Paid Ads: A Manifesto For Marketing Engineers

What's happening

With the power of Claude Code, OpenClaw, and Codex, it's unsurprising that Marketing Engineering is becoming its own category. Marketers who can build — who can connect APIs, architect systems, and automate entire workflows — are carving out a role that didn't have a name six months ago. We think that's exciting. And we think paid ads is one of the most under-explored places to apply it.

At Multiply, we've spent the last year building AI agent systems purpose-built for paid media. In many ways, our AI-native, expert-led managed service represents a version of what Marketing Engineers might try to build for their own organizations.

Everyone who's run paid media knows that the old model is too slow and too expensive to compete on: humans manually refreshing creative, pulling reports by hand, updating audiences one by one, and trying to evaluate and iterate from unclear results week after week. So we built systems to handle it.

What we've learned feels relevant to this emerging category. Specifically: what does Marketing Engineering actually look like in paid ads? Where does it add the most value? And where can it go badly wrong?

The overhead problem

Most B2B marketing teams spend the majority of their paid media time on work that isn't strategy. We know because our team have all done it. Pulling reports. Refreshing creative. Updating audiences. Reconciling attribution.

None of that is where the real value lives. The value lives in the decisions — what to test, when to test, who to target, what message to lead with. But those decisions get squeezed into whatever's left after the overhead.

Marketing Engineering should focus its efforts on attacking the overhead, of which there is plenty: automate everything around thinking, so the human can do more strategy and fewer tedious manual tasks. There are two ways to think about what that looks like in practice.

Part 1: Acceleration

Creative refresh

The problem: all ads decay. What worked in January stops working by March. Audiences tune out, relevance drops, performance falls. Most teams know this and still refresh creative quarterly at best, because producing, reviewing, and trafficking new assets is slow.

The opportunity: build a system that generates creative variants, routes them for approval, and pushes to the platform on a weekly cadence. The result isn't that any single refresh is transformative — it's that campaigns keep learning instead of going stale.

• Creative generation: AdCreative.ai, Pencil
• Templatized production: Canva API, Figma API
• Workflow automation: Make.com, n8n
• Publishing: LinkedIn Ads API, Google Ads API

A/B testing at scale

The problem: the bottleneck is almost never ideas — it's execution. Setting up tests, monitoring them, calling winners, and cutting losers all takes manual time.

The opportunity: build infrastructure that runs testing continuously, calls statistical significance automatically, and cuts underperformers without waiting for a human to log into a dashboard.

• Campaign management: LinkedIn Campaign Manager API, Google Ads API
• Performance data: Supermetrics, Funnel.io
• Decision logic: n8n or Make.com (e.g. pause an ad if CPL exceeds a threshold after N impressions)
• Experiment logging: Airtable, Notion

Audience maintenance

The problem: ICP lists drift. Closed-won data changes. New segments emerge. Campaigns keep running against yesterday's audience profile.

The opportunity: wire your CRM to your ad platform so audiences update automatically. Win a new customer in fintech? Your fintech segment reflects it. A cohort of churned accounts? Suppressed before you spend another dollar against them.

• Source of truth: HubSpot, Salesforce
• Platform sync: LinkedIn Matched Audiences, Google Customer Match
• List building and enrichment: Clay, Riveter
• Sync logic: Zapier, Make.com

Lead quality feedback loops

The problem: leads come in, land in the CRM, and the campaign never hears what happened next. It keeps optimizing toward form fills, not toward the leads that actually became pipeline.

The opportunity: build the plumbing — CRM to enrichment to lead quality scoring to campaign signal — so the platform's optimization target is 'leads that close,' not 'leads that click.'

• Enrichment: Clearbit, Apollo
• Deal tracking: HubSpot, Salesforce
• Signal push: n8n or Make.com via offline conversions (LinkedIn, Google)
• Tracking infrastructure: Google Tag Manager
• Data layer: Segment

Part 2: Invention

Once the recurring work runs itself, what would you build that your team has never had time to attempt?

Buyer language, injected in real time

The problem: every sales call contains better ad copy than anything written in isolation. Customers say exactly what they're worried about, what made them book the demo, what almost stopped them from signing. That language almost never makes it into ads, because processing call recordings at scale takes time no one has.

The opportunity: build an agent that processes recordings continuously, extracts recurring phrases and objections, and surfaces them as copy inputs for active campaigns. Messaging grounded in real buyer language consistently outperforms messaging built from internal assumptions.

• Call recording and transcription: Gong, Chorus, Fireflies.ai
• Theme and phrase extraction: OpenAI API, Claude API
• Copy bank: Airtable, Notion
• Orchestration: n8n, Make.com

Competitive monitoring with automatic response

The problem: competitors change pricing, launch features, and update positioning. If your campaign is still running the same comparison messaging from six months ago, you're losing deals without knowing why.

The opportunity: a system that monitors competitive surfaces in real time, judges when something meaningful has changed, and stages a response — updated talk track, revised ad copy, refreshed comparison page — for human review before anything goes live. The strategic work still happens with a human. The monitoring and drafting don't have to.

• Ad intelligence: Semrush, SpyFu
• Traffic and positioning: SimilarWeb
• Custom scraping: Playwright, Puppeteer
• Judgment layer: OpenAI API, Claude API
• Alerts: Slack webhooks
• Orchestration: n8n

Dynamic ABM surround

The problem: most ABM ad programs are a static list and the same three creatives running until someone remembers to update them.

The opportunity: build a system where list changes propagate to campaigns automatically. An AE adds a new target account and it's in the ad audience within 24 hours. A deal closes and the account is suppressed. Intent signals shift and messaging adjusts.

• List building: Clay
• Intent data: 6sense, Bombora
• Dynamic list updates: LinkedIn Matched Audiences API
• Deal stage source of truth: HubSpot, Salesforce
• Sync automation: n8n, Make.com

Attribution that actually closes the loop

The problem: last-click is lying to you. The LinkedIn ad someone saw three times before they Googled you and booked a demo doesn't get credit, so you under-invest in what's actually driving pipeline.

The opportunity: multi-touch attribution that connects impressions to pipeline. It's not a new idea, but actually building it — connecting ad platforms, CRM, enrichment data, and revenue outcomes — is the kind of project that sits on roadmaps for years and never gets done.

• Out-of-the-box MTA: Triple Whale, Northbeam, Rockerbox
• Lightweight approach: Google Analytics 4 with custom conversion events
• In-ecosystem: HubSpot revenue attribution
• Data layer: Segment
• Custom model: BigQuery or Snowflake plus dbt

Where it goes wrong: a story about $30K and 174 phantom leads

I want to be direct about the failure mode here, because it's real and we see it often. One of our consultants recently worked with a client who had been running their LinkedIn campaigns through OpenClaw: full automation, no human in the loop, max delivery on, audience expansion on. This sounds smart, and exactly what your team might want a Marketing Engineer to build.

The platform optimized for what it was told to optimize for: lead form fills. It got very good at finding people who would click. It got very bad at finding people who would buy. Audience expansion pushed beyond the ICP. Max delivery went wherever it could convert, regardless of fit.

This isn't a technology failure. The tool worked exactly as designed. It's a judgment failure. Nobody with marketing sense was watching. The platforms will optimize toward whatever signal you give them, and if that signal is wrong, they'll optimize harder in the wrong direction. The cardinal rule of AI prompting holds: garbage in, garbage out.

When that division holds, automation compounds. When it doesn't, you get 174 phantom leads.

Seem like a lot?

It is. What we've just described — acceleration work, invention work, and the judgment layer that keeps the whole system from optimizing you into a ditch — is a serious marketing engineering project. Most marketing teams won't have the niche understanding of paid media, nor the time and headcount to stand it up from scratch properly. Even the ones that do will likely underestimate how much of the work is ongoing maintenance rather than one-time build.

If you've got the engineering appetite to build it in-house, great — the playbook above is a real one and we'd love to compare notes. If you'd rather skip to the output of a Marketing Engineering function without standing one up, that's what we do at Multiply.

What we're still figuring out

• Attribution is genuinely hard. Most companies don't have the data hygiene to support a clean multi-touch model yet.
• Buyer language injection works better in some categories than others. It depends heavily on sales call volume and consistency.
• Competitive monitoring generates noise if the signal detection isn't tuned carefully.
• Lead quality feedback loops require CRM hygiene that most companies haven't done the work to establish.

Marketing Engineering in paid ads is a real opportunity. It's also genuinely early. The teams that treat it as an experiment — build, measure, adjust — will figure it out. The ones that hand it to a tool and check the dashboard once a month won't. We'll keep sharing what we learn.