measurement changed, and reading that as a change in performance would be wrong.
Composition is the real finding, not volume. In January, not one submission carried a campaign code — the forms were inherited, mostly whitepaper downloads, attributable to nothing. From April, every single submission does, and the old forms have been retired. That is the difference between collecting names and running an engine.
05 · The team
Five roles, each holding what used to be four or five.
The headcount did not shrink because people got faster at typing. It shrank because the work between having the idea and shipping the thing — the production, the coordination, the chasing — mostly stopped existing.
Story and brand
Narrative, positioning, brand system, the CEO readout.
ConventionallyBrand lead, copy chief, content strategist, internal comms.
Growth and digital
Website, AEO and SEO, paid programs, webinar programs, marketing analytics.
ConventionallyWeb team, SEO agency, paid media agency, marketing ops, analyst.
Product marketing
Use cases, product launches, positioning source of truth, enablement.
ConventionallyTwo to three PMMs, plus a competitive intelligence analyst.
Social and video
Channel strategy, podcast, live shows, video production end to end.
ConventionallySocial manager, producer, videographer, editor, motion designer.
Operations
Program management, timelines, asset status, launch checklists, systems.
ConventionallyMarketing ops team, project manager, production coordinator.
06 · The flow
One loop, run over and over, on everything.
There is no separate AI workflow sitting beside the real one. This is the whole method, and it applies equally to a landing page, a campaign, a video, a board update and a competitor page.
Ideate
The only genuinely scarce input. What is worth saying, to whom, and why now. No agent decides this, and no agent should.
Human, always
Ground
The load-bearing step, and the one most teams skip. Thirteen systems are wired into the session, so the agent reaches the billing record rather than a summary of it — then the dataset is frozen before a word is written. An agent that cannot invent a number does not need to be talked out of inventing one.
Thirteen systems of record
Execute
Render it. A page, a press release, a competitor audit, a narrated video, a pull request — all off the same frozen dataset, which is why the format is close to free. The output is a merged commit or a finished file, not a suggestion to be actioned later.
A commit, not a draft
Prove
Read the result back through the same systems that built it. The agent that made the funnel reports the disappointing number as readily as the good one.
Measured, not estimated
Step one is the only part that did not get cheaper. Which is exactly why it is now the whole job.
07 · The ground truth
The knowledge is already there. Reach further down.
Most marketing teams are not short of data. They are short of access to it at the moment they are making something. Surface metrics are cheap and frequently wrong; the answer sits in the layer underneath.
This is the part that actually matters, and the part that is easiest to skip. Everything else on this page — the multiples, the page count, the videos — is downstream of the four layers below. Get the source of truth wired in and the outputs become cheap. Skip it and you have a very fast way to publish things that are wrong.
Brief, brand system, positioning source of truth
Notion · Drive · encoded brand rules
What every channel actually did this week
GA4 · Search Console · LinkedIn · YouTube · Peec
Who raised a hand, and where they came from
HubSpot · Common Room · G2 · Slack
Billing-grade truth. Every published number resolves here.
Warehouse · CRM · platform data
The discipline that makes the speed defensible: before a printed guide went out to a named account list, the ask to the data layer was explicit — aggregates only, state the sample size you used, flag anything you would not want quoted. The answer came back that one set of totals could not be defended. Those totals were dropped.
And out the other side
One dataset in. Four kinds of thing out.
This is the test of whether an evidence layer is real. If it only produces prose, it is a writing tool. If the same frozen dataset can leave as a web page, a press release, an audit and a narrated video, it is infrastructure.
- Web
- Landing pages, guides, comparison pages
- Next.js, shipped by pull request
- Written
- Press releases, briefs, weekly reporting
- Assembled from the systems, not retyped
- Analysis
- Competitor audits, brand and claim reviews
- Re-run on a schedule as the market moves
- Video and audio
- Ninety-second explainers and how-tos
- Scenes timed to a generated narration track
The video work is the clearest case. Rendered as code rather than cut in a timeline, an explainer is versionable and regenerable — when the product changes you re-run it instead of re-shooting it. Early and on low volume, average view duration on the short explainers has run at roughly 49% against a ~28% baseline. That retention signal is the reason the format is still being tested, and the volume is not yet there to call it more than a signal.
08 · Who reaches what
The overlap is the team. The gaps are the roles.
Andrew Brown, who runs growth, and me — measured the same way, against the same evidence layer. Every cell is a read from a system of record in a working session. If you are trying to work out what an AI-native team actually looks like structurally, this is the diagram I would want to see — not a tool list.
Three systems are genuinely shared
Notion, Slack and Peec come out near-even between us. Those are the ones carrying shared truth — plans, decisions, and how the brand appears in AI search. If a team is going to wire anything in first, it is these.
And two are a single point of failure
Analytics and CRM are lopsided — heavily mine, barely theirs, despite equal access. That is not a tooling gap, it is a concentration risk: if I stop, the team loses its read on performance and pipeline. Access was never the hard part.
Browser and preview automation is excluded — it is the majority of all traffic and it is verification, not evidence. One unidentified connector on my side is also excluded rather than guessed at.
09 · The working record
Nine weeks inside the machine. One person’s log.
Ninety-nine working sessions across fifteen parallel workstreams, 18 June to 18 August. Not chat transcripts — a log of what was actually read, built, queried and shipped.
Everything in this section is from my machine alone. The rest of the team runs the same way on their own, and none of it is counted here — so read these as one marketer’s floor, not the team’s total. The git numbers at the top are the team’s.
Where the evidence came from
2,445 reads from systems that already held the answer — my sessions only. A further 3,858 calls drove a browser to verify the result, which is checking rather than evidence, so it is not counted here.
Calls per system · my sessions only
What the agent actually did
18,995 actions, mine alone. Mostly building and checking, not writing prose.
Actions by type · my sessions only
10 · The honest part
What this does not do.
A page that only lists wins is marketing about marketing. These are the four things that have actually bitten.
Output is raw material, not a finished story
It arrives structured and fast, and it is still not the story. Someone has to decide what to cut and what is actually worth saying. Shipping the first draft is how numbers drift.
Speed moves the bottleneck, it does not remove it
Two weeks from paper to full program was deliverable. Fact-checking it in two weeks was not. Build capacity outran data quality — a sequencing failure, not a tooling one.
No agent owns a claim
Every published number needs a source, a capture date and a named human approver. Competitor claims are the highest-risk content any team publishes, and agents write them at scale.
A missing leader is still a missing leader
This does fill seats — but only underneath a leader who is already there. Where we have no brand lead, we have a gap, and no amount of agent capacity closes it.
The shape of the next hire
We do not need five brand designers. We need one, with AI.
The same is true everywhere. Not five operations people — one or two, with AI. What this model scales is not headcount, it is leverage per person, and the constraint moves accordingly: every function needs someone good enough to direct it, and after that the volume takes care of itself.
Which is why the gaps that hurt are leadership gaps, not capacity gaps. Where we are missing a leader — brand, right now — no amount of agent throughput closes it, because there is nobody to point it anywhere. Scale intelligent people with AI, not more people.
11 · If you run a team
Four changes that did the work.
- 01
Give the agent hands, not opinions
An assistant that drafts copy saves an hour. An agent that reads the warehouse, builds the page, opens the pull request and measures the result replaces a workflow. The difference is write access, and it is the whole difference.
- 02
Make everyone able to ship
All five of us have GitHub and Vercel. When the person with the idea is also the person who can publish it, the handover — and the week it costs — disappears. This is a permissions decision far more than a training one.
- 03
Wire the evidence in before you ask for output
Every hour spent connecting a system to the place the work happens comes back the first time someone needs a number at 6pm on a Thursday. The data was never the problem. The distance to it was.
- 04
Move the gate to the front
When production gets this fast, sequencing becomes the failure mode. Freeze the numbers before anything is built against them. We learned that the expensive way, once.
12 · Outside the team
The obvious objection is that this needs a company behind it. It does not.
Everything above is bounded to one marketing team over one measured window, and it should be read that way. So here is the same method with no team, no budget and no stakeholders — three websites I have written and published on the same stack, one of which is my mum’s garden maintenance business on the Mornington Peninsula.
- techseekinghuman.ai
The podcast — 41 episodes, plus the rendered video format
- daveanderson.com.au
This site. Including the page you are reading
- dirtgirls.com.au
Garden maintenance, Mornington Peninsula. My mum's business
A whole category of tooling stopped being necessary
All three run on Claude, GitHub and Vercel. No WordPress, no plugins, no hosting account, no theme to maintain. That removes about $500 a year, which is not the interesting number — the interesting part is that the line item disappeared rather than got cheaper. There is no CMS to log into because the site is a repository.
And one of them writes itself
Dirt Girls has an agent that researches and publishes gardening and maintenance advice to the site on its own. Nobody briefs it and nobody publishes it. It is the smallest possible version of the argument on this page, running unattended for a business with no marketing function at all.
The video renderer generalises the same way. Two fifteen-minute episodes of Tech Seeking Human are fully rendered with no editor and no timeline — whether AI takes your job by 2030 and when AGI actually arrives. Same pipeline as the ninety-second product explainers, just more of it, and built by one person rather than a team.
The measure that matters is not how much AI a team uses. It is whether the claim survives an audit, and whether the thing actually ships.
Both are now countable. That is the part I would take to any team I worked with next. If you want the argument rather than the evidence, it is in You have six months.