Turning a week of netlinking into an hour

The monthly netlinking routine is roughly seven steps. Some of them are pure computation and should never be done by a human again. Others are judgement calls that automation makes worse.

Knowing which is which is the whole design problem.

What should be automated

Four steps are mechanical. They follow a formula, they produce the same answer every time, and doing them by hand at scale is where quality quietly degrades.

Pulling and scoring Search Console data. Aggregating page-level impressions, CTR and position over 28 days and computing an opportunity score is arithmetic. There is no judgement in it.

Excluding recently-targeted URLs. Checking whether a page received a link in the last two months is a lookup against your own history. A human doing this from memory will get it wrong.

Filtering marketplace inventory. Applying Trust Flow, traffic and referring-domain thresholds across a few hundred listings is exactly what software is for. So is deduplicating host domains across every line of an order.

Tracking delivery. Polling each order until the article is published, capturing the article URL and publication date, and updating status. Nobody should be refreshing a marketplace dashboard by hand.

Together these are the bulk of the time cost and essentially none of the intellectual value.

What should stay manual

Two things resist automation, and forcing them through anyway produces expensive mistakes.

The final confirmation before money leaves. Automated preparation should produce a draft, never a purchase. The gap between "here is what I would buy" and "I bought it" is where a human catches the page that is no longer relevant, the host that looks wrong, the budget that drifted.

This matters more on some platforms than others. Where an order creates a cart that a human pays later, an automated mistake is recoverable. Where a confirmation is an immediate purchase, it must require an explicit, itemised approval, never a single "confirm everything" button that can be misclicked into a month's budget.

Strategic adjustments. A client launched a new service page. A competitor overtook you on a key query. A page you have been pushing finally ranks and should be retired in favour of the next one down. None of this is in the data yet, and an automated system will happily keep optimising for last quarter's reality.

The shape that works

The pattern that holds up is prepare automatically, review manually, execute explicitly.

On the first of the month, a scheduled job runs for every configured site: pull GSC, score the pages, exclude recent targets, pick hosts within budget, generate anchors, and write a draft order. Nothing is bought. Nothing is charged.

When you sit down, the drafts already exist. Your work is review, meaning you adjust the targets that need it, reroll a host that looks wrong, rewrite an anchor, add or remove a line. Then confirm, per order, with the total in front of you.

Two properties make this safe.

Preparation must be idempotent and re-runnable. If a site was added mid-month, or the scheduled run failed, you need a "prepare this one now" button that produces the same result. A job that only works on the first of the month is a job that silently skips clients.

Failures must be visible. When a marketplace has no inventory in a category, or a language check rejects every candidate, that has to show up in the review UI as an empty or short order, not as a log line nobody reads. The most expensive automation failure is the quiet one.

The anchors question

Anchor generation sits awkwardly between the two categories, and it is worth being precise about why.

Writing an anchor is a judgement call. It needs to read naturally, match the page's actual topic, and differ from every anchor already used for that URL. That is exactly the kind of task a language model does well, and doing it by hand a hundred times a month is exactly where humans start reusing phrases.

So generate them automatically, but treat the output as a draft that a human scans. Reading twenty proposed anchors and fixing two takes three minutes. Writing twenty from scratch takes forty. The review step is what keeps quality up, and the generation step is what makes the review affordable.

Give the generator the constraint it needs to be useful: the page's top queries from Search Console, and the list of anchors already spent on that URL. Without the history it will produce variations that are individually fine and collectively repetitive.

What you get back

The realistic outcome of automating the mechanical steps is not that netlinking becomes free. It is that the per-client cost stops scaling with client count.

The forty minutes per client per month becomes roughly five minutes of review, plus the occasional deeper look at a client whose situation changed. Twenty clients goes from a week to an afternoon, and more importantly the twentieth client gets the same page-selection rigour as the first, because a formula does not get tired at client fourteen.

The side effect worth naming: because every order is recorded as it goes out, the client report stops being a monthly reconstruction and becomes something you can generate at any moment.