noon-ads

# Noon — Ad Manager > **PREREQUISITE:** Read `../noon-shared/SKILL.md` for login, page > structure, and common patterns. ## 0. FIRST — declare what this task is for **Before any browser action, call `vibe_seller_declare_ad_task`.** This is a precondition, not a courtesy: a report with no declaration behind it is refused. Two things follow from the declaration and cannot be changed afterwards — how much the completeness gate asks of you, and whether the user gets a review console. Read them off what the person actually asked for, not off what you expect to find once you look. | They asked for | `kind` | `scope` | |---|---|---| | "审计一下广告" / "review our ads" | `audit` | omit `combos` — whole store | | "复核 widget-006 在 amazon 的广告" | `audit` | that platform+country, those campaign ids | | "把这三个词的出价降下来" | `audit` | those campaigns | | "帮 widget-006 建关键词广告" | `create` | the market you are creating in | | "执行刚才确认的调整" | `execute` | the campaigns being changed | | "我们 SA 的 ACOS 大概多少" | `investigate` | the market asked about | **Numbers or changes?** That is the line between `investigate` and `audit`, and it matters more than the table above. If the answer the person wants is a FIGURE — what did we spend, what is our ROAS, how did last month go — that is `investigate`, however many markets it spans: you are reading, not proposing. If they want to know what to CHANGE, or asked "is this right?", that is `audit`, and it opens a review console. Observed live: a plain "how much did we spend on ads this month, and what came back?" was declared a whole-store `audit`, which put five marketplaces of drill work behind a question that wanted four numbers. And the other way round, observed in CI: "review the keyword bids on <campaign>" was declared `investigate`, so the bid recommendations it produced reached the user with no console to approve them on. **If you got it wrong, correct it upward.** If you declared `investigate` and the work turned out to produce decisions — a table telling someone to raise, cut, pause or negate — call `vibe_seller_declare_ad_task` again with `kind: "audit"` and the same scope, or a narrower one. That is the ONLY kind change allowed inside a turn, and it is allowed because it only ever makes you owe more: the coverage obligation appears and the user gets the console. Going the other way (`audit` → `investigate`), adding a marketplace, or widening the campaign list are all still refused, and still need a new message from the user. Do not instead delete your recommendations to fit the declaration — the person asked what to change. **There is no `edit` kind.** A request to change specific bids is an `audit` whose scope names those campaigns — the user still reviews the change before it is applied, and the scope is what makes it small. **The scope trap:** omitting `combos` means EVERY marketplace this store sells on, and you will be held to all of them. Omit it only when the request really is store-wide. A store selling on five marketplaces has been asked for all five because a one-product task left the field out. **If the request names a product, not a campaign** ("widget-006 的广告"), declare the MARKET now and the campaigns later. You cannot know campaign ids before you have looked, and declaring a market with no campaign list means **every campaign in it** — which is how a one-product request turns into an audit owing a whole marketplace. So: declare `{kind, combos, products}` first, enumerate the campaign list, find the campaigns carrying that SKU family, then call the tool AGAIN with the same kind and the same combos plus `campaigns: [...]`. That second call is a **narrowing refinement** and is accepted within the same turn. **Refining may only ever remove reach.** Same kind, same marketplaces, and a campaign list going from "all of them" to a named subset. You cannot add a marketplace, change the kind, or widen the campaign list — a wider scope needs a new message from the user. If a gate asks for something outside your scope, say so in your result and let the user redirect you; do not try to re-declare around it. **When the user sends a NEW message that changes what you are doing, declare again.** That is a new phase, and it is the only way a declaration changes. Two cases you will hit often: - "现在把刚创建的广告复核一下" after a `create` phase → declare `audit`, and its scope may name the campaigns you created earlier in this same task. - The review console submits the user's decisions back as a follow-up message → declare `execute`, scoped to the campaigns that submission actually names. --- Mechanics for noon Ad Manager. The actual *thinking* — when to create / tune / kill a campaign, how to research keywords — lives in the three reference files (see § 11). **URL**: `https://admanager.noon.partners/en-{cc}/home?mpCode=noon&project=PRJ{project_id}` Left nav is an **icon rail** (not a text `ul[role=menu]`): **Campaigns**, **Recommendations**, **Budget**, **Vantage**, **Billing**, **Settings** — the labels render as `role=menuitem` spans, so match by text, not by `ul[role=menu] a`. > ⚠️ **The Ad Manager was redesigned (verified live 2026-07-21).** The > `/home` landing is now an **Overview** dashboard (KPI panels + > promos), NOT the campaign list. The "Ad Manager" card carries three > tabs — **Overview | Performance | Campaigns** — and the campaign list > lives under the **Campaigns** tab at `…/home?…&tab=campaigns`. A page > that also shows a **Sale Event Optimizer** widget (event budget/bid > boosts) and a **Recommendations** panel is the current build. See § 2 > for how to enumerate the list — the old paginator is gone. ## 1. Campaigns Overview Metrics Overview-tab KPI cards: | Metric | Formula / Notes | |--------|-----------------| | ROAS | Return on Ad Spend = Revenue / Spends | | Revenue | Total ad-attributed revenue | | Spends | Total ad spend | | eCPC | Effective Cost Per Click | | CTR | Click-Through Rate = Clicks / Views | | Orders | Ad-attributed orders | | Clicks | Total clicks on ads | | Views | Total ad impressions | | ATC | Add To Cart count | | CvR | Conversion Rate (shown on campaign detail) | A time-series performance chart below lets you toggle any metric. ## 2. Campaign List — Filters, True Totals, Full Enumeration Open the **Campaigns** tab (`…/home?…&tab=campaigns`). Filters above the list: - **Search** box (by campaign name) - **Date range** (default `Last 30 days`) - **Ad Type**: `All types` / `Product` / `Brand` / `Display` - **Targeting**: `All targeting` / `Auto` / `Manual` - **Status count control** (segmented): `Live N` · `Paused N` · `All N` + a `More status filter` dropdown. The counts are the **true totals** for the current filter — read them directly; there is no page math. - **Export all campaigns** (top-right of the list) — a list-level bulk export (distinct from the per-tab Export Data in § 7). Columns (horizontally scrollable): Campaign, Status, Budget, Revenue, ROAS, Ad Spend, eCPC, Orders, Views, Clicks, ATC, … Actions. ### Enumerate EVERY campaign — the paginator is gone The Ant-Design pager (`li.ant-pagination-item-N`, "15 items per page") **no longer exists**. The list is now a **lazy-loaded, inner-scroll table**: only ~15–20 rows render on first paint, and **`window.scroll` does nothing** — you must scroll the list's own container until every row loads. Skipping this silently under-counts: an unscrolled read that sees the first ~20 rows can miss more than half the Live campaigns when the `Live N` chip is much larger, which then fails the completeness gate. > ⚠️ **Enumerate LIVE from the scrolled list — never from a pre-existing > or downloaded export file.** A leftover `Campaign_*.csv` / > `Export all campaigns` file in `~/.vibe-seller/downloads/` (from a > prior run, or a first-paint export before you scrolled) captures only > the rows that were loaded when it was written — typically the first > ~20. Drilling that file makes the audit look done at `20/20` while the > account has far more Live campaigns. (Live failure this fixes: an agent > reused a 20-row export and reported noon SA `20/20` when the `Live` > chip showed **45**.) **Any campaign set whose count is below the > `Live N` chip is stale — re-enumerate by scrolling (below); and if you > do use `Export all campaigns`, first scroll the list fully, then verify > the file's row count equals the chip before trusting it.** **What you owe is what you DECLARED.** `./AUDIT_TARGETS.json` lists every country the store is configured for — the MENU, written by the server from Settings. Your `vibe_seller_declare_ad_task` call turns part of it into an obligation: the countries you declared each need their own `AUDIT_SCOPE.json` combo entry (step 4) AND their own `## noon <CC>` report section; declaring no combos at all (a whole-store audit) means you owe every country in the file. Do NOT expand past your declaration because the menu is longer — reporting on a market nobody asked about is an out-of-scope gap. A country with genuinely no Live campaigns is still written down — an entry with `"active_ids": []` and `"total_active": 0`, plus a section saying so; 可以为空，但不能不写。Omitting a declared combo is a `[基线]` gap that blocks submission. Phase 1 (Discover) MUST, per country: 1. **Read the true total** from the status chips — the `Live N` / `All N` numbers are your completeness target. The chip *label* (`Live` today) is English on the noon partner console, but rely on the **number** and the **`/campaign/details/` links** — both language-neutral — so this works whatever the seller-market locale; don't key on the label word: ```bash browser-use <<'PY' print(js("return JSON.stringify([...document.querySelectorAll('*')].filter(e=>e.children.length<=2 && /^(Live|Paused|All)\\s*\\d+$/i.test(e.textContent.replace(/\\s+/g,' ').trim())).map(e=>e.textContent.replace(/\\s+/g,' ').trim()))")) PY ``` 2. **Scroll the list container to the bottom** (match its class by the `CampaignListRevamp_` prefix — the hashed suffix changes per build; fall back to any inner `overflow-y:auto` scroller taller than its viewport). Repeat until the campaign-link count stops growing: ```bash browser-use <<'PY' for _ in range(12): js("[...document.querySelectorAll('*')].filter(e=>{var s=getComputedStyle(e);return (s.overflowY==='auto'||s.overflowY==='scroll') && e.scrollHeight>e.clientHeight+50;}).forEach(e=>e.scrollTop=e.scrollHeight)") wait(1) print("links:", js('return document.querySelectorAll("a[href*=\\"/campaign/details/\\"]").length')) PY ``` 3. **Extract IDs** with the still-valid `a[href*="/campaign/details/"]` read (§ 3) and de-dupe. Only when the distinct count matches the chip total (e.g. `Live N`) is the manifest complete. Re-run this whole loop **after every country switch** (`/en-{cc}/`). 4. **Persist the scope — required for audits.** Append this combo to `./AUDIT_SCOPE.json` at the task root, with the de-duped ids **and** the `Live N` chip number: ```json {"combos": [ {"platform": "noon", "country": "AE", "total_active": 45, "total_active_source": "chip:Live 45", "active_ids": ["C_DEMO0001", "C_DEMO0002"]} ]} ``` The server requires `total_active == len(active_ids)`, and rejects the scope when they disagree. That is deliberate: the chip is rendered by the server and does **not** depend on how far you scrolled, so a half-scrolled list (20 ids, chip 45) is caught as stale instead of being accepted as a complete `20/20` audit. If they disagree, keep scrolling — don't "fix" it by editing the number down. Every id you list must then get its own `### <id> | … ` drill block in the report. Full field reference: `amazon-ads/references/audit-quickref.md` Step 1. **`total_active_source` is required, and for noon it is the chip reading** — `"chip:Live N"`, the number you read in step 1. Without it the combo is a `[基线]` gap. Reason: `total_active == len(active_ids)` only proves the two numbers agree, not that either was *observed* — trivially true when both come from the same parse (observed live: a run declared a 12-campaign marketplace as `4/4` because its script silently dropped files it couldn't read, and every check passed). The chip form isn't verifiable from disk the way Amazon's `"bulk:<file>.xlsx"` is, but writing the reading down turns an invented total from an omission into a claim the reviewer can check against the live page. Read the chip; don't back-fill it from the id count. ## 3. Campaign Detail Page **URL**: `admanager.noon.partners/en-{cc}/campaign/details/{campaign_id}?mpCode=noon&project=PRJ{project_id}` Campaign ID format: `C_{alphanumeric}` (10 alphanumerics after the underscore, e.g. `C_XXXXXXXXXX`). **Campaign ID extraction** from the campaigns list (IDs are in `<a href>` attributes, not visible text): ```bash browser-use <<'PY' print(js(""" var links = document.querySelectorAll('a[href*="/campaign/details/"]'); var data = []; links.forEach(function(l) { var m = l.href.match(/\\/campaign\\/details\\/([^?]+)/); if (m) data.push({name: l.textContent.trim(), id: m[1]}); }); return JSON.stringify(data); """)) PY ``` > ⚠️ **The link extraction above only returns the rows currently > rendered.** The list lazy-loads on inner scroll, so a raw read > captures ~15–20 of what may be many more. **Enumerate the full set > via the § 2 procedure** (read the `Live N` / `All N` chip totals, then > scroll the list container until the distinct link count matches). > Treat a single unscrolled read as a spot-check, never the full set — > under-counting here fails the completeness gate. **Campaign names can be misleading.** Verify actual products via the **Products tab** — do not trust the campaign name. A campaign named "mouse004 Auto" may target keyboard SKUs, not a mouse. Header shows: campaign name, Status badge, Budget, **Top-of-Search boost** (displayed as `Top Slot: N%` between Budget and Bidding Strategy for manual campaigns with TOS configured), Bidding Strategy, Running From date, Last Updated. **Top-right action icons:** a date-range picker, **pause/resume**, **duplicate**, and a round **blue pencil = Edit** (opens the campaign editor — see § 9 for adding/removing keywords & negatives). `target_filter` query param on the Targets tab switches the view: `target_filter=all` (positive keywords, default) vs `target_filter=negative` (negative keywords). **Brand Ads have different CTR/ROAS norms.** Brand Video ads measure view-through differently — never compare CTR directly to Product Ads. A `brand video` CTR of 0.23% is not "weak" vs product ad peers at 1.7–6.7%. Compare Brand Ads only against other Brand Ads, or against the brand ad's own historical ROAS. **Brand Ad "Creative" row.** The Products tab may show a "Creative" row (logo/video element) that accumulates clicks with zero attributed orders — noon can't attribute conversions to the creative. If Creative spend > 20% of campaign budget, flag it; calculate ROAS both with and without the Creative row for true product performance. KPI cards (same 10 metrics as overview, scoped to this campaign). Performance chart with metric toggles: ROAS, Revenue, Spends, eCPC, CTR, CvR, Orders, Clicks, Views, ATC. **4 Sub-tabs on campaign detail:** | Tab | Purpose | |-----|---------| | Products | SKUs in this campaign with per-SKU metrics | | Placements | Ad placement performance | | **Targets** | **Keywords** with match types and bids |