Product case study — Paula Ward

Competitive intel
that maintains itself.

CompeteIQ is a system I designed and built end-to-end — from PRD to a running production pipeline — that monitors competitors on a cadence, synthesizes what changed using Claude, and keeps a living artifact current with a human always in the loop.

A quick note on what this is: it isn't a slick battlecard for GTM teams — though they're welcome to reference what's inside it. The primary audience is the product team, and the real point is showing how I, as a product leader, use AI to design and build a workflow solution instead of just asking for one. What follows walks through both the process of building it and the system itself. Total time to build: less than a day.

01 — The Problem

Competitive intel is important and nobody keeps up with it.

Every product team agrees competitive awareness matters. Almost none of them have a reliable system for it — just a Google Doc someone updated once, three quarters ago.

CompeteIQ is a recreation of a problem I lived in a previous role. Not every company has a dedicated competitive intelligence team, or budget for an off-the-shelf tool — mine didn't. Keeping up with competitive movement was already a struggle for a product team with actual product work to do. It got harder when we expanded into new markets: new markets meant new competitors, and the ad hoc approach we'd been getting away with stopped being good enough.

PRD v0.3 — Problem Statement

Product teams struggle to maintain consistent visibility into competitive movement. Competitive reviews are time-consuming, inconsistently executed, and typically reactive rather than proactive. By the time a team synthesizes a competitor update, the signal is stale and the window for action has narrowed.

I've owned those documents. In the worst cases, maybe they're updated once or twice a year — during annual planning, or when building a business case. Usually requiring a massive overhaul from a blank slate. They're irrelevant within days. The worst is hearing about competitor news a month late.
The default state

Manual Googling, ad hoc Slack shares, or an analyst tool that still dumps the synthesis work back on you.

The real cost

Not the research — the upkeep. Someone has to remember to check five sources, every cycle, forever, with no reminder when it slips.

The failure mode

No living record of current competitor state — so every review starts from zero instead of from "here's what changed."

02 — The Process

I didn't start with a tool. I started with a document.

The build is the least interesting part of this story. The decisions underneath it are the part that's actually PM work — so before the architecture, here's how I got there.

PRD excerpt: Problem Statement and Opportunity
PRD_v0.3 — Problem Statement & Opportunity, before a single ticket existed
Ticket backlog, Epic 1: Setup and Configuration
Ticket backlog, Epic 1 — nine pages, written before any code

The PRD isn't pictured as a finished document — it went through three full rewrites before a single ticket existed. Each pass cut ambiguity that would otherwise have turned into back-and-forth during the build.

Four decisions that mattered more than any line of code

DECISION 01

A spreadsheet, not a UI

The system needed a source of truth for which competitors to track. Building an admin UI would have taken a sprint nobody asked for.

From the ticket: "Google Sheet config approach keeps this accessible without a UI."
DECISION 02

Baseline before monitoring

It's tempting to jump straight to "detect changes." But you can't detect a change against nothing — the baseline had to ship first, even though it felt like just setup work.

From the ticket: "Critical dependency for FR-03: Claude needs a baseline to compare against."
DECISION 03

Human-in-the-loop, not autopilot

An AI that silently rewrites a shared doc is a system nobody trusts for long. Every signal needed a person to say yes before it became permanent.

From the PRD: "...a human-in-the-loop review step to accept or reject changes into the shared artifact."
DECISION 04

Bi-weekly, not weekly

Weekly runs produced too little real movement per digest and started training people to skim past the email. Bi-weekly gave each run enough signal to be worth five minutes.

From the Zap itself: literally named "CompeteIQ – Bi-Weekly Run"
✦

Claude was a thought partner through all of it, not just a code generator — scoping the PRD, stress-testing the weekly-vs-bi-weekly cadence call, arguing through whether the artifact belonged in Drive or a purpose-built repo (Drive won: competitive intel dies the moment it lives somewhere people don't already look), and reasoning through what "human-in-the-loop" should actually gate. The decisions above are mine — but I made better ones faster with a collaborator who'd push back.

03 — The Architecture

Five tools, one loop, one human checkpoint.

Google Sheets
Google Apps Script
Zapier
Claude (Sonnet)
Gmail
Google Drive / Docs

Config lives in Sheets. Orchestration and scraping run in Apps Script. Zapier is the scheduler. Claude does synthesis, scoring, and baseline generation. Gmail is the interface. Drive is the system of record.

I was already familiar with Zapier, Claude, and Google Docs & Sheets — Google Apps Script was the new piece. I debated the orchestration layer with Claude, then hit a wall: building parts of the automation entirely in Zapier would have meant upgrading off the free tier, and this project had a $0 budget. Claude suggested Google Apps Script instead — free, and it integrated cleanly with the Google tools already in play.
ZAPIER Every 14 days COMPETITOR CONFIG Google Sheet GOOGLE APPS SCRIPT PIPELINE Scrapes 6 sources CLAUDE Synthesizes and compares to baseline GMAIL DIGEST Scored & grouped HUMAN IN THE LOOP Email Accept / Reject accept always logged ARTIFACT UPDATED Google Doc LOG FILE Google Sheet
Trigger & config Collection & synthesis Human decision Persisted state
04 — The Build, Walked Through

What actually happens every other Monday

Real screenshots and a real recording from the running system — not mockups. As I built the Apps Script pipeline, I tested at every step along the way, in the execution log, before trusting it to run unattended.

1

Configure once

Up to ten competitors, five URLs each, in a Sheet. Adding or dropping a competitor is a row edit — no deploy, no ticket to me.

Competitor config Google Sheet
CompeteIQ_Config — Competitor Config tab
2

Zapier fires the run

A scheduled trigger kicks off the pipeline every other Monday, reads the config, and hands rows to the pipeline step by step — no server to babysit. The same Zap can also be triggered ad hoc, so I'm not locked into the cadence when something — a launch, a board prep, a rumor — means I need fresher signal sooner.

Zapier Zap overview, CompeteIQ Bi-Weekly Run
CompeteIQ – Bi-Weekly Run, Zap overview
3

The pipeline scrapes and Claude synthesizes

Apps Script pulls each competitor's site, blog, careers page, G2 and Capterra profiles, then hands the raw pull to Claude alongside the existing artifact — asking it to find what's actually new.

Apps Script pipeline with execution log
Execution log — 2 signals collected for Applause, 4 confirmed after synthesis
4

A digest lands, not a dashboard

Every signal is scored for relevance and grouped by competitor. No login, no tool to check — it shows up where product teams already work: their inbox.

Signal digest email
Signal digest — 20 signals flagged across 5 competitors, 0 scraper errors
5

One click, and it's real

Accept writes the signal straight into the competitor's artifact in Drive. Reject logs it to a feedback sheet instead of just disappearing — so the system has a record of what didn't matter, too.

● LIVE CAPTURE
Actual accept flow, recorded from the running system
05 — The Artifacts

What the system actually produces

Not a demo environment — the real output from a real run.

The competitor artifact, before and after

Accepting one signal from the digest above updates the News & Trends section of Applause's artifact automatically — timestamped and attributed to its source.

Before
Competitor artifact before accepted signal
After one Accept
Competitor artifact after accepted signal
What changed A single accepted line — "[9/14/2026] October 2026 — CEO Aatish Salvi publishes first major thought leadership piece..." — inserted into Section 5 (News & Trends), dated and sourced, with zero manual formatting.
06 — What's Next

Where I'd take it from here

Distribution

Slack delivery option

Email works for a solo PM. A team rollout needs the digest to land where the team already triages work.

Learning loop

Feed rejections back into scoring

Every Reject is already logged to a feedback sheet. It's currently a record — it should become a training signal for relevance scoring.

Coverage

Add pricing-page diffing

News, hiring, and review sentiment are covered. Pricing page changes are some of the highest-signal, lowest-noise competitive moves — and currently not monitored.

Data

Pull in internal signals

Right now every source is public. The more complete synthesis brings in internal tools and data too — win/loss notes from CRM, for one.

Artifact quality

Update in context, not just append

Accepted signals currently get appended to the relevant section. The better version reasons about the artifact as a whole and integrates the update cleanly into existing context — not just tacks it on.