Your Shopify Store Is Bleeding Money Every Second. Where Is the Leak? - Sidharth Sahni, Helium

Someone is in your Shopify store right now. They clicked your meta ad. They landed on a page. They left. No purchase. No email captured. Nothing. Most founders look at that and think "bounce rate, normal." Then they go buy more traffic.

Sidharth Sahni thinks that's one of the most expensive mistakes a D2C brand makes. He's the co-founder and CTO of Helium, a platform that watches what happens inside every session of your Shopify store and uses AI to adapt the storefront in real time for each visitor.

Helium sits in the middle of the stack. Recommendations, dynamic merchandising, attribution tracking, a custom pixel that fingerprints devices instead of users, and a root cause analysis engine that traces revenue down to a single homepage banner. Sidharth and his co-founder Shrey started Helium in 2023, are backed by Mirae Asset Ventures, and Sidharth is on the Forbes 30 Under 30 Asia 2025 list.

This conversation goes deep into the actual stack of personalization on Shopify. Why PDP and collection pages matter and homepage barely does. Why ROAS is the wrong number to optimize for. Why the most ignored leak in any Shopify store is at the catalog level on Meta itself, before anyone even lands on the website.

If you're running a Shopify store with sub 2,000 SKUs and you've ever wondered why optimization feels like throwing darts in the dark, this one's for you.

Key Takeaways: The Real Anatomy of a Shopify Leak

The Optimization Hierarchy:

• PDP is the most important real estate, collection page is second, homepage is the worst place to optimize because only 5% of traffic lands there
• Most brands optimize the third fold of the homepage, which sees 0.2% of total traffic, then wonder why nothing changes
• The leak often starts before the website, on Meta itself, when catalog ads point to out-of-stock variants

Why ROAS Is the Wrong North Star:

• ROAS is an outcome metric, not an input metric, you can't optimize for it directly
• Most brands focus on ROAS because it's the only number Meta gives them they're confident is correct
• Real optimization requires breaking ROAS into sub-metrics like traffic quality, page-level conversion rate, and visit-count cohorts

What Brands Are Blind To:

• They don't know if changes they make actually moved the needle, attribution at the banner-level barely exists in standard tools
• They optimize for low-traffic real estate while ignoring the pages where 80% of revenue happens
• They tend to chase quantity of traffic instead of quality of traffic, which is far easier to fix

Q: How did Helium actually start?

Sidharth: I've loved building since school. Made websites, weird animations, a galaxy generator in three days, weird projects, 80% of which I cannot show anyone. I still build for fun. I make personalized games as gifts for close friends. I took part in 20 to 50 hackathons through college, won a lot of national and international ones.

Then I got to know Shrey. We were working with an app enabler that builds apps for D2C brands, and we both got stuck on the same idea. We thought, app is fine, that's a retention channel, but website has so much opportunity. Website is so static and bland right now, and AI was starting to grow. So we said let's jump the gun. That's where Helium came around almost three years ago.

The name came from Shopify's framework Hydrogen, which is built on top of React for headless storefronts. We saw it as one-upping them. We started building a web builder on top of the modern stack and realized the bigger problem was that liquid is slow and hard to change, and that the same page everybody sees is broken. People coming from this ad should see something from this ad. People coming from X location should see something they can relate to. Those were the two notions we started with. Then the personalization rabbit hole pulled us deeper.

🔥 ChaiNet's Hot Take: Helium started by trying to one-up Hydrogen. They ended up in the personalization rabbit hole. That's the pattern with most useful infra companies. The original idea isn't where they end up, but the original idea was the only way to discover what the real problem was.

Q: Without Helium, what actually happens when someone clicks a meta ad and lands on a Shopify store?

Sidharth: It starts even before the website. Catalog ads, which most big brands run, are 50 product carousels. As soon as somebody clicks, especially on apparel, the product is often out of stock or only one variant is available. Just XS, just large. Five of the variants are wrong. Nobody converts because their size isn't there. So before even reaching the website, it starts at the meta level: are the products you're showing even sellable?

When you land, on the PDP, the user might like it 60-70%. Most people, especially in apparel, scroll down looking at recommendations. Without Helium, the PDP is static, just whatever you've kept. Most likely "recently viewed" or basic Shopify-native recommendations. On the collection page, it's a manually-ordered list saying find what you're looking for.

For a large catalog brand, on a kurta collection of 2,000 products, what do you even show, and to whom? With Helium, within half a second of arriving, we fingerprint the device. Not GDPR-violating tracking, just the device. We say, OK, this is from a newer iPhone, a good tier device, coming from Mumbai or Delhi, in the afternoon, on this collection. Then we reorder based on what people in this cohort tend to prefer. As you scroll, it dynamically changes again.

🔥 ChaiNet's Hot Take: The biggest leak in your Shopify store isn't on your Shopify store. It's the moment someone clicks a meta ad and lands on a product page where their size doesn't exist. Before you optimize anything else, fix your catalog. Out-of-stock variants in your meta carousel are paying for clicks you can't convert.

Q: How does the cohort analysis actually work?

Sidharth: We do it at a city level for the top 10 cities, then split the rest into city tiers, tier 1, tier 2, tier 3. We do it at a device tier level too, iOS but older iPhone 11 is a tier 3 device, not tier 5. We split across these and run it for each individual collection at a regular cadence.

We auto-filter out things like wrong sizing, products with no inventory, anything where the catalog is not maintained. As the user scrolls, we calculate intent towards different things. Did they spend more time on black colors, on solid patterns, on a particular silhouette? We try to incorporate that as they scroll further so the niche tightens. The idea is they get more and more precise based on the intent they showed earlier in the session.

🔥 ChaiNet's Hot Take: This is where personalization stops being a buzzword. Helium isn't doing 50/50 A/B tests on hero text. They're rebuilding the collection page in real time based on city, device tier, and behavioral intent inside the session. The math is different. The output is different. And the conversion lift is in a different ballpark.

Q: Do general optimizations work for smaller brands, or do they need this level of personalization?

Sidharth: General optimizations totally work. If a brand is selling on their own website, they might also be on Mintra or Amazon. They have a sense of what tends to sell and they do a manual sort. Most medium scale brands don't have enough traffic for advanced personalization. A 2,000 to 3,000 SKU catalog with 50 to 70 orders a day is unevenly split. If you do a non-manual best-selling sort, the 70 selling SKUs come up, the system pushes those harder, and you end up with a self-reinforcing loop where most of your catalog never sees the light.

So yes, simple best-selling sort or manual ordering works until a certain scale. To optimize further, it starts breaking. You need cohort and intent based reordering.

🔥 ChaiNet's Hot Take: Start manual. Go basic. Don't pay for personalization until you've outgrown best-selling sort. Most founders try to skip the manual stage because it sounds unsophisticated, but the data Sidharth describes only works if you have enough volume to compute it. Personalization is a scale problem, not a startup problem.

Q: For a smaller brand with under 2,000 SKUs, what are the two or three most important things to do today?

Sidharth: First is investing in setting up a good catalog. Facebook catalog specifically. And optimizing your PDP. PDP is the core real estate any website has. Almost any scale brand should optimize for it.

That's where the maximum traffic lands and where the buying decision actually happens. Most other pages you can deprioritize.

🔥 ChaiNet's Hot Take: Catalog first. PDP second. Everything else later. If you remember nothing else from this conversation, remember the order. Most founders waste months redesigning their homepage. Sidharth is telling you the homepage is the worst place to spend your optimization budget.

Q: A lot of founders think conversion is a traffic problem. Just buy more clicks. Is that right?

Sidharth: Not really. Traffic is one of the core things you'll do, but it shouldn't be the primary cause of most conversion problems. If you have 0.1% conversion and you bring 20,000 sessions, you'll get 200 orders, which sounds like a lot. But your ROAS will be 1.5 to 2.5, depending on category. It won't scale.

The first diagnostic is whether you're bringing the right people. Most websites are set up decently OK. So the question is, am I evaluating the type of audience, not just the quantity? After that it's about the website itself. Are there even trust markers, would this reach me, X day delivery, returns, etc. Then visit-count cohorts. How many visits does it take for a person to actually convert? Most brands don't track that.

It boils down to traffic, but quality, not quantity. And it's hard. Platforms have made it really tough to track. Google has stopped giving user-level data. Meta doesn't give user-level data. We have a custom pixel that tracks a device rather than a user. We calculate intent on the fly. These cohorts can be used as a seed audience, or an older ad that worked well can be a seed audience to get better traffic if quality is the problem.

🔥 ChaiNet's Hot Take: Buying more traffic to fix bad conversion is like fixing a leaky bucket by pouring in more water. Sidharth's diagnostic is sharper. First, audit who you're attracting, not how many. Second, ask how many touches it takes to convert. If your answer involves "more spend," you're probably not asking the right question yet.

Q: Brands talk about ROAS like it's the only metric. Is that right?

Sidharth: ROAS is high level. It has traffic quality inside it, conversion rate inside it, retargeting strategy inside it, catalog quality inside it, and so on. It's an outcome, not an input. It's a good number to monitor at a regular cadence, but you can't throw a dart and say I will increase ROAS right now. You optimize the sub-parts. Optimizing each contributes to the main number.

The problem is most Shopify brands don't even have GA set up properly. Their Facebook pixels are not getting half the signals. So the brand is essentially saying I don't know if half these numbers are accurate, I won't focus on improving the quality of these numbers, I'll just optimize for ROAS because it's the only number I'm confident is correct. And then they ignore everything else.

🔥 ChaiNet's Hot Take: ROAS is the metric you trust because it's the only one Meta lets you trust. That's not a strategy, it's a default. The real winners audit their analytics stack, fix the broken pixels, and start optimizing the sub-metrics that actually drive ROAS. Trust the inputs, not just the outputs.

Q: Tell me about the RCA engine you built.

Sidharth: It's a custom pixel and a tracking system we built in-house. We were doing personalization, going deeper, and realized this is fundamentally a data problem. We needed to track way more than we did to even do better personalization.

That became the rabbit hole. Now we have an RCA engine that's a flowchart. ROAS at the top, then all the metrics that contribute to ROAS. Click conversion rate, you see the pages and which has what conversion rate. Click product, it shows trends, splits into iOS, Android, and traffic sources. Click Facebook, it splits further. A whole tree to traverse.

It's hard to consume manually because it's so deep, so we have an agentic layer on top of it. We call them anomalies. The system says my X-source audience with this device or region has gone down, conversion rate is down, here's where they're landing, here's how to optimize. We send these reports daily. My 10 to 20 ads on Meta, these three saw a drop because this happened.

We trace data down to the level of "homepage banner that points to this collection contributed this much directly to revenue, and this much indirectly." If the user clicked the banner, went to the bestseller collection, and bought, that's direct. If they went, saw a recommendation that doesn't exist in bestseller, and bought elsewhere, that's indirect.

🔥 ChaiNet's Hot Take: Most analytics tools tell you what happened. Helium's RCA engine tells you why, then automatically generates the fix. Banner-level attribution sounds insane until you realize most brands have no idea which homepage banner actually drives revenue. They guess. They redesign. They guess again. Banner attribution is a moat.

Q: There's a community question. What are brands completely blind to?

Sidharth: Brands tend to have a lot of tools and a lot of things they change themselves, but they don't know if their change actually caused something to go up or down. They added a size predictor, a chatbot, X feature, Y feature. Did it impact anything? They have no idea.

Many times brands optimize the wrong audience or the wrong fold. They work on things that won't move the needle, like the fifth fold of some particular page that 1 to 2% of audience reaches, or the third fold of homepage where homepage is only 5% of traffic and the third fold is 0.2%. You're optimizing for such a small bucket that even with a big lift, the ROI is invisible.

🔥 ChaiNet's Hot Take: "Did the chatbot you installed last quarter actually change conversion?" Most founders can't answer that. They added it, hoped it worked, and moved on. Sidharth's point is brutal. Without attribution, every change is a prayer. And most prayers go to the wrong page.

Q: As a CTO, where do you use AI in your daily life?

Sidharth: It's taken over. Nobody on our team is writing most of the code by hand. 95% is shipped through AI. You sit on top architecturally, but that's the better place to sit and just ship faster.

It hasn't always worked perfectly. Sometimes I've optimized the prompt, optimized the workflow, fed everything I can, and a state-of-the-art model still won't give the quality I want. That happens often, especially in creative generation for ads. But structurally, I set up a system first. We have a root file that points to every sub-file, so the AI gets the right context and doesn't have to re-read everything.

I personally use Antigravity more than Claude Code, but I do use both. Antigravity needed a bit of optimization but I like it because it's fast. When I have a spec, I do a bigger spec, a phased plan if I'm making something new, then split it into smaller models. So I have four parallel Antigravity sessions running.

🔥 ChaiNet's Hot Take: 95% of code shipped through AI, four parallel sessions running. This is what the new senior engineer workflow actually looks like. Architectural review of AI output replaces handwritten code. Speed and parallel capacity are the new productivity ceiling. The bottleneck is no longer keystrokes, it's how many sessions you can think about at once.

Q: How has AI changed your hiring plans?

Sidharth: I've stopped having interns. I used to give them work. I led teams of 10 to 12 with junior and senior devs at another company. Since coming here, AI has drastically removed the need for junior hires.

It goes both ways. We tend to ship more vertically than horizontally, so deep product knowledge matters more than horizontal hands. Interns don't really align with that, three months isn't enough to ship vertically.

Maximum impact is at the senior level. Seniors know what they want to do, the technologies that work, the dogs and functions, so they can quickly review what AI wrote and ship. Juniors don't know the depth, so they can't tell if AI's output is good or bad. So juniors and mid-levels need to either go deep into knowing every library and SDK, or pivot toward being a tech-enabled product person who can think of what to build, not just code.

🔥 ChaiNet's Hot Take: Senior engineers got 10x'd. Junior engineers got squeezed. The bridge for junior talent is no longer "learn to code." It's "learn to architect" or "learn to think like a product person." If you're early in your career and only learning to write code AI can already write, you're competing with the worst use of your time.

Q: Why is my Shopify store sending me WhatsApp messages I never gave my number to?

Sidharth: It's a gray area. There are providers like GoQuick and a few others. They sit on your website and if you've bought from any website, they have your information. So you don't actually log into Noise or another Shopify website. You log into GoQuick when you give your number at checkout. GoQuick is part of this website and that website. So when you arrive on a new website that has GoQuick installed, it knows who you are.

It's a loophole. Personally, I don't think it should be illegal. Google does the same thing. Opens an iframe, saves cookies, and that cookie lives across every website with GA. So you'd have to call Google bad too.

This was originally an RTO hack. Brands wanted to reduce return-to-origin rates, so they used pin-code level data. This pin-code returns more, this pin-code returns less. So I won't show cash on delivery here. Now they extended to sending WhatsApp messages, which is where it's starting to bug users.

🔥 ChaiNet's Hot Take: The infrastructure that helps brands reduce RTO is the same infrastructure that's now sending you unsolicited WhatsApp messages. Cross-site cookies, network effects across Shopify stores, and a regulatory gray zone that nobody is rushing to close. Useful tools become creepy tools when scope creeps. Watch the gray areas in your own stack.

Final Thoughts: ROAS Is Not Strategy

Sidharth's parting reframe: "ROAS is an outcome, not an input. You optimize the sub-parts."

The bottom line: The story Sidharth tells is the real anatomy of a Shopify business. The optimization hierarchy is PDP first, collection page second, homepage almost never. The biggest leaks happen before the website even loads, in Meta catalog ads pointing to out-of-stock variants. Most brands chase ROAS because it's the only number Meta lets them trust, while ignoring the sub-metrics that actually move it.

What Helium is really building is attribution down to the level of a homepage banner. Direct purchases, indirect purchases, and an agentic RCA layer that tells you which Meta ad to scale and which to kill. That's not personalization software. That's the missing analytics stack the platforms refused to give you.

For founders, the practical takeaway is brutal but useful. Audit your catalog before you audit your store. Fix your pixel before you fix your funnel. Stop optimizing the third fold of your homepage. Start tracking visit-count cohorts. And if you're using AI to ship product like Sidharth's team, set up architecture and root files so you can review fast instead of typing slow.

The next 12 to 24 months in Shopify will reward brands who treat their site like a system instead of a brochure. Real-time personalization, banner-level attribution, agentic optimization layers. The ones still A/B testing hero text in 2026 are going to wonder why their ROAS keeps falling.

Q: How can people connect with you and learn more about Helium?

Sidharth: You can check out gethealdim.co to see what the platform actually does. We work with brands like GoNoise, building landing pages and personalization stacks. I'm active on LinkedIn and you can connect with me there. We might also have my co-founder Shrey on a future episode for more on the business side, but if you're a Shopify founder thinking through any of this, do reach out.

Final words: Sidharth's worldview is simple. Stop optimizing what doesn't move. Start measuring what does. ROAS is an outcome, not a lever. Homepage is real estate, not strategy. The catalog leak in your Meta ads is bleeding more money than your conversion rate ever will. The brands who win the next decade of Shopify will be the ones who treat their store as a measurable system. Trace revenue down to the banner. Personalize down to the cohort. Attribute every change. Everything else is throwing darts in the dark and praying.

---