How I think about content strategy

Content strategy got harder in two ways at once. Producing content became nearly free, which collapsed the value of producing it. And the click that inbound was built on started disappearing, which broke how we measured it.

Most teams responded to the first change and ignored the second. They doubled output, kept the same distribution and measurement model, and six months later there's more content and less pipeline.

The way I organize the work now is two engines. Different jobs, different constraints, different ways they break, one shared quality gate. Most content strategies I look at are one engine wearing the other's clothes.

Engine one: programmatic

Programmatic exists to capture demand that already has search volume behind it. Integration pages, comparison pages, category and use-case templates, glossary and listicle formats. One template, a structured data source, a page per row.

The thing people get wrong is treating production volume as the constraint. It isn't. Getting indexed is, and staying out of the AI-content penalty box is.

I want to be careful about how that gets stated, because it usually gets repeated as a hard Google mechanic and it isn't one. There's no published limit. What exists is a working cadence that operators running these factories have landed on, roughly ten to fifteen pages a week, and the reason it holds isn't that Google refuses the eleventh page. It's that publishing templated pages faster than that is what gets a domain flagged as AI production at volume. Treat it as a safety rule you're choosing, not a ceiling somebody set for you. I got into why the operators who win at this actually win in AI Scales Expertise, and the short version is that every constraint in a good factory is a scar from someone doing the work by hand first.

Domain authority sets a second ceiling. Below roughly DA 40, publishing thin templated pages at volume is how sites get hit by core updates, not how they grow. When I scoped this for a domain sitting at 35, the honest plan started at two or three pages a week against a brief asking for five to ten, and the first move was auditing what already existed. Five hundred-plus indexed pages, several clusters cannibalizing each other. Raising the average quality of those was worth more than any new page we could have shipped that month, which is a boring recommendation to make to someone who hired you to publish more.

Niche selection outranks volume by a wide margin. The factory amplifies existing search demand. It can't create demand where there is none, which sounds obvious written down and is still the single most common way these programs burn a quarter.

Then there's the tell nobody expects. Anything in the pipeline that should look stochastic has to be computed, not asked for. Ask a model for a random number between five and eight and you get six or seven nearly every time, and that regularity is a detectable programmatic footprint across a few hundred pages. Vendor selection, slug patterns, palette, ordering, all of it belongs in code. The model writes.

That last point is the one I'd carry into any version of this. The guideline files, the argument spine, the banned-tells list, the voice canon, those are the asset that accumulates and every bad draft is a chance to update one of them. The prompts themselves are disposable.

Engine two: thought leadership

The second engine exists to create demand and to be worth citing. Completely different job, and it can't be templated, which is exactly why it's defensible.

The failure mode shows up immediately when people try to automate this. AI can emulate the structure of an argument all day, and the structure is exactly what fools you, because developing a thesis and defending it is the part it can't do. So it takes your idea and does a lap around it, restates it three ways, and you finish reading a very well-organized document that didn't say anything.

That's why interviews are the starting point. A person developing an argument out loud goes "I do this, and because I do this, I do this," and the argument accumulates as they talk. That progression is what a first draft from a cold prompt is missing, and no amount of prompt engineering puts it back.

So my pipeline for this engine starts with research instead of drafting. Curated authorities first, meaning my own wiki and the sources I've already vetted, then external search to fill gaps and, more importantly, to go find the material that contradicts my hypothesis. Then a brief gate where the thesis and the angle get named. Then a draft. Then the same deslop and craft edit the programmatic engine runs.

A piece anchored in a real number, a real customer sentence, or a real experiment is a piece nobody else can produce, and increasingly that's the only kind that earns its keep, because anything anchored in the model's priors is something everybody already has.

Which engine gets the topic

The decision tree is short.

Programmatic if there's verified search volume, a repeatable template, and a structured data source you can defend the accuracy of. Miss any of those three and you're about to publish pages that are thin, unrankable, or wrong. Comparison pages are the classic trap here. High commercial intent, obviously templatable, and a hallucination risk until somebody builds the competitor data matrix underneath them. That dependency is a real reason to defer a play, not a reason to be brave about it.

Thought leadership if the piece rests on a claim only you can make, a number only you have, or a position you're willing to be wrong about in public. Also for anything where getting cited matters more than getting ranked.

Some of the strongest work sits in the seam, where a first-party dataset gets templated into something programmatic-adjacent. That's usually where I'd spend the marginal hour.

The shared gate

Both engines fail the same way without an editorial layer with teeth, and this is the part teams skip, because it produces nothing new.

When every team runs the same models, trained on the same corpus, prompted with the same playbooks, output converges toward the mean. So taste becomes the constraint, meaning the judgment to spot mid work and kill it, to pick the angle nobody else is running, and to notice when polished output is generic positioning in a nice suit.

The operational version of taste is unromantic. A deslop pass that strips the tells. A craft edit that applies an actual lexicon of how I talk, so it's catching voice drift and not just grammar. A written list of constructions that get cut on sight. Mine bans the not-X-but-Y flip, the punchy one-word fragment setups, the staccato rule of three, corporate power verbs, and the balanced two-part maxim as a closer, because a too-neat closer is almost always a symptom of an idea that never got developed.

Worth naming a real tension I haven't fully resolved. There's a good argument that familiarity beats originality in ad copy specifically, that proven structural templates outperform novelty because audiences prefer what feels familiar. I think both hold, and the split is familiar in structure, distinct in substance. But I hold that loosely, and anybody telling you distinctiveness always wins hasn't run enough ads.

The click is going away, and it's going away invisibly

Roughly two thirds of Google searches now end without a click. Around half of B2B software buyers start their research inside an AI chatbot, and most of them take the shortlist the model hands them without adding vendors to it.

That last part is the strategic point. If you're absent from the answer, you're absent from the deal, and the loss is invisible in your reporting. The shortlist forms in a conversation that fires no pixel, then the buyer shows up later as branded or direct traffic. Your dashboard shows a healthy direct channel and an underperforming content program, when what actually happened is your content did the work and got no credit for it.

The compensating upside is real. Visitors arriving from AI referrals convert at meaningfully higher rates, because the model already qualified them before they showed up.

Being in the answer is a different discipline than ranking

Answer engine optimization gets treated as SEO with a new name. That rename is the most common mistake in the category right now.

Search was zero-sum. Short queries, three or four words, same ten links, so you fought for position against a fixed number of slots. AI prompts run much longer, more like twenty-plus words, which creates effectively unlimited inventory and makes the old slot-one fight mostly irrelevant. The game is being the right answer to a specific, context-rich question.

That changes what to produce, for both engines. No single page ranks, because the model synthesizes across sources, so volume of genuinely citable material compounds and the old cannibalization worry mostly evaporates. Clear extractable structure matters more than keyword placement, being quotable matters more than being comprehensive, and distinctive framings and named frameworks get cited because they're easy to attribute, which is the thought leadership engine paying into the programmatic one.

Most of the mentions models make of you come from off your own site, which means the channels that feel least measurable are doing more of this work than your content calendar is. The cheapest reconnaissance available is running your fifteen highest-intent buying questions through the major models once a month and logging where you're named and where you're absent. Takes about ten minutes and it tends to reset the roadmap.

The uncomfortable part is that this is demand creation moved into the least trackable stage of the funnel, which is exactly why it stays underfunded relative to what it does.

Distribution is the constraint, not production

Content nobody sees is a failed demand creation program wearing a content program's reporting.

The habit worth building is reformatting one idea for every medium instead of publishing once and moving on. The same argument becomes a post, a document ad, a section of a talk, a newsletter piece, a page. That's one thesis meeting people where they already are, and it's cheaper than generating five new ideas at the same quality.

Proof beats claims, and you can manufacture proof

Two moves I lean on hard.

Show, don't tell. Instead of describing what you could do for someone, build the finished artifact and hand it over first. A sample built for their company, a working table with their data, the actual deliverable. AI makes this viable per prospect at a scale that used to be reserved for a handful of logos, and it converts because it's substance where the alternative is an identity token pasted into a template.

Performative proof. We get rated on a timeline rigorous attribution can't meet. The bar that matters is producing something good enough that a skeptical exec has to concede the point and look closer. Worse proof today often beats better proof in three months, because the budget conversation is happening now. The honest version calls itself directional and doesn't pretend to be an attribution platform.

What I'd look at first

• Whether the content has a point of view somebody could disagree with.
• Whether anything published last quarter contains a number or a sentence only this company could have produced.
• Whether the programmatic cadence is a number somebody chose on purpose, or just how fast the machine can write.
• Whether the guideline files get updated when a draft comes out bad, or whether somebody just fixes that draft.
• Whether ideas get reformatted or published once.

And whether anyone has asked the models what they recommend in your category. That one's free.