Thirty minutes of setup, then a loop that measures your pages, writes what you queue, and comes back later to tell you whether it worked. Choosing the work stays with you.
Read-only, takes about thirty seconds. Penlid starts a permanent archive: the last few days first, then it walks backwards a window at a time to Google's 16-month floor, so this week works tomorrow and the history fills in behind it. From this moment your data outlives Google's retention window. Free forever, on every plan, including no plan.
Penlid crawls your pages to learn your voice, your audience and your internal link graph. You confirm or correct what it inferred, add any phrases you never want used, and set your publishing destination.
Penlid proposes the buying-intent questions your customers actually ask. You edit them — they're your prompts, and prompt choice shapes the measurement, so you should own it. Sampling begins immediately.
Set the monthly model budget for the account. It is reserved before every paid job, so two jobs running at once cannot walk through the same cap. Two things are not settings: every draft waits for a person to approve it, and drafts below the 85% fact-score threshold, or with any claim contradicted or unchecked, cannot be approved.
Pulls the newest Search Console data Google will serve — currently three days behind, which is Google's lag, not ours — appends it to the archive, and re-scores every page for decay — adjusted against your site-wide trend so a seasonal dip doesn't masquerade as a problem. Pages below the reliability floor are marked unknown rather than guessed at.
Runs every tracked prompt once per engine, every day — not weekly, and that daily cadence is what each plan's monthly sample allowance is built from. An engine is sampled only if it is both configured on this deployment and inside your plan's engine count, so a deployment with five keys still samples three for a Free account and four for Core. Claude, with live web search, is always one of them; the rest come online as their key and adapter are configured. Records each answer, each cited source and the model version that produced it, and aggregates them into a citation rate with a 95% Wilson interval. The report names the engines that actually answered, and the free check names the ones this deployment can sample right now.
The dashboard marks every page healthy, slipping, decaying or lost, so you can see which ones are losing ground. Deciding what to write and what to refresh is yours. Once you queue an article, the rest runs on its own: it researches, drafts, optimises, and fact-verifies every checkable claim, then waits for your review.
Before anything goes live, Penlid records exactly what changed and picks a matched set of pages it is deliberately not touching, matched on traffic magnitude and trailing trend. Sharing the target's first URL path segment (the closest thing to a topic cluster a URL alone tells us, not a content analysis) counts as a modest bonus toward being picked, never a requirement — on a flat site, or where the nearest pages by traffic sit elsewhere, the controls come from elsewhere and the match is still a real one. This is the step that makes the next one possible, and it can only be done at the moment of the change.
Once the window has elapsed and Search Console's three-day reporting lag has cleared, Penlid compares the changed pages against their frozen controls over the same calendar days, and stores the estimate with its 95% prediction interval, p-value and control count. Every evaluated change gets an answer — improved, declined, inconclusive, or contaminated when a control was itself touched inside the window.
Automation that can publish to your site and bill your card needs limits that are enforced, not advisory.
Connect Search Console and get the decay dashboard and permanent archive free, forever. Add writing when you're ready.