A 1,097x Discrepancy - How I Ended Up Here (Part 1)

The platform said we made $82K from TikTok ads. The actual number was $75.

Not $75,000. Seventy-five dollars. A 1,097x discrepancy between what the advertising platform claimed and what the data showed. I’d been signing off on this spend every month because I trusted the reporting. The reporting was fiction.

That discovery changed how I think about every number in a business. But to understand why it mattered, you need to know who found it and how.

I co-founded Omnidesk, a premium standing desk brand in Singapore, in 2018. We grew it past eight figures in annual revenue, bootstrapped, no outside capital. At peak, we had showrooms, a full team, operations across Singapore and Australia, and a product line ranging from $699 entry-level desks to one-of-a-kind artisan resin pieces at $1,500+.

I came from SAP consulting. Contact center implementations first, then Hybris ecommerce. The work wired my brain for systems thinking, data flows, integration points. Advanced data management was my favourite subject at university. I’ve always needed clean data, structured properly, organized so it tells you something at a glance.

That background matters because it explains why the problems I’m about to describe bothered me so much, and why I had the instinct to go looking.

The frustration didn’t arrive all at once. It accumulated.

Processes that worked fine when the company was small hadn’t scaled. Reporting was manual and time-consuming. Sales data, stock movement, marketing performance: the information existed in our systems, but turning it into something useful required me to do it myself. The tools and processes we’d built during growth hadn’t evolved to match what the business actually needed.

The gap was clearest in advertising. We were spending real money on performance ads every month, but we didn’t have a reliable way to validate whether the platform-reported numbers matched reality. I could sense the discrepancies. Without the right tools to cross-reference the data, they were impossible to pin down.

Even basic operations had calcified. At tradeshows, the sheer number of manual processes made it feel like 2005, not 2025. The technology existed to do it differently, but the process had never been updated because it technically worked.

The root cause was always the same. I had the instinct to see what was wrong with the data, but not the bandwidth to fix it all myself. The questions I needed answered required pulling from multiple systems, cross-referencing, and interpreting. And the tools we had weren’t designed for that kind of work.

In January 2026, I started experimenting with AI. Not out of desperation. Out of curiosity.

The same month, I’d signed up for a public speaking course. A resolution to learn something new, to invest in myself outside the daily grind. The AI exploration had the same energy. Not a reaction, but a decision to stop doing things the way I’d always done them.

I’d been using ChatGPT on and off for a while. It was useful in a general sense. But it felt like I was using it wrong: asking questions and getting answers, like a slightly smarter search engine.

Then I found Anthropic’s Claude. Specifically, Claude Code.

The difference wasn’t subtle. It was like the gap between a $50 Bluetooth speaker and a proper audio setup. ChatGPT gave me answers. Claude gave me a workshop. I could feed it data, ask it to analyze, have it build tools, iterate in real time. It wasn’t a chatbot I was querying. It was a collaborator I was building with.

I started with the thing that had been frustrating me the most: the numbers.

The first audit was TikTok.

We’d been running TikTok ads as part of our performance marketing mix. The platform’s attribution dashboard showed roughly $82K SGD in revenue from our campaigns. That’s a solid number. If you looked at the dashboard alone, TikTok was a meaningful revenue channel.

I pulled our Shopify data and filtered for orders with TikTok UTM parameters. The actual tracked revenue was $75.

TikTok was claiming credit for revenue it had nothing to do with, using view-through attribution and broad matching to inflate numbers that had never been cross-referenced against actual sales data.

We’d been spending money on this channel every month based on numbers that were essentially made up. There was no process in place to validate platform attribution against actual orders. The platform reported a number, that number flowed into our reports, and I signed off because I trusted the process.

The process was broken.

The second audit was Klaviyo.

Our email marketing hadn’t evolved much, despite a few platform migrations over the years. The setup that worked when the list was small had stayed the same as the audience grew. Bulk campaigns sent to the full list. Same message, same timing, same approach. The list was being treated like a megaphone: louder equals better.

I used Claude to segment the list by actual behavior. Purchase history, browsing patterns, cart activity, engagement signals. What came back changed how I thought about email entirely.

Our “HOT abandoned cart segment”, people who had added items to their cart recently and shown high intent signals, was generating $3.30 in revenue per recipient. The bulk campaigns sent to the entire list were generating $0.10 per recipient.

That’s a 33x difference. We’d been diluting our best-performing audience with 28,000 unengaged profiles who hadn’t opened an email in months. Those profiles weren’t just dead weight. They were actively hurting deliverability, distorting metrics, and costing us money every month just to store.

We suppressed the unengaged profiles, retired the static lists, built new behavioral segments, and created exclusion templates so the high-intent audiences never got buried under bulk sends again.

The entire Klaviyo infrastructure rebuild took a weekend.

I want to be clear about what happened here, because the lesson isn’t “AI is magic.”

The lesson is that I’d been running a company for seven years with data infrastructure I never questioned. I trusted the platforms to report accurately. I trusted the processes we’d built during growth to still be working all the time.

None of that trust was earned. It was just inertia.

What Claude gave me wasn’t intelligence I didn’t have. It gave me speed. The ability to pull data from multiple sources, cross-reference it, and surface discrepancies in hours instead of weeks. I could have found all of this manually. But it would have taken so long that I never would have started.

One person with domain knowledge and the right AI tools could do in a weekend what hadn’t been done in years. Not because the team was incompetent. Because the tools they had weren’t designed for the questions I was asking. And they didn’t have the context to know which questions to ask.

I didn’t need better people. I needed better tools.

The free section told you what I found. This section shows you how to find it yourself. Every audit below was done by one person (me) using Claude, with no custom integrations and no dedicated analytics team. If you can export a CSV, you can do this.

The principle is simple: never trust platform-reported revenue. Always cross-reference against your source of truth (for ecommerce, that’s usually Shopify or your payment processor).

The process: 1. Export your ad platform’s attribution report for the last 90 days. Every platform: Meta, TikTok, Google, whatever you’re running. 2. Export your Shopify orders for the same period, including UTM parameters. 3. Feed both into Claude and ask it to match attributed revenue against actual tracked orders. 4. Calculate the discrepancy ratio for each platform.

What you’ll likely find: Meta and Google will be inflated but in a believable range (2-5x is common with view-through attribution). TikTok and newer platforms tend to be wildly inflated. The gap tells you which channels are actually driving revenue and which are taking credit for revenue that would have happened anyway.

The decision framework: Any channel with a discrepancy above 10x deserves immediate scrutiny. Above 50x, you’re probably funding fiction. Cut it and see if revenue actually drops. (Ours didn’t.)

Most ecommerce email lists are treated as a single audience. They’re not. Your list contains at least four distinct groups, and treating them the same is leaving money on the table while actively hurting your deliverability.

The segments that matter: 1. HOT (high intent): Added to cart in the last 7-14 days, browsed multiple product pages, opened recent emails. These are your highest-value recipients. 2. WARM (engaged): Opened or clicked in the last 30-60 days, previous purchasers, active browsers. Solid audience for campaigns. 3. COOL (fading): No engagement in 60-120 days. Worth a re-engagement attempt, but don’t send them everything. 4. COLD (dead weight): No opens, no clicks, no site visits in 120+ days. Suppress them. They’re hurting your sender reputation and costing you money.

The overhaul process: 1. Export your email platform’s engagement data (opens, clicks, purchase history, last activity date). 2. Use Claude to segment profiles into the four categories above. 3. Suppress the COLD segment immediately. Calculate how much you save monthly. 4. Build separate flows and campaigns for HOT and WARM. Different messaging, different frequency, different offers. 5. Create exclusion templates so bulk campaigns never override your targeted segments.

Our results: HOT segment at $3.30 RPR vs $0.10 for bulk. That gap is probably sitting in your list right now.

This one is about your Meta Ads (or whatever your primary paid channel is), and specifically about ROAS trends over time.

Here’s what the advertising frustration actually looked like from the inside: weekly reports that said roughly the same thing every week. Same recommendations, same format, same conclusions. ROAS numbers that never matched what I could see in Shopify. No clear triggers for when to scale spend or cut it. We were spending significant money monthly with no reliable way to know if it was working.

What to look for: 1. Pull your ROAS by month for the last 12-18 months. Our trajectory went from 8-10x in 2024 to 3-6x by late 2025. 2. Break it down by campaign type. Prospecting vs retargeting. Brand vs conversion. The averages hide the story. 3. Check for audience overlap across ad sets. This is where spend gets wasted silently. Multiple campaigns bidding against each other for the same people. 4. Compare platform-reported ROAS against your actual Shopify revenue (same methodology as Step 1).

The uncomfortable finding: View-through attribution was inflating our Meta ROAS significantly. The platform counted a conversion if someone saw an ad and purchased within a window, even if they never clicked. When we stripped view-through and looked at click-through attribution only, the numbers told a different story.

Across the three audits above, we identified roughly $1.23 million in annual cost recovery. That’s not new revenue. It’s money we were already spending that was either wasted (TikTok), leaked (unengaged email profiles), or misallocated (overlapping ad audiences).

The breakdown: - Channels with fictional attribution: cut entirely, redirect budget to proven channels - Email list cleanup: immediate savings on platform costs, plus improved deliverability driving higher revenue per send - Ad spend reallocation: same total budget, but concentrated on campaigns with verified ROAS

Most businesses have this kind of number hiding in their operations. They just don’t have the tools or the time to find it.

You don’t need a data team. You don’t need custom dashboards. You don’t need a six-month analytics project. You need one person who knows the business well enough to ask the right questions, and an AI tool that can process the data fast enough to answer them.

The person with the deepest operational context is almost always the founder. And until recently, that person didn’t have the tools to act on that context at speed.

Now they do.

Here’s what I keep thinking about: how many founders are signing off on numbers right now that nobody has cross-referenced against reality? Not because they’re careless. Because the tools to check were never fast enough to make it worth doing.

What’s the number in your business that everyone trusts but nobody has actually verified?

Next week: Part 2, “The Business AI Was Built to Fix.” To understand why those audit results mattered, you need to understand what I was auditing.