Marketing run as software for a tech brand — versioned AI skills, a GA4-driven loop, confidence gating

The problem most tech brands recognize

You have a product good enough to grow on, and no marketing department to grow it. So the work falls on engineers between sprints, or on an agency that bills you to hand-write blog posts one at a time and nudge ad budgets in a dashboard once a week. Either way you’re paying people to do work that doesn’t need a person: write the page, tag it for search, decide which campaign gets the next dollar, read the numbers, repeat. It’s mechanical, it’s repeatable, and it never stops needing to be done. That’s the shape of work that should be software, not a hire.

This brand had exactly that problem and named it plainly: grow without standing up a team, and without renting one by the month. The product already existed and already worked — what was missing was the distribution layer around it. The brief was to make content production, SEO, ad allocation, and measurement run as a system instead of a payroll line. That framing is what separates this from a launch project: there was no audience to build from a cold start, but there was a real product, real positioning, and a real budget that needed to be spent against evidence rather than instinct.

The timing matters. Gartner projects that by 2028, 90% of B2B buying journeys will be influenced by AI agents, touching more than $15 trillion in purchases. When buyers arrive through an AI answer instead of a ranked link, being citable by those systems stops being optional — and that’s not a problem you solve by hiring one more content writer. It’s a problem you solve by changing how pages are built and how they’re sourced, which is an engineering and process question, not a headcount one.

The constraints we worked under

Four constraints shaped every decision, and they pulled against each other in ways worth being explicit about.

No marketing headcount, and no appetite to add one. Whatever we built had to run on its own and stay cheap to run, which ruled out anything that needed a human in the loop on every page or every budget change. “Automated” couldn’t mean “unsupervised,” though — a tech brand’s name is on every page that ships, and wrong output going live unwatched was off the table. Those two constraints fight: the cheap, hands-off path is to let the model publish freely, and the safe path is to review everything by hand. The whole design is an attempt to find the narrow band between them where the system runs itself but never ships something nobody would stand behind.

It also had to be durable. White-hat only, nothing that gets clawed back when a search or answer engine tightens its rules. We treated any tactic whose value depends on a loophole as a liability rather than a win, because the cost of a manual action or a de-listing lands on the brand’s name, not on a campaign you can quietly retire. And the underlying engine had to be auditable end to end — if the system decided to publish a page or move a budget, we had to be able to say why after the fact, without reverse-engineering a black box.

How we built it

We defined the stack around the brand first, then built the automation on top of it. The split matters: the site is a normal, boring, maintainable content site, and the intelligence lives in the pipeline that feeds it. That keeps the part that holds the brand’s name simple and inspectable, and isolates the moving parts where things can go wrong.

Server rendering as a hard AI-visibility constraint

The content layer is server-rendered with the Next.js App Router and deployed to the edge via OpenNext on Cloudflare. Server rendering wasn’t a default we reached for out of habit — it was a requirement we could point at. An analysis of 500M+ GPTBot fetches found AI crawlers effectively don’t execute JavaScript, so a client-rendered SPA looks like an empty shell to them. The page that renders perfectly in a browser is invisible to the exact systems Gartner is describing.

The alternative we considered was a client-rendered app with a prerendering or hydration shim for bots — detect the crawler, serve it a snapshot. We rejected it on two grounds. First, it’s a maintenance tax that compounds: every new page type needs its snapshot path verified, and the failure mode is silent — content simply stops being seen, with nothing in the logs to tell you. Second, serving bots a different payload than users is exactly the kind of tactic that ages badly when an engine changes how it crawls. Rendering real HTML on the server for everyone is the version with no special case to maintain and nothing to get penalized for. It costs a little more at build and serve time and removes an entire class of “why isn’t this page getting cited” debugging.

Content as versioned skills behind a regression gate

Content is produced by custom skills — small, single-purpose LLM routines, each tuned for both classic search and AI answer engines. The important word is versioned. Each skill is a tracked artifact with its prompt, its parameters, and an evaluation set that defines what “good output” means for that one job. When a skill changes, it runs against that eval set before anything it produces can publish, and a regression — output that drops below the bar the previous version cleared — fails the gate and blocks the release rather than shipping and waiting for a reader to find it.

This is the piece that makes “continuously produced” safe to leave running unattended, and it’s a direct answer to the cheap-versus-safe tension above. The naive version of marketing-as-software is one big prompt that writes everything; it’s cheap to build and impossible to trust, because a small change to the prompt silently shifts the quality of every page and you find out in production. Splitting the work into narrow skills means a regression is localized — you know which job degraded — and the eval gate means quality is a number that has to clear a threshold, not a judgment call someone has to make on every draft. The cost is real: each skill needs its own eval set built and maintained, and that’s work. We paid it deliberately, because an eval set is the only thing that lets the system publish without a person reading every page, and that hands-off publishing was the point.

One GA4 signal driving two downstream consumers

GA4 is the single signal driving the whole loop, and “single” is a design decision, not a convenience. The same analytics that tell us which content is working also drive a downstream ad-allocation step. Content priority and ad budget both move off one measured source of truth, which means they can’t drift apart the way they do when a content team reads one dashboard and a paid team reads another and the two quietly optimize for different definitions of “working.”

Concretely, the signal fans out to two consumers from one place: the content side reads it to decide what to produce or refresh next, and the allocation side reads the same numbers to decide where the next ad dollar goes. We considered letting each side keep its own metrics — paid optimizing to its platform’s conversion events, content optimizing to engagement — and rejected it because that’s how the two ends of a funnel end up pulling against each other. Funneling both through GA4 costs some fidelity (a platform’s native conversion view is richer than what GA4 sees) but buys coherence: there is exactly one definition of a win, and both the content engine and the budget engine are optimizing the same one.

Confidence gating and an audit trail instead of a black box

Two properties keep “fully automated” honest. First, every automated decision is logged and traceable: you can ask why a given page shipped or why a given budget moved and get an answer from the record, not a shrug. That auditability was a requirement, not a nice-to-have — a system that holds the brand’s name has to be inspectable after the fact, and a model that can’t explain itself is a liability the moment a decision looks wrong.

Second, the decisions are confidence-gated. When the signal is clear the system acts on its own; when a call is low-confidence — a thin or noisy signal, an ambiguous result — it’s escalated to a person instead of being made anyway. This is where the “automated but not unsupervised” constraint actually lives. The gate is tuned to send the genuinely uncertain calls up and let the obvious ones through, so a person spends their attention on the handful of decisions where judgment changes the answer, not on rubber-stamping the ones the data already settled. Get the threshold too loose and the brand ships things nobody reviewed; too tight and you’ve rebuilt the manual process you were trying to replace. Most of the operational tuning on this build was finding that line.

GEO as a content discipline, not a trick

On GEO we treat being citable as a property the content earns, not something bolted on after. The lever with the most evidence behind it is sourcing: the original GEO research (KDD ‘24) found that adding citations, statistics, and quoted sources lifts a page’s visibility in generative answers by up to about 40%, with “cite sources” the single most effective tactic. So the skills write that way by default — claims carry sources, numbers carry provenance. The same habits that make a page trustworthy to a human reader are the ones that make it citable by a model, which is why this folds into the content skills rather than being a separate optimization pass.

The discipline is white-hat by construction. No scraping, no doorway pages, no serving the answer engines something different from what a reader sees. Search providers have folded manipulation of generative answers into their spam policies, and any tactic whose value depends on the engine not noticing has a short shelf life — exactly the kind of liability the durability constraint ruled out from the start.

The result

Marketing that runs as software: visible across both search and AI answers, producing and optimizing on its own, with people stepping in only where judgment actually changes the outcome. The brand got the output of a marketing team without the headcount, and without handing the work to an agency to grind out by hand. Because the intelligence lives in versioned skills and the loop reads from one measured signal, the pipeline compounds quarter over quarter instead of needing to be rebuilt — and the same engine, re-pointed by config and skills rather than rewritten, would do the same job for the next product the brand wants to grow.