Traffic planner that thinks before it clicks

The client operated in a space where organic rankings weren't won by content alone. Behavioral factors — click-through rates, dwell time, bounce patterns — played a measurable role in how their pages performed in search. Their team had been managing this manually: picking keywords, estimating how much simulated traffic each page needed, and feeding those numbers into execution tools one by one. The process was slow, error-prone, and disconnected from actual SERP data. By the time a plan was ready, the rankings it was based on had already shifted.

planning workflow

We built a planning module that ties keyword research, SERP analysis, and traffic modeling into a single workflow. It starts with keywords — an operator loads target queries, and the system pulls fresh SERP data via XMLStock to see where the client's pages currently sit, who occupies the top positions, and what the click distribution looks like for each query. Claude then analyzes the competitive landscape and suggests traffic volumes and behavioral patterns that would look organic for a page at the target position. The output is a structured plan: keyword, target URL, daily session count, expected dwell time range, and a confidence score.

conservative defaults

What made this delicate was the nature of the work itself. Behavioral factor simulation lives in a grey area — the system needs to produce plans that are effective but conservative enough to avoid detection. We built configurability into every layer. Operators can set aggression levels per keyword, cap daily volumes, define cool-down periods between campaigns, and flag any plan for manual review before it goes to execution. The system defaults to cautious — it would rather under-deliver than trigger anomaly detection on the search engine side.

tiered serp refresh

The dependency on fresh SERP data created a cost problem. XMLStock charges per query, and checking positions daily for hundreds of keywords adds up fast. We implemented a tiered refresh strategy: high-priority keywords (top 20 positions, high traffic potential) get daily updates, mid-tier keywords refresh every three days, and long-tail queries update weekly. This cut API costs by roughly 60% without meaningfully impacting plan accuracy. The system also caches SERP snapshots, so if an operator re-runs analysis within the refresh window, it uses stored data instead of burning another API call.

variance modeling

Training the behavioral models was an iterative process. Early versions produced plans that were technically valid but obviously artificial — perfectly uniform session durations, identical bounce rates across all keywords. We added variance modeling: each plan now includes randomized ranges rather than fixed numbers, with distributions that mirror real organic traffic patterns. The models improve over time as the system tracks which plans led to actual ranking changes and feeds that data back into future suggestions.

results

The result is a planning layer that reduced plan creation time from a full day of manual work to about twenty minutes of review and approval. More importantly, it connected planning to reality — plans are based on current SERP data, not last week's spreadsheet. The honest limitation is that this is still a tool that requires experienced operators. The system suggests, but a human decides. We deliberately kept it that way — fully automated behavioral simulation without human judgment is a risk the client wasn't willing to take, and neither were we.

The takeaway: the value wasn't in automating execution — it was in automating the thinking that happens before execution. A bad plan executed perfectly is still a bad plan.

Frontend & API: Next.js 14 (Route Handlers), Prisma, PostgreSQL

SERP Data: XMLStock API with tiered refresh and caching

AI Analysis: Anthropic Claude API for competitive landscape modeling