Merchandising trend research inside an AI assistant

Killed my Google Trends scraper the same afternoon. Clean payloads, low latency, honestly better than I expected on free.

Keyword trend reports for clients. Free tier forces us to batch for weekly digests, which turned out fine. Upgrade path exists when we outgrow it.

~20 min to a working mcp hook. json shape is sane, docs match what ships

We pipe weekly series into BigQuery for a few brand cohorts. The old Selenium path was mostly babysitting selectors when Google moved something. This is boring in the best way. Uptime's been solid.

When leadership asks why we're shipping something, these screenshots land. Wish the dashboard had saved views; API's the strong half.

Cursor + MCP, one try. Not a trends person; my side project emails me when a niche term spikes week over week and that's all I wanted.

Growth endpoints before I write up notes. Normalization isn't always 1:1 with the raw Google UI (fine), but it's consistent run to run.

Multi-source ranked lists into a notebook, straightforward. When I typo a param the error payload is readable. Sounds small. It isn't.

Does what it says. I knocked a star because onboarding assumed I already knew MCP wiring; a copy-paste block for Claude Desktop would have saved me 15 minutes.

tiktok momentum vs paid spend in one looker sheet. unglamorous tbh — first rollout we didn't fight over about halfway through…

Retries: predictable. No random HTML in responses (scrapers, I'm looking at you). Team key rotation would be nice; we rotate keys manually for now.

Retail buzz before filings. Faster than twelve tabs and a spreadsheet nobody trusts.

Meme or real trend? One call usually tells me. Still spot-check Search Console when it matters.

Workshops: how do you ground LLMs in something fresher than training data? I show this. Engineers who hate glue code actually nod at the MCP bit.

Client reporting. Finance finally stopped asking which line item this maps to. Peak hours can drag a little. Acceptable.

I wired this behind a small CLI for contributors who want trend context in issues. Keeping the surface area tiny matters for OSS, and the schema has not churned on me yet.

30-day window, daily pull, internal scoreboard. The screenshot-from-Trends fight in Slack basically ended.

We are pre-revenue, so free tier discipline matters. I hit the cap once during a brainstorm where everyone wanted to try random keywords. Learned to batch smarter.

Security review: HTTPS, scoped keys, no sketchy redirect hops we could find. Sounds basic. It isn't, in this category.

I do not need this daily, but when App Store rank shifts look weird, having Reddit and news context in one place saves me from context switching across six apps.

Blowing up for real vs loud on one platform. Doesn't replace reporting. Keeps my ledes from lying.

playwright job died every time google sneezed. same shape every week now.

Seasonal demand spikes line up with what we see in Amazon search interest here. Merch team stopped sending me screenshots from random tools that never matched.

Decks: fine. I still export to Sheets by hand so minus one star. Direct connector someday maybe.

Steam concurrents plus Reddit chatter in one workflow beats our old spreadsheet ritual before milestone reviews.

Quick pulse on whether a feature name is confusing people in search before we ship copy. Cheap sanity check compared to a full survey.

Monitored from Grafana via a thin wrapper. p95 stayed under our SLO budget last month. One noisy day during a holiday but nothing alarming.

Narrative fights in meetings got shorter once we could point at the same trend line everyone agreed on. Sounds silly until you have lived through it.

Using normalized series as a weak prior in a forecasting experiment. Citation-friendly timestamps in the payload made reproducing runs less painful.

Approved for our pilot group after a quick vendor review. Would love SAML, not a blocker for our size.

YouTube interest + TikTok hashtags in one view. Makes sponsor conversations less hand-wavy. I can point at something.

Cron before standup, Slack gets a blurb. Afternoon to wire. Two quarters without drama.

Public-interest stuff: search interest is a rough attention proxy. I still hit primary sources. One signal among several, not the whole story.

Runs in a VPC egress-only subnet with allowlisted domains. Fewer exceptions to explain to auditors than our last vendor.

Spotting when a topic is about to flood Discord saves my team from reactive moderation fires. Not perfect, but directionally right often enough.

Lean team. ROI isn't subtle. I wouldn't rebuild our old scraper unless legal made us.

Examples in the docs match what the MCP actually returns. You would be surprised how rare that is in this category.

Pager stayed quiet. When something upstream flaked once, the error string told me which parameter to fix without opening logs first.

Students use it for coursework demos. Budget is tight so free tier matters; we coach them to cache aggressively.

Helps prep talking points when retail interest in our name swings after earnings. Not material disclosure, just context for Q&A prep.

Rolled into our internal CLI. On-call hasn't paged for this integration once.

Good first pass before we pull filings. Not a substitute for fundamentals, just faster triage.

claude reads the json, i fix the content. split workflow that actually works

Mostly lines up with what we hear on weekly sales calls. I still check inventory before I trust it for a buy.

Pulled a trend summary into ChatGPT for ad copy angles. Corny but it saved a two-hour brainstorm.

Client asked for a trend slide Thursday 5pm. This existed so I didn't have to fake it.

Windsurf picked up the MCP manifest without me hand-editing JSON. Small win, I'll take it.

Retail attention vs what management said on the call. Useful tension to surface before the write-up.

We stopped exporting from five different tools. One chart now goes in the Monday deck and nobody argues about whose export is newer.

Webhook → trend pull → Slack. Boring pipeline. That's the whole compliment.

Copilot suggested a call pattern that matched your error schema. Saved me a round trip to the docs.

Solid API. Dinged for dashboard polish. I'm not an engineer, I live in the UI half the day.

Serde-friendly enough I didn't write a custom deserializer. High praise from a grumpy systems person.

Response sizes stay small enough for mobile hotspots. I hate APIs that dump megabytes for a sparkline.