ChatGPT Ads Landing
Build, deploy and verify static landing pages for ChatGPT / OpenAI Ads campaigns — any advertiser, any vertical — using the chatgpt-ads-landing engine (pages// + page.json → build → check → zip) and an optional compatible site-hosting platform. Covers intake of client material, page content, OpenAI Measurement Pixel events (oaiq), UTM/oppref forwarding, per-country geo routing (time zone + IP), footer privacy links, cookie consent for UK/EU, packaging, first-deploy URL backfill, and live verification (pixel really reaches OpenAI, crawlers get 200, live == build). Use this whenever someone wants to 做/改/上线/检测 an OpenAI 或 ChatGPT 广告落地页, mentions chatgpt-ads-landing, oaiq / OpenAI pixel, OAI-AdsBot, "pixel 能触发吗", "UTM 生效了吗", "是否符合 OpenAI 落地页要求", 一键部署 / 自动上线, 落地页体检, 按国家/IP 跳转, or a new client's 落地页 for ChatGPT ads — even if they don't name the skill.
ChatGPT Ads landing pages (chatgpt-ads-landing)
One engine, many clients: each campaign is a pages/<slug>/ folder (static index.html +
page.json), built into a self-contained site, zipped, and deployed as one site on a
site-hosting platform. The bundled assets/starter-page is the reference starting point — copy it when you need a full example.
Why it is built this way: OpenAI reviews the landing page with its crawler (OAI-AdsBot), so
the page must be fully readable as static HTML and must never redirect itself; everything
dynamic (country routing, pixel events, consent) is progressive enhancement in lp.js.
Where things are
- Repo: ask the user where their landing-page repo is . If one exists, work there — it may be newer than this
skill's copy (
diff -rq <repo>/shared <skill>/assets/engine/sharedto tell). - No repo on this machine →
node <skill>/scripts/init-repo.mjs <dir>(copiesassets/engine/, refuses non-empty dirs), thenpnpm installin it. - New page →
node <skill>/scripts/new-page.mjs <repo> <slug>(copiesassets/starter-page/, a vertical-neutral page whose blanks are填【…】— check.js fails until all are filled). - Verify anything live →
node <skill>/scripts/verify.mjs <url> [--dist <repo>/dist/<slug>] [--proxy http://<your-proxy>]. - Re-read OpenAI's rules →
node <skill>/scripts/fetch-official.mjs --proxy …. - Deploy + platform audit via the open API →
node <skill>/scripts/publish.mjs login | deploy | audit | ship(device-code login once; then no manual zip upload). See platform-deploy.md §Open API.
<skill> = this skill's directory. Prerequisites — check them before the first run and tell
the user what is missing: Node ≥ 22, pnpm, Google Chrome, and for any check against OpenAI an
overseas network exit. Ask the user for their proxy URL rather than assuming one (for example
http://127.0.0.1:7890, whatever their local proxy listens on); without a proxy, a system-wide overseas VPN works and --proxy can
be omitted. Run network scripts with the sandbox disabled so the proxy is reachable.
References — read the one you need, when you need it:
| File | Read when |
|---|---|
references/page-config.md | writing/editing page.json or index.html (fields, tokens, content rules) |
references/pixel-events.md | mapping the client's conversion events; pixel debugging |
references/openai-requirements.md | any "does this meet OpenAI's requirements" question; category limits |
references/privacy-consent.md | choosing consent countries / privacy links per market; banner behaviour |
references/platform-deploy.md | packaging, first deploy, URL backfill, platform limits |
references/lessons.md | before starting — the mistakes already made once and the user's preferences |
Workflow A — new client page
- Intake. Collect, and ask only for what is missing:
- brand + campaign, offer facts, eligibility/terms, FAQ (source: brief, Feishu doc via lark-doc, a design HTML, or the client's site);
- target countries and, per country, the client's tracked landing URL (with UTM) — use them verbatim; plus a fallback URL for everyone else;
- pixel ID and the Ads Manager data source's event list (snippets are fine);
- per-country privacy (and cookie) policy URLs from the client's site;
- whether the client's logo may be used (default: our own mark);
- category check: is the vertical allowed (openai-requirements.md §2)? Flag early.
- Scaffold.
new-page.mjs <repo> <slug>. - Content. Fill
index.htmlfrom the material only — no invented numbers, headings or claims; leave填【…】and ask when something is missing. Keep JSON-LD word-for-word with the visible copy. If the client supplied a design, follow it; otherwise the starter's structure (hero with h1 + facts + CTA → benefits → details → FAQ → footer). - Config.
page.json:routing.sitesper country,ctas(hero=geo, one fixed CTA per country if the page lists countries),links.privacy(+cookies),consentfor UK/EU targets,pixelEventsmapped from the client's events (pixel-events.md). New country? Make sure its time zones are inshared/geo.js. - Local check.
pnpm dev(http://localhost:4123/) andnode scripts/check.js <slug>; thenverify.mjs http://localhost:4123/<slug>/ --proxy …for pixel/geo/consent. Stoppnpm dev(dev.js) before step 6. dev.js rebuildsdist/on every file change;publish.mjs ship/package.jsrewritepage.json(url, siteId) and rebuild at the same time, so two builds writedist/concurrently and images/favicon get overwritten → check fails ("引用的本地资源都存在" ✗) and nothing is deployed. Never "fix" it with--force; kill dev.js, runnode scripts/build.jsonce,check.jsmust pass, then ship. - Deploy. Preferred —
publish.mjs ship <repo> <slug>: creates the site, backfillsurl+deploy.siteIdinto page.json, redeploys with the real URL, then runs the platform's landing-page audit (countries =routing.sites).publish.mjs whoamifirst; if not logged in, runpublish.mjs loginand tell the user to open the printed link and approve (they need the platform's app_id and API address — ask, or setBV_APP_ID/BV_API_BASE; this step is optional and only works with a compatible platform). Creating a site is outward-facing: confirm the site name / subdomain with the user before the firstship(deploy --dry-runshows the plan). Manual fallback (no login possible):node scripts/package.js <slug> --force→ user uploads via 新建部署 (entryindex.html, 加密访问 off) → user sends the assigned URL → put it inpage.jsonurl→node scripts/package.js <slug>(must pass) → user uploads with 部署新版本. Always give the exact zip path and say older zips are obsolete. - Live verification.
verify.mjs <url> --dist <repo>/dist/<slug> --proxy …— the platform audit does not test pixel delivery, geo routing or consent. Report per group; if the exit is CN, say pixel delivery is unverified and ask for an overseas node rather than presenting skipped checks as passes. - Hand-off. Changelog (template below) and what the client still owes (copy review, banner text, legal sign-off).
Workflow B — change an existing page
Edit → publish.mjs deploy <repo> <slug> --remark "<what changed>" (or package.js <slug> →
user uploads with 部署新版本) → verify.mjs --dist. For content updates, change
content only and leave styling alone unless asked. Adding a country = routing.sites +
fixed CTA + links entries + time zones in geo.js (+ consent.countries if UK/EU).
Workflow C — "does it work / is it compliant?"
Run verify.mjs first; answer from its output. For a fuller content review add
publish.mjs audit <url> --countries … (the platform's 6-dimension audit; works for any URL,
not only our sites). Its items carry basis: only openai_official items are OpenAI rules —
training items are the training framework's recommendations and must be labelled so. For compliance, separate three kinds of
statement and label each: OpenAI requires (quote openai-requirements.md / fetched text),
law in that market usually expects (privacy-consent.md, not legal advice), our
checklist recommends. Never escalate a recommendation to "OpenAI requires".
Reporting results to the user
- Lead with the verdict, then a short table (event/country → result). Distinguish "verified live", "verified locally", "not verified (why)".
202frombzr.openai.com= OpenAI accepted the event; matching to a conversion is confirmed in Ads Manager, not by us.- Name things the user can act on: exact zip path, exact platform button, exact DevTools path.
Changelog template
## <Client> <page> 更新记录
站点 / 上传包 / 对比版本
### 新增 (user-visible behaviour first, one bullet per capability)
### 修复
### 工程改动 (files touched, one line each)
### 验证 (what was verified, where: local / live, with which exit country)
### 已知限制
### 待办 (deploy, client confirmations, git commit)
- Downloads
- Security audit
- Review
- Last updated
- 2d ago
- Current version
- v1.0.1
- License
- MIT-0
