Every campaign goes through the same four-step pipeline: live research, a strategy brief, a single tactical pass, and schema validation. This page explains each step and the reasoning behind the architecture. The same pipeline runs for every user; there is no premium path that produces a different quality of output.
Step 1: Live web research
Before any generation happens, the tool runs three web searches built from your inputs:
- your ideal customer and their pain point, looking for marketing strategies that already work on that audience;
- your product category, looking for competitor campaigns and examples;
- your campaign goal, looking for best practices for that objective with your audience.
Each search returns up to five results. The results are deduplicated by URL, capped at 15 sources, and passed to the strategy stage as evidence. The sources are also stored and listed in your report, so you can see exactly what grounded the strategy.
Two design choices matter here. First, research is treated as untrusted input: the retrieved text is fenced off inside the prompt and the model is instructed to analyze it, never to follow instructions found inside it. Second, research failure never blocks generation. If the search provider is down, the campaign still generates from your inputs alone; the report simply lists no sources.
Step 2: The strategy brief
The first model call produces only the strategy layer: the positioning statement, exactly three messaging pillars each with a proof point, a channel mix ranked by expected ROI for your budget tier, three to five key differentiators, and a campaign theme and concept. It also writes a short preview of each tactical section, which forces the strategy to be concrete enough to execute against.
This stage runs on a heavyweight model tier because it is the highest-leverage call in the pipeline: every downstream deliverable inherits its quality. The prompt bans generic marketing filler ("unlock growth", "game-changer") and requires the output to reference your actual product and pain point.
Step 3: One tactical pass, not seven
All seven tactical deliverables are generated in a single model call that receives the full strategy brief. This is deliberate. Generating sections independently is faster to build but produces campaigns that drift: emails that contradict the positioning, budgets that fund channels the strategy never recommended. One pass with the brief in context keeps the whole campaign internally consistent, and it is enforced with hard rules:
- budget allocation may only use channels named in the brief's channel mix, and the percentages must sum to exactly 100;
- CTAs and KPIs must follow the campaign goal you selected;
- email send days must respect your sales cycle length;
- hooks, angles, and phrasing must not repeat across sections.
This stage runs on a high-throughput model tier tuned for large structured output, because it writes the bulk of the report: roughly 20 posts, 5 full emails, 10 LinkedIn posts, and 4 more structured sections in one response.
Step 4: Validation, or an honest failure
Both stages are validated against strict schemas before anything is saved: exactly 4 calendar weeks, exactly 5 emails in the right stage order, exactly 10 LinkedIn posts, budget percentages that sum to 100, KPI benchmarks inside realistic caps, and so on. Output that fails validation is retried once with the exact validation error included in the prompt. If it still fails, the report is marked failed and the retry button regenerates the entire campaign.
The result is a binary contract: you either get a complete, schema-valid report with all eight sections, or a clear failure state. There is no silent half-report. A watchdog also reclaims any run that dies mid-pipeline, so a failed generation is reported as failed rather than leaving the progress screen spinning.
Repeatability
Identical inputs within 24 hours return the same stored report rather than a new generation. Beyond that window, or with any input changed, a fresh run generates a new campaign. Like all generative output, two runs with the same inputs will differ in wording while following the same strategy shape, because both anchor to the same goal, budget, and research landscape.