The AI feature was straightforward. The data layer was the actual product work.

The situation

Brand Metrics provides brand lift measurement for digital campaigns. Creating survey questions for those campaigns had always been a manual process, and by 2025 it had become a genuine workflow problem.

The workflow had evolved on its own over time: first Excel with complex formulas, then isolated ChatGPT conversations, then ChatGPT Projects for team access. Each step helped, but none of it was integrated with the platform. Customers were re-entering the same campaign context every session and copy-pasting results back into their workflow, with no version control and no connection to their actual campaign data.

The opportunity was clear: build AI Survey Generator directly into Brand Metrics, eliminating the external dependency and using campaign metadata that already existed in the platform to produce better output than any generic AI tool could.

What I found

The discovery work surfaced something that changed how we approached the build. Survey question generation looks like a text generation problem, but the quality of the output is almost entirely determined by the quality of the input context. Brand name, product category, creative format, industry vertical, ad copy from the actual campaign creative: the more of this the model has access to, the better the questions it produces.

The problem was that this metadata lived across multiple systems and wasn't consistently structured. Some of it was in the Brand Metrics platform, some in Google Ad Manager, and some existed only as manual customer input. Without a clean data layer, even well-engineered prompts would produce output that was generic rather than campaign-specific.

"Metadata completeness is a quality metric. We couldn't treat it as an infrastructure concern and separate it from the product thinking."

This reframed the whole project: the AI feature required a metadata architecture first, and prompt engineering came last.

The approach

Working directly with Engineering, we designed a metadata architecture that could support AI Survey Generator and serve as the foundation for future AI features on the platform.

We validated the API against existing production campaigns before writing any prompt logic. That validation surfaced data gaps early and gave us a realistic picture of what Phase 1 could reliably support and what would need automation before we could scale.

Phase 1 shipped in 3 weeks: static metadata with manual entry as fallback, basic prompt engineering, and beta access for select customers. No GAM automation, no dynamic classification. The core workflow working reliably, with the architectural scaffolding for subsequent phases already in place.

Manual fallbacks in Phase 1 were a deliberate design choice. They kept the feature working when upstream data was incomplete and let us validate customer demand before investing in the automation that Phase 2 would deliver. The 3-week timeline was achievable because the architecture decisions were correct upfront.

Quality assurance from day one

We defined what good output looked like during discovery, not after launch. Relevance, clarity, and alignment with brand measurement methodology were the three dimensions that mattered, and beta testing with select customers served as the first evaluation loop.

Customer feedback shaped prompt iteration directly. Metadata completeness also became a measurable quality signal during beta: campaigns with richer context produced noticeably better output, which confirmed the data-layer-first approach and gave us a clear signal for what to automate next.

Formal benchmarking and production observability dashboards were scoped for Phase 2. Phase 1 needed to generate real production usage before there was anything meaningful to measure against.

The outcomes

Beta customers adopted AI Survey Generator immediately after launch and the external ChatGPT dependency was gone. The phased architecture meant the team could continue iterating on subsequent phases without rework, and the feature validated AI as a monetisation direction for Brand Metrics at a moment when the business was actively looking for new revenue streams.

"The prototype shipped in 3 weeks. The architecture it sits on will carry every AI feature that follows."

What this taught me

AI features require data infrastructure thinking from the start, because the quality of what a model produces is a function of the quality of the context it receives, and that context is always a data problem before it's a prompt engineering problem. Metadata architecture is a product decision.

Phased rollout with deliberate fallbacks is how you ship quickly without locking yourself into decisions you'll need to undo later. The manual entry fallback in Phase 1 was a design choice that kept the feature working reliably while Phase 2 automated the parts that actually needed automation.

Building evaluation criteria into discovery changed how we made decisions throughout the build. Knowing upfront what good output looked like meant we could use beta feedback as signal rather than noise, and it gave the observability work in Phase 2 a clear target to measure against.