Running a marketplace with AI: content, SEO and ops on autopilot

The platform layers — ingestion, enrichment, search, checkout — are what people ask about. But what determines whether a small team can actually operate a marketplace is everything around them: content that brings traffic, SEO that doesn't rot, monitoring that catches problems, demos that sell the thing. We run that layer largely with AI agents, under one rule that makes it safe: AI authors, deterministic systems execute, humans gate.

The rule: AI authors, deterministic systems execute, humans gate

Everything below follows the same pattern. The AI does the slow, creative, one-time work: write the article, design the storyboard, analyze the audit. A deterministic pipeline does the execution: publish on schedule, render the video, run the crawl — repeatably, cheaply, with no model in the hot path. And a gate — sometimes automated checks, sometimes a human — decides what ships. Autonomy isn't granted by trust; it's granted by track record, one gate at a time.

Content at scale, with a citation gate you can fail

Commerce traffic starts with content, and content is where AI generation earns its terrible reputation. Our pipeline produces articles at a scale a small team never could — hundreds per campaign, scheduled at a steady two per day, natively multilingual — but the interesting part is the gates, because a citation gate you can fail is what separates a content pipeline from a content cannon:

• Evidence before claims. The pipeline resolves every reference and fetches real source material — abstracts from medical databases, regulator publications — first. Then a model grounds each factual claim against that evidence with three verbs: ground it (keep, cited), soften it (rewrite to what the evidence supports), or drop it.
• Quarantine, not publish-and-hope. An article with unresolvable references, or one where too many claims had to be dropped, doesn't ship. It goes to quarantine for a human look. Fabricated citations — the classic LLM failure — are caught precisely because every reference must resolve to a real source.
• A tiered source whitelist. Peer-reviewed sources outrank regulator guidance, which outranks curated tertiary sources — and the lowest tier is never allowed to be the sole support for a number.
• Grounded variants for the strictest cases. For some campaigns we constrain generation to a single vetted source per article — the model may only restate and structure what that source says. Hallucination drops to near zero because there's nowhere for it to come from.
• Images with a no-reuse ledger. Every article gets a unique, topic-matched, properly licensed image, enforced by a global ledger. (The ledger exists because an early bug put the same photo on two hundred articles. Gates get built from scars.)

Does it work for SEO? Generated content that is thin gets punished; generated content that is gated behaves like content. Search engines rank usefulness, not the tool that typed the draft.

Product demo videos an agent can re-render

Every feature needs a demo, and demos rot faster than docs. We open-sourced our answer: aidemo, an engine where a coding agent turns a one-line brief — "record a 45-second demo of the checkout flow" — into a declarative storyboard, and a headless pipeline renders a narrated, captioned MP4: synthesized voice first, then a deterministic browser recording that replays the exact scripted actions, then word-level captions, then composition (trim idle time, sync scenes to narration, auto-zoom, music).

The design choice worth stealing mirrors this whole article: there is no LLM in the capture loop. The AI authors the storyboard once; the render is a repeatable machine step. Which means demos re-render in CI whenever the product changes — same storyboard, fresh footage, ~zero marginal cost — and one recording can ship in multiple languages, since only narration and captions localize while the footage is shared. Marketing collateral that maintains itself, because it was authored once and executed deterministically.

SEO maintenance: a crawler that emails you at 06:30

SEO for a marketplace isn't a launch checklist — catalogs churn daily, and every churn can break something Google sees. Three automated layers keep ours honest:

• A daily site audit. Every morning, a crawler walks every page — sitemap plus discovered links — and diffs reality against intent: broken links, redirect chains, canonical mismatches, pages that accidentally went noindex, sitemap drift. The findings arrive as an email digest before the workday starts. Most days it's boring; the days it isn't, it caught a regression the same morning it shipped instead of three weeks later in a traffic graph.
• Structural SEO as platform features. Dynamic sitemaps, structured data on every product page, instant-indexing pings on content changes, an llms.txt for AI crawlers — built once, maintained by the platform, inherited by every marketplace on it.
• Measuring visibility inside AI answers. A growing share of product discovery now happens inside ChatGPT and AI-generated search overviews, so we probe them monthly: does the brand appear for the queries that matter, is it cited, how does that shift over time? You cannot optimize a channel you don't measure — and most teams aren't measuring this one yet.

Weekly ops reviews and the monitoring budget

Once a week, an agent runs a read-only sweep of production: firing alerts, uptime, error rates, infrastructure headroom, ingest job health, order flow, recent deploys, cost anomalies. The output is a findings log with an owned to-do list — the same document every week, so nothing discovered ever silently disappears. It's the junior-SRE job nobody staffs at a small company, done reliably because it's a checklist and checklists are what agents are best at.

Watching real sessions (the bug analytics can't see)

Dashboards tell you that conversion dipped; session replay tells you why. We capture anonymized, consent-gated session recordings with friction scoring — rage clicks, dead ends, abandoned checkouts bubble to the top of a review queue. The class of bug this catches is the kind no error tracker ever will: everything returns 200, and the experience is still broken. Our favorite specimen: certain catalog requests with long query strings were being silently rejected by a firewall rule at a size threshold — no error in any log we owned, invisible in aggregate analytics, obvious in thirty seconds of watching a real session fail to load results. The fix was one request-shape change; finding it without replay might have taken a quarter.

What stays human

Honesty requires the other list. What we deliberately do not automate: vendor relationships and commercial terms; pricing and promotion decisions; anything legally binding (the compliance machinery enforces rules humans chose); production deploys (agent-prepared, human-triggered); and the final gate on any content that carries the brand's name into a customer's feed. The agents make the team senior — they don't make it optional. And every automated gate above exists precisely so that the humans in the loop spend their attention where judgment matters instead of where checklists do.

Frequently asked questions