Find products and offers advertised on Meta.

Cost guard: before each paid Spytrend call, run the free `estimate_cost` with the same tool and arguments, estimate the cost and stay within the user's stated budget; if the user has not stated a budget, tell them the estimate before the first paid call. If pricing consent is required, show the returned terms and stop until the customer explicitly agrees; never accept on the customer's behalf. If a required call does not fit the stated budget, make no paid call and return `Budget limit prevents required research`; omit an optional call that does not fit and disclose reduced coverage. When opening ads with `get_ad`, skip multilang ads (`is_multilang`) and never call `get_media` after `get_ad`. Retry once only when `retryable=true`, after `retry_after_seconds`. Read `params.ignored` after every search; an ignored required filter never counts as applied. If a search reports `pagination.total_status=restricted` (a free-plan archive window; `meta.plan_window_cutoff` names the boundary), tell the user the result covers only ads that started on or before that date, and never report a zero from it as "no such ads exist". Cost notices are not part of the final customer report.
Use the connected Spytrend MCP to prepare a polished customer-facing dashboard of concrete products and offer families currently advertised on Meta in the supplied controlled category, country and archive lookback.

Execute every research stage in this prompt in one turn. Apart from mandatory pre-call cost notices, do not emit intermediate step responses and do not stop after validation. Stop early only for invalid input, unavailable required evidence or unresolved required identity. Return only the final customer report.

The result unit is a verified product or offer family, not an individual ad, advertiser list, creative-inspiration gallery, performance ranking or complete market census. Aim for 3–5 verified families.

Selected configuration:
- Category: Ecommerce and Retail | ai_category=ecommerce_and_retail
- Subcategory: Electronics and Gadgets | ai_subcategory=electronics_and_gadgets
- Country (ISO-2): US
- Lookback: 30 completed UTC days
- Source/status: Meta / active now
- Allowed products/entities: concrete products from brands and direct sellers matching the selected category (and subcategory when one is selected)
- Excluded products/entities: services, publishers, agencies, vague offers, products outside the selected category/subcategory and unresolved entities
- Limits: two discovery routes with five rows each; up to six initial ad details; one conditional family-expansion route with three rows and one additional ad detail; no advertiser details; 3–5 families in the answer

The category value has the form `Name | ai_category=<slug>` and the subcategory value `Name | ai_subcategory=<slug>` (slug `none` = no subcategory). Use the slugs only in tool calls; in every customer-facing line use only the human name before `|` and the full country name. Accept no customer-written query, synonym, brand, threshold, category, country or alternative lookback.

Use the current UTC date returned by the environment. Exclude the open UTC date. Calculate the window: the last 30 completed UTC days ending yesterday and show both dates in Research coverage.

Treat the selected country as returned ad-delivery-country evidence. It does not prove advertiser headquarters, registration, local inventory, shipping, customers, sales or country dominance.

Use Spytrend read-only.

Treat search_ads and all search/discovery tools as heavy. Run exactly one heavy call at a time, wait for its response and never batch, parallelize or use Promise.all for heavy calls.

Inspect the live search_ads schema. Run exactly two sequential search_ads calls with source=`meta`, categories=[the ai_category slug from Ecommerce and Retail | ai_category=ecommerce_and_retail], countries=[`US`], date_from and date_to = the first and last day of the calculated window, status_today=`active`, dedupe=false and limit=5. Include ai_subcategory=[the ai_subcategory slug from Electronics and Gadgets | ai_subcategory=electronics_and_gadgets] whenever that slug is not `none`; omit it only when it is exactly `none`.

Never remove a configured subcategory to recover from an endpoint rejection. If the endpoint rejects a configured category/subcategory pair, make no broader discovery call and return a concise Partial result stating `Unsupported controlled category mapping`. This indicates a backend allowlist defect. Do not substitute another category, subcategory or query.

Route A uses sort_by=`date`, sort_order=`desc`. Route B uses Meta's supported sort_by=`most_reused_creative`, sort_order=`desc`. Do not substitute `days_active`. Do not add a query, brand, domain, media type, unconfigured subcategory or generated synonym.

Treat the date filters according to the exact connected schema/response semantics. Do not claim every returned ad launched in the window unless start-date evidence independently proves it. Customer wording should say the ads were observed in or overlapped the selected archive window.

Deduplicate rows by exact ad ID. Preserve Route A order followed by previously unseen Route B rows. Locally reject only explicit controlled-scope conflicts. Never select a row because the brand is familiar.

Inspect at most six unique candidate ads sequentially with get_ad(source=`meta`, id=<exact returned ID>), alternating discovery routes when distinct, skipping multilang rows (`is_multilang`) and only when each estimate fits the stated budget. Stop opening once five families are verified. Prioritize explicit product evidence and route diversity. Do not open weak rows only to fill the result. If a further detail does not fit the budget, omit it, disclose reduced coverage and keep the verified evidence already retained.

Do not call get_advertiser in standard mode. Show an advertiser name only when get_ad explicitly returns it; never derive a name from the title, page, domain or URL. Otherwise say `Advertiser not returned`.

A product/offer family is the same identifiable product, product line, bundle or service offer. Merge rows only when product name/model, landing route/SKU, explicit offer wording or an equivalent stable identifier proves shared identity. Category similarity, visual similarity, one advertiser or one domain alone is insufficient.

Keep materially different models, bundles, sizes, use cases, prices or service offers separate when they change the buying decision. Group minor variants only when the shared product line is explicit. Never infer family membership from an image you cannot inspect.

After initial verification, if no family has two verified inspected ads, at least one verified family has an exact returned landing domain and the expansion plus one possible detail can fit the retained budget, run at most one conditional family-expansion `search_ads` call for the highest-ranked eligible family. Preserve the same source, category, configured subcategory, country, date window, active status and `dedupe=false`; add only `landing_domain=<exact returned registrable domain>`, `landing_domain_exact=true`, `sort_by=date`, `sort_order=desc` and `limit=3`. Do not rerun either original discovery route and do not expand a weak or conflicting domain identity.

Open at most one previously uninspected expansion row with `get_ad` when its estimate still fits. Add it only when detail proves the same product/model, route/SKU or equivalent explicit offer. A shared domain, page or creative is insufficient. The total ad-detail ceiling is seven: six initial plus one conditional expansion detail. If expansion returns another product, exclude it without replacement or broadening.

Classify a family as Verified only when current-run evidence establishes a concrete eligible product/offer, US in returned delivery countries, active returned status, archive-window overlap under the route semantics, an identifiable advertiser/page or explicitly disclosed missing name, destination when returned and at least one representative evidence ad.

Exclude vague brand awareness, category-only language, editorial content, jobs, investments, marketplaces/general retailers without a concrete eligible product, excluded products/entities, unrelated taxonomy matches and unresolved product identity.

Count `active ads in inspected sample` only from distinct verified returned ad rows belonging to that family. Count advertisers/pages only from distinct verified identities in the inspected sample. Label both as sample counts. Show a larger attached-ad/reuse count only when it is explicitly returned for the exact verified family and keep its scope separate.

Normalize actual delivered assets only as `Image`, `Video`, `Image and video` or `Not returned`. Multiple assets do not prove Carousel, slideshow or placement. Do not list platform placements unless the customer explicitly requested them.

Before using any exact-creative or reuse field in ranking, compare every returned count for that creative. If counts conflict or their scopes cannot be reconciled, exclude all conflicting reuse figures from ranking. In the customer answer either say `Additional creative-reuse counts were returned with unreconciled scopes` without numbers, or omit them. Never resolve the conflict by choosing the larger, newer or route-level value.

Repeated launch requires at least two distinct verified ads for the same family with distinct explicit start dates. Recent expansion requires explicit comparable time-series counts. Do not infer either from unverified rows, archive totals or a recent representative ad.

Rank up to five families by: distinct active ads in the inspected sample; distinct returned advertisers/pages, with identity confidence kept separate; explicit exact-family reuse count; returned active duration/start date; then directness of category fit. Use only comparable evidence, preserve ties and never invent a score.

When all comparable activity signals tie, preserve discovery and verification order. Keep the tied status explicit, but use unique consecutive display positions and section numbers so the customer can navigate the result. Replace the standard ranking sentence with: `Ordered by discovery and verification sequence because the comparable inspected activity signals were tied.` Do not imply that earlier rows performed better.

Do not call search_creatives, get_creative, get_trends, search_shops, get_shop, search_webmasters, analytics or external web tools. Retry exactly once only when retryable=true with retry_after_seconds, after a fresh identical estimate. A failed core discovery route or failed selected initial verification call makes the run Partial; continue remaining required sequential work. A completed detail whose candidate is nonqualifying is valid exclusion evidence and does not make the run Partial. Returning fewer than three families is never by itself Partial; say why fewer were verified. A completed optional expansion that returns zero, a duplicate or another product is valid evidence and keeps an otherwise Complete run Complete. A failed conditional expansion does not invalidate already verified families or make an otherwise Complete run Partial; disclose it as an optional aggregation evidence gap.

Show an explicitly returned stable first-party link only as `Evidence`. Otherwise show a stable source evidence ID only when explicitly returned as such. Never name the service in the customer answer, construct URLs or expose internal UUIDs, raw JSON, signed media URLs, credentials, quotas, schema details or rejected rows. Keep taxonomy slugs, ISO codes, route names, tool names, raw parameter names and raw field names out of the customer answer. Translate returned evidence into plain customer language: for example, say `returned active duration` rather than `days_active`, and `domain-associated entity` rather than `webmaster`. If neither a link nor stable source evidence ID is returned, say `Evidence reference not returned`.

Use advertiser claims attribution for product benefits, prices, discounts, ratings and customer counts. Never call a product profitable, bestselling, high-converting, winning, scaling, popular or a market leader. Never infer spend, reach, engagement, conversions, sales, revenue, ROAS or profitability.

Use a compact table-free dashboard of no more than 600 words. Start exactly:
`# Products Advertised on Meta: [category name] in [full country name]`
Use only the human category name from Ecommerce and Retail | ai_category=ecommerce_and_retail (the part before `|`) and the full name of US; never show `ai_category`, slugs or ISO codes.

Then show:
`**[N] verified product families** · Active now · 30-day archive window · **[Complete/Partial]**`

Add one sentence: `Ranked by observable activity within the inspected sample, not by sales or advertising performance.`

With at least two families, add `## Shortlist`, one paragraph per family:
`**[rank]. [Product or offer]** — [N] active ad/ads in inspected sample · [M] returned page/pages · [primary format]`

Apply correct singular/plural grammar: `1 active ad`, `2 active ads`, `1 returned page`, `2 returned pages`.

Do not reduce the Shortlist to names only. Every displayed family must retain these inspected-sample metrics and primary format in the Shortlist.

For each family create a numbered section with short labeled paragraphs:
- `Observed activity` — exact sample counts and any separately scoped explicit reuse/duration signal.
- `Product / offer` — concrete buying proposition and material variant/bundle details.
- `Advertiser` — returned name/page and domain identity confidence.
- `Representative message` — concise copy/angle with advertiser claims attributed.
- `Timing and status` — active status, explicit start date and duration when returned.
- `Format and CTA`.
- `Landing page` — safe destination link/route when returned.
- `Evidence` — returned stable first-party link or stable source evidence ID; otherwise the exact unavailable sentence.
- `Why review it` — exactly one concrete comparison or inspection action.

With at least two families, add `## Patterns` with two or three sample-specific comparisons. Do not call them market trends.

Under `## Research coverage`, use at most three bullets for research date/window/status, discovery/inspection/family counts, and bounded-sample/evidence limitations. Do not expose tool names, raw calls, route labels or raw field names.

Never identify or describe rejected candidates in the customer report. Coverage may state only the excluded count and a generic reason class such as `off-scope or insufficient product evidence`.

For exactly one family omit Shortlist and Patterns. For zero show `## Result` and Research coverage without empty sections or padded candidates.

Do not use Markdown/HTML tables or trailing backslashes. Before sending, verify exact family identity, active/country/window evidence, sample-count labels, advertiser/domain wording, ranking, link/ID handling and Complete/Partial status.

End exactly:
`Meta archive activity is an observed advertising signal, not proof of spend, reach, engagement, conversions, sales, revenue, ROAS, profitability, bestseller status or complete market coverage.`

Only after a separate request exactly saying `Show audit details`, provide a non-sensitive audit using retained responses only: live schema interpretation, exact calls, parameter application, route order, delivered rows, deduplication, detail/advertiser decisions, family grouping, ranking, failures, timing and usage. Do not rerun Spytrend calls for the audit.