DEJAN's outreach pipeline runs an adversarial simulation that pairs an generative Red Team agent with an independent, blind Blue Team evaluator. Operating with high-reasoning Gemini models equipped with live web search and URL context tools, the Red Team drafts realistic off-site articles tailored to custom editorial guidelines. Crucially, the system prompt of the Red Team contains the full, embedded Blue Team rubric, forcing the writer to actively anticipate and counter the detection heuristics used to flag unnatural promotional links while sourcing authentic citations across the web.
Before an article is judged by the detector, it passes through structural validation layers:
banned-link verdict.The sanitized article is then handed to the Blue Team detector cold at zero temperature. Mimicking a human editor or webmaster, the detector evaluates the page to classify external links into organic filler versus commercial "money links." If the detector correctly isolates the client's target link, the draft is flagged as detected and folded into the diagnostics log with the detector's specific rationale. Only articles that successfully bypass detection, retain their target placement, and comply with all safety boundaries receive a passed verdict and are archived into the property's content library.
For post-generation review and manual draft refinement, operators can deploy the Outreach Coach. This secondary pass ingests the detector's findings and translates structural red flags into concrete writing advice. The coach outputs discrete before-and-after suggestions, strictly enforcing that every proposed replacement corresponds to a verbatim substring in the original draft to ensure automated replacements never corrupt surrounding prose.
Referenced by