How I Built a Content System With AI — 8-Week Case Study
Thirty-nine million views in eight weeks, and what actually produced them
Between 20 July and 13 September 2026, a brand my wife and I started from nothing did 39.5 million views on Instagram, reached 12.9 million unique people and gained 44,198 followers. In the same eight weeks it acquired a 54-article website, a brand system, a media kit and an email programme.
Two people. No team, no agency, no studio, and not a dollar of paid amplification.
I am not writing this because the numbers are large. I am writing it because the interesting part is not the growth — it is that a two-person operation produced the volume of assets normally associated with a marketing department, and that most of that gap was closed by workflow rather than effort. That part transfers. The growth might not.
Everything below is inspectable. The brand is My French Husband. The site is live, the articles are indexed, and you can view source on any of it.
What the brand is, briefly
My French Husband is a comedy-documentary account: an American wife filming her French husband's reactions to American life. Supermarkets, malls, beach wagons, coffee creamer.
The format matters for one reason. Cultural contrast is a self-renewing content engine with a debate mechanic built into it. Every episode poses a question two audiences answer differently, which means the comments do not merely react — they argue, and they supply the next episode. Ninety-seven per cent of our views came from people who did not follow us. That is a distribution characteristic of the format, not a reflection of a clever posting schedule.
I mention this because it is the honest boundary of the case study. The format did the distribution work. The system did everything else.
The thesis
Distribution has become the cheap part. Reach is abundant and getting cheaper; a phone and a genuinely good format will out-distribute a media budget, and we have the numbers to show it.
What remains expensive is converting distribution into assets you own — searchable answers, an email list, a brand that survives a platform change, a site that an AI answer engine will cite. That conversion is an operations problem, not a creative one. It is where most brands with strong social performance quietly leak everything they earn.
The system below exists to solve that one problem.
The stack, and what was rejected
Not a tool list. Every line is a decision.
Filming: a phone. No camera, no rig, no lighting. The single most useful piece of equipment we own is a small tripod and we forget it most of the time. A format that rewards immediacy is actively damaged by production value.
Editing: Edits, Meta's own app. Free, and built for the format we publish in. It handles captions, cuts and export. For a two-person operation posting reels, a desktop suite would buy us almost nothing and cost us a step.
Demand validation: vidIQ, before filming rather than after. Several of our articles exist in the shape they do because the search demand was measured first. Filming a day into a topic nobody searches for is the most expensive mistake available to a small team.
The website: hand-built static HTML. No CMS. This is the decision people question most. There is no WordPress, no framework, no build step. It loads instantly, there is nothing to patch, no plugin surface to be hacked through and no monthly fee. The cost is real and worth naming: site-wide changes need a script rather than a click. That cost is exactly what the workflows below are for.
Monetisation: ShopMy and LTK in parallel, because the audience genuinely splits across both apps.
What we deliberately skipped: scheduling tools, because posting takes a minute and the platforms reward native uploads. A desktop editor. Anything with a subscription that replaced a decision rather than a task.
That last line is the filter. Most stack bloat is a tool bought to avoid making a choice.
The AI workflows
This is the part worth reading, and I am going to be specific, because "we used AI" is not a method.
Comments become briefs. Published reels generate comment summaries. Those summaries name the questions an audience is already asking — whether French women secretly get Botox, which French foods Americans refuse to eat, what a French husband makes of Home Depot. Each becomes an article brief with the demand already demonstrated. We are not guessing at topics; we are transcribing them.
Articles generate from a data structure, not a blank page. Each article is a dictionary — headline, summary, verified facts with sources, body sections, comparison table, FAQ — rendered through one template. The consequence is that all 54 articles carry identical Article and FAQPage schema, a table of contents with stable text-derived anchors, a share row, a newsletter block and a unique 1200×630 social card. Not because anyone remembered. Because the template cannot produce a page without them.
Shared components have exactly one definition. Footer, navigation, share row: each lives in one canonical file and is stamped across every page by a script. Change the file, run the script, and 70 pages update identically. This sounds trivial. It is the difference between a site that stays coherent at 54 articles and one that quietly diverges at fifteen.
Assets generate from the same sources they document. The icon library, brand marks, favicons at four sizes, the brand book as a web page and as an eleven-page PDF — all generated from one folder of SVGs. The brand book cannot show an icon that does not exist, and cannot drift from the assets, because it reads them at build time.
Verification is inside the loop, not after it. Every page is parsed for unclosed tags, every block of structured data is JSON-validated, every internal link is resolved against the filesystem, and interactive pages are executed in a headless browser before shipping. Colour contrast is calculated rather than eyeballed.
This fifth workflow is the one most people skip and the reason the system holds. Generation without verification does not save time; it moves the error downstream to where it is more expensive.
What AI did not do. It did not decide what the brand is, what was funny, which claims were unsupported, or what to cut. Every judgement call in eight weeks was human. What AI removed was the cost of consistency — the reason small teams historically produce inconsistent output is not that they lack taste, it is that applying taste 54 times is expensive. That cost is now close to zero. The taste is not.
If a vendor tells you their AI replaces the judgement, they are selling you the wrong half.
What broke
A case study without failures is marketing. Four real ones.
Placeholder copy nearly shipped. A page went to the verge of publication with ONE HONEST SENTENCE ABOUT YOUR OWN EXPERIENCE GOES HERE still in the body. Generated systems fail silently in a way handwritten ones do not; a human writing that page would never have left the words there.
A stale component broke 23 articles invisibly. The article generator kept its own pasted copy of the share row. When the shared version was updated, the generated articles silently kept the old one, and the share buttons were dead on 23 pages until a browser test caught it. The fix was structural — the generator now reads the component at build time rather than holding a copy. One definition, or drift is inevitable.
Two analytics exports disagreed by nine million views. Both were labelled "last 90 days," taken a minute apart. One was still loading. Had I used the first, I would have understated the account by roughly a quarter in a media kit sent to brands.
A page was blocked two ways at once, which achieved the opposite. A section was set to noindex and disallowed in robots.txt. A crawler obeying the disallow never fetches the page, so it never reads the noindex — and can still list the URL from inbound links. The block was producing precisely the outcome it was meant to prevent.
That last one is not an AI failure. It is an ordinary technical SEO error, and it is in this list because the interesting failures in a fast system are usually the boring ones.
The number I am least happy with
1,700,585 profile visits produced 3,122 external link taps. Just under two tenths of one per cent.
Twelve million people watched. 1.7 million came to look at the profile. Roughly three thousand went to the site.
That is the weakest link in the funnel by a wide margin, and it is the one entirely within our control — a bio link and a caption habit, not an algorithm. I include it because it is the most instructive figure on this page. A brand can be excellent at distribution and still be poor at conversion, and the two are separate disciplines. Most of the social-first brands I have worked with are strong at the first and assume the second follows. It does not.
What transfers to a brand, and what does not
Transfers directly:
Generate from one source. Any asset that exists in two places will diverge. This applies to footers, disclaimers, product copy, PDP modules and brand guidelines equally.
Verification in the loop. Automated checks for schema validity, broken links and rendered output cost a day to build and pay back permanently. Almost no mid-size commerce team does this.
Validate demand before production. Cheaper at every scale, and the discipline survives contact with a marketing calendar.
Answer-first structure. Every article opens with a direct answer before the argument. That is what gets cited by AI answer engines, and it is a formatting decision, not a writing one.
Does not transfer, and I will not pretend otherwise:
Speed. Two decision-makers with no approval chain, no legal review and no brand committee is not a repeatable organisational condition, and those structures exist for reasons — most of them good. Inside a company the honest equivalent is not "move as fast as a couple on a sofa." It is reducing the number of decisions that need a meeting by making the system enforce the standard, so review is spent on judgement rather than on consistency.
That is the actual transferable claim: push consistency into the system, so humans are only spending attention on things that need a human.
Questions I get asked about this
Did AI write the articles? AI generated them from structures I specified, with facts I sourced and verified. The judgement — topic, argument, what to cut, which claims were unsupported — was human throughout. The distinction matters: AI removed the cost of consistency, not the cost of thinking.
Why a hand-built site instead of a CMS? Speed, security surface and cost. A static site loads instantly, has nothing to patch and no monthly fee. The trade-off is that site-wide changes require a script. For a site whose content is generated anyway, that trade-off is strongly favourable.
Does the social audience actually convert to the site? Barely, so far — around 0.18% of profile visits become link taps. That is the honest answer and the current focus of work.
Is this repeatable? The workflows are. The growth is one instance, and I would not claim a two-person account doing 39.5M views in eight weeks is reproducible on demand. What I will claim is that the asset production — 54 articles, a brand system, a media kit, an email programme, in eight weeks — is a function of method rather than luck.
How long did the site take? It was built alongside the account, not before it. The two ran in parallel, which is a different claim from one having caused the other.
If this is the kind of problem you have
I work with commerce teams on exactly this gap: brands with real distribution and no system for converting it into owned assets that search and AI answer engines will surface.
The most useful first engagement is usually a content system audit — a fixed-scope review of how your content is produced, where it diverges, what an AI answer engine can actually cite from it, and which three workflows would remove the most recurring cost. It takes about two weeks and ends in a document your team can execute without me.
Get in touch and tell me what you are producing and how often it goes wrong.
