Growth & Search
We run Google, Meta and Microsoft paid media against contribution margin and tested incrementality — and we build the server-side tracking and product feeds the platforms actually learn from. Engineering first, media buying second.
Most paid media relationships break at the same point: the dashboard says 6x ROAS and the bank balance says otherwise. ZenMagix provides PPC management services from Mumbai for companies that have already been through that and want the arithmetic to reconcile. We are a software engineering firm that runs paid media, which means the parts other agencies outsource or ignore — server-side tagging, offline conversion imports, product feed pipelines, margin modelling in the warehouse — are the parts we do ourselves. Platforms have taken most of the bidding levers away. What is left to compete on is signal quality, feed structure, creative volume and honest measurement. That is an engineering problem before it is a media-buying one, and it is where we spend our time. We work with ecommerce, SaaS and lead-generation businesses across India, the UK, the US and the Gulf.
ROAS is the number agencies report because it is the number that always looks good. It ignores cost of goods, returns, shipping, payment fees and the customers who would have bought anyway.
Take a 6x ROAS at ₹100 revenue. If COGS is ₹42, shipping and fulfilment ₹14, payment gateway ₹2, and 12% of orders come back, the ₹16.67 of media spend that produced it is sitting against roughly ₹20 of gross profit before anyone has paid a salary. Push volume and the marginal order is loss-making long before blended ROAS moves. The fix is not complicated, it is just unglamorous: get product-level COGS and return rates into the warehouse, join them to spend and orders, and report contribution margin after media by campaign, product and cohort.
Then bid to it. Lead generation has the same disease in a different costume — cost per lead falls, sales close rate falls faster, and nobody notices for a quarter because the CRM and the ad platform never speak. We import qualified-lead and closed-won values back into the platforms so bidding optimises toward revenue that exists rather than form fills that do not. The uncomfortable consequence is that our reports sometimes show a smaller number than the last agency's. It is the true one.
Attribution tells you which ad got credit. Incrementality tells you what would have happened if the ad had never run. Only one of those is a business question.
Branded search is the standard example. It converts beautifully, it costs little, and a meaningful share of it is people who typed your name and would have arrived regardless. Retargeting is worse — you are frequently paying to reach buyers already halfway to checkout. The only way to know your share is to switch something off and measure the difference. We run geo holdout experiments: matched sets of cities or states, treatment and control, spend suppressed in control for two to four weeks, results analysed with GeoLift or CausalImpact against a synthetic baseline rather than a naive before-and-after.
Where geo splits are impractical, we run PSA or ghost-bid tests on Meta so the holdout group sees a placeholder and the lift is measured cleanly inside the platform's own auction. Most agencies never run one, and the reason is straightforward: a test that shows 40% of reported conversions were not incremental is an argument for reducing the spend they are paid a percentage of. We run them quarterly and calibrate the whole media mix model against the results, so the numbers you plan with have been checked against reality at least once.
Automated bidding genuinely beats manual CPC when it is fed good data. Performance Max genuinely hides where your money went. Both things are true, and the job is working within that.
Smart Bidding won on merit. No human adjusts bids by device, time, audience and query context across millions of auctions per day. Handing that over was correct. Performance Max is a different bargain: real reach across Shopping, Search, YouTube, Display, Gmail and Discover, in exchange for a channel report that will not tell you what the network split was or which search terms paid. The levers that survive are structural. Asset group segmentation by margin tier or product category — not one campaign containing your whole catalogue.
Feed exclusions and custom labels to steer budget toward products that actually make money. Brand exclusion lists so PMax stops harvesting branded search and calling it prospecting. Search themes and audience signals as seeds. Conversion value rules for geography and new-customer acquisition. And, critically, the value you send in: a conversion action carrying real margin makes Smart Bidding solve the right problem, while one carrying gross revenue makes it very efficiently find your worst-margin SKUs. We also keep standard Shopping and Search running alongside PMax where the data justifies it, so there is a controllable comparison rather than a single black box you are obliged to trust.
Bidding algorithms are only as good as the conversion signal they receive. Fixing that signal is software work, and it is the part most media agencies quietly subcontract.
Browser-side tracking has been degrading for years — ITP, ad blockers, consent banners, iOS. The response is server-side: Google Tag Manager server container on Cloud Run, first-party endpoint on your own subdomain, events forwarded to Google Ads, Meta Conversions API, GA4 and your warehouse from one governed place, with deduplication and event IDs handled properly rather than double-counting every purchase. Enhanced Conversions and hashed first-party identifiers restore match rates the browser lost.
For anything with a sales cycle, offline conversion imports close the loop: CRM stage changes and closed-won values flow back on a schedule so the platform learns from revenue, not from a thank-you page. Product feeds get the same treatment. Merchant Center disapprovals, missing GTINs, thin titles, wrong availability and stale prices cost more real money than most bid strategies ever will, and they are fixable with a pipeline rather than a spreadsheet — title construction from structured attributes, supplemental feeds for margin and stock labels, automated disapproval monitoring, and hourly price and availability updates via the Content API. This is ordinary engineering. We just do it in-house, which is why it gets done.
What you get
COGS, returns, shipping and fee data joined to spend and orders in BigQuery, with contribution margin after media reported by campaign, product and cohort. This becomes the number we manage to, and you keep the models.
Server-side GTM on your own subdomain, Conversions API, Enhanced Conversions, consent handling, event deduplication and offline conversion imports from your CRM. Built, documented and owned by you, not rented from us.
Automated feed generation and enrichment, supplemental feeds for margin and stock labels, disapproval monitoring with alerting, and hourly price and availability sync. Applies to Google Shopping, Meta catalogues and marketplace feeds.
Day-to-day build, structure, bidding, budget allocation, audience and creative testing across Search, Shopping, Performance Max, Demand Gen, Meta and LinkedIn — with a written test log so you can see what was tried and what it showed.
Designed geo holdout or PSA tests each quarter with power analysis, synthetic-control results and revised budget guidance. Weekly reporting in Looker Studio on margin, marginal CAC and test status rather than screenshot ROAS.
How we work
Two to three weeks. We reconcile platform-reported conversions against your backend orders, trace every tag and event, audit feed health and Merchant Center status, and quantify the gap. You get the findings whether or not you hire us.
Server-side tagging, deduplication, Conversions API, offline imports and the margin model go in before campaign work starts. Optimising a campaign on broken data just gets you to the wrong answer faster.
Account restructure by margin tier and intent, value-based bidding targets set against contribution margin, feed pipeline live, creative testing cadence established. Spend is held flat while the new measurement is verified against backend numbers.
First incrementality test runs in month two or three. Budgets move to where marginal margin is positive, and scale decisions reference tested lift. Quarterly retest, because channel dynamics and your product mix both change.
Every phase ends at a decision point you can stop at — see how that works across fixed-scope projects, embedded pods and retainers.
Stack
Questions
Flat monthly fee, scoped to the work. Percentage-of-spend pricing pays us more for spending more, which is a poor incentive when the honest recommendation is often to cut a channel. We will quote percentage-based if you insist and it suits your finance team, but we will tell you why we think it is the worse structure.
Roughly ₹5 lakh, about $6,000, per month. Below that, agency fees consume too much of your gross profit to justify. The better move is a fixed-fee setup engagement — tracking, feeds, account structure, bidding targets — then run it in-house with a quarterly review. We do that work and say so upfront.
Feed and tracking fixes often move numbers within four to six weeks because they compound immediately. Structural changes need six to eight weeks for Smart Bidding to relearn, and performance usually dips first. A defensible incrementality read takes a full test cycle, so month three is the earliest honest verdict.
We do not cut it on a hunch, we test it. A geo holdout with branded search paused in control markets for three weeks shows what organic recaptures. Sometimes it is nearly everything and the budget moves; sometimes competitors bid on your name and the spend is defensive. Either way you get an answer, not an opinion.
You do, always. We work inside your Google Ads, Meta Business Manager, GTM and cloud accounts with our own user access, never a reseller MCC that holds your history hostage. Server-side containers run on your infrastructure, warehouse models live in your BigQuery project, and offboarding means revoking access.
Yes, and it is the more common case. We have implemented server-side tagging and conversion pipelines on custom Next.js, Laravel, Django, WordPress and headless commerce stacks. Because we are a software engineering firm, the data layer, the API integration and the feed generator are jobs our own developers do rather than tickets we raise with yours.
Questions about cost, timelines, IP ownership and data residency are answered on the general FAQ, and how this practice came out of blockchain infrastructure explains why we build the way we do.
Related
The full growth stack, coordinated rather than sold in disconnected retainers.
Getting cited by ChatGPT, Perplexity and AI Overviews — measured, not promised.
Engineering-led SEO: crawl budget, rendering, structured data, Core Web Vitals.
See also: see how paid media fits the wider digital marketing programme · keep the creative pipeline full with AI ad creative automation · reduce dependence on paid clicks with technical and content SEO
Tell us what you are trying to build. We will tell you honestly whether we are the right team for it, and what it would realistically take.
