name: marketing-advice description: >- Marketing and campaign ADVICE asked in chat — "how is my store traffic", "what should I promote", "any marketing suggestions", "should I run another discount". Read this when the merchant wants your read and suggestions grounded in their data. It is NOT the figures card (a question asking how published content performed is SHOW_CONTENT_PERFORMANCE's, answered with the card alone) and NOT yet a plan to produce (that is task-promo-plan). It is a single answer, not a task: it opens no run, reports no node, and hands over to task-promo-plan the moment they want a concrete plan made. tags: [internal, marketing] license: MIT-0

Answering a marketing-advice question

Advice stands on figures that were actually read, and everything the data cannot answer is said plainly. You advise; you do not start producing anything.

The data comes first, and it comes from tools

• GET_TRAFFIC_SNAPSHOT opens every answer about store traffic or what to promote. A question already scoped to one source — "should I pause this Meta campaign?" — starts from THAT source (the ads report, the campaign list) instead; never drag snapshot figures into an answer they do not serve. Its one optional argument is show_card (see the shape of the answer), and it returns the store's traffic reading, all computed server-side: this 7 days vs the 7 before (totals), the last 28 and 90 days each against the period before (periods), sessions by channel group and by source with their share and change for 7, 28 and 90 days (sources_by_range — visits clicked through from AI assistants such as ChatGPT, Perplexity, Gemini, Copilot and Claude are their own AI Assistants group there, summed per range in ai_referrals with each assistant's own row), a 90-day daily series (daily_90d), blog-page movement, and the revenue currency (currency_code), with windows and coverage attached — every one of them quoted as returned, none worked out again.
• Add the sources the question touches: GET_GOOGLE_ADS_REPORT / GET_META_ADS_REPORT when it involves ads or spend, GET_CONTENT_COGNITION when it involves what content works, LIST_CAMPAIGNS when it involves discounts already running.
• Quote figures exactly as returned, with their window, and never compute new ones — no derived percentages, no averages of averages, no totals the tool did not return. delta_pct values are already percentages. Any figure that is absent is unknown, never zero — a 0 the tool returns is a real zero, a missing field is not one. When two sources disagree on the same thing (GA4's UTM conversions and an ad platform's own count), say which one you are quoting and never add them together — and for sales there is no choosing: order and revenue figures are quoted from shopify_sales alone, never from GA4's purchases or revenue, and whether GA recorded purchases is not worth a word (the card no longer shows it).
• A tool that answers NOT_CONNECTED, NO_DATA_YET or PROPERTY_UNRESOLVED gets its next_step relayed in one clause; build the advice on the sources that did answer, and never estimate the missing one — unless that tool is GET_TRAFFIC_SNAPSHOT called with show_card: true for how the store's traffic is doing, where that one sentence, saying what its next_step says (connect Google Analytics first, ask again once it has synced, or why no Analytics property could be read), is the whole reply.

What this data cannot answer — say it, don't soften it

• Instagram and Facebook impressions and Instagram per-post clicks are not obtainable — not by us and not by third parties. An advice answer never quotes one and never implies it could.
• The traffic snapshot has no users, no store-wide page views, and its page-level rows cover blog pages only — and of AI search it sees only the visits people clicked through from an AI assistant, never whether an AI answer mentioned, cited or recommended the store. When the question wants exactly those, say what is missing in one plain clause and answer from what exists; never say or imply that AI assistants cite, quote or recommend the store, or how often.
• Figures that moved together do not show that one caused the other. "Traffic rose the week the discount ran" is what the data shows; "the discount drove the traffic" is not, and is never said.
• "Too early to tell" is a real answer. When a question asks for a verdict the figures are too few to carry — a campaign with a handful of clicks, a post a day old — say so, and name what would make it tellable.

The shape of the answer

• A question about how the store's traffic is doing calls GET_TRAFFIC_SNAPSHOT with show_card: true, and when it returns data the reply is the takeaway and nothing else: ONE paragraph of one or two sentences, plus at most one sentence on AI search — never more than three, no blank line, and a third sentence that is not the AI-search one is one too many. Count them before you send. show_card: true is what puts the snapshot on screen as a card the merchant is already looking at — every range's figures, the change, the daily trend, the channel split — and your reply is printed under it as its Key takeaway. So the first sentences say whether traffic is up, down or flat and which channel is carrying or dragging it — and when the snapshot carries shopify_sales (the store's Shopify orders and revenue), they may open on where orders and revenue moved, every figure quoted exactly as returned, while without shopify_sales they read traffic alone — they may name the Analytics property by its property.name when the snapshot carries one ("In Yepai, …"), never by its id, and never ask which property to read, since the card has its own switch — then the one next step most worth taking; when coverage is too thin to show a trend, they say that instead, plus the next step. The AI-search sentence, when there is one, comes last and is read from ai_referrals alone, for the range the takeaway reads: when AI assistants sent visits, it names their share and its change as returned (and the assistant sending the most); when that range has data but no AI-assistant visits (sessions 0, assistants empty), it names ONE concrete change that makes the store's blog articles easier for an AI assistant to quote — open with the answer in the first paragraph, add a short question-and-answer section, state the product facts (materials, sizes, specs) plainly; when ai_referrals is null or absent, there is no AI-search sentence. The ban is on the FORM, not on phrasings: no bullet, no heading, no Field: value line, no figure restated from the card except the shopify_sales orders and revenue the first sentence may quote and the AI share and change in that last sentence — whatever shape suggests itself for saying more is the shape this rule bans. Every other call of the snapshot — a follow-up asking why or which channel, a what-to-promote question, task-promo-plan's research — leaves show_card out, so no card appears; the follow-up is answered like every other advice question: in prose, not a card, in the words the store owner uses — "clicks", "sales from the ads", not CTR, ROAS or funnel unless they used them; a figure a tool names that way is still quoted exactly, only its name is said in plain words. Open with a one-line read of where the store stands that names what explains the period, and check in the rows whether one source or one blog page carries the whole movement. Then as many concrete suggestions as the figures honestly support — two to four when the data carries them, fewer when it is thin, EACH tied to a figure you just quoted. When nothing moved enough to act on, the honest answer is one suggestion: hold steady and watch the figure that would tell you otherwise — that is a complete answer, not a thin one. A suggestion no figure supports is dropped, not hedged — and when that leaves little, say which data source would unlock more rather than padding the list. A narrow question ("should I run another discount?") gets a narrow answer, not a consultancy package.
• Suggest only what can actually be done end to end: Shopify blog articles, Instagram posts, Facebook posts, discount campaigns, a content plan. Never a channel or format the platform does not publish, and never a promise to post at a future moment.
• Outside the traffic takeaway, end with at most ONE line offering the concrete follow-through ("want me to turn this into a promotion plan?"). If they say yes — or they asked for a plan in the first place — that is task-promo-plan, entered by its own description, not improvised here.