There is a version of this argument that goes nowhere, where one side says the craft is dead and the other says nothing has really changed. Both are wrong in the same boring way.
What actually happened is that the work split. Some of it turned out to be mechanical and got handed off. Some of it turned out to be judgement and stubbornly did not.
Knowing which half you are holding at any given moment is most of the skill now.
What manual used to mean
If you did this ten years ago, you remember the rhythm. Pull the keyword export. Sort it. Group things by hand into clusters that felt right. Open competitor pages in twenty tabs and read them properly. Write the brief. Argue with someone about the title tag.
It was slow and it was also where you learned the category. You could not spend two days inside a competitor’s content without forming opinions about what they were bad at.
That knowledge was a by-product of the grind. Which is worth remembering, because when the grind goes away, the by-product goes with it unless you deliberately replace it.
The volume problem broke first
The thing that made manual work stop scaling was not laziness. It was surface area.
A site that once needed forty pages to cover a category now needs those forty pages to answer questions in a dozen phrasings each, stay current as the category shifts, and hold together as a set rather than as a pile. Add the fact that visibility now spreads across several answer engines with different appetites, and the audit that used to take a week takes a week per surface.
Then the drop happened. A lot of large content sites watched their search traffic slide over the past year, sites that had done nothing wrong by the old rules, and the honest reading is that the old rules stopped describing the situation. Plenty of the practitioner writing on why traditional search traffic is dropping lands in the same place: the pages were fine, the demand was fine, and the layer sitting between the two changed what it needed from them.
Manual process could not re-audit a large site against a moving target fast enough to respond. That is the actual case for automation, and it is a narrower case than the pitch decks suggest.
Where the machines genuinely earn it
The tasks that hand off cleanly share a trait. They are ones where being roughly right across everything beats being exactly right about a few things.
- Crawling a large site and flagging structural problems, broken internal paths, orphaned pages, thin duplicated clusters
- Clustering thousands of queries by intent instead of by string match
- Watching how you get described across different assistants and noticing when a description shifts
- First-pass competitive coverage mapping, who has answered what
- Log analysis and index monitoring at a volume no person will read
Notice these are all observation tasks. Wide, repetitive, dull, and genuinely improved by not being done by a tired human on a Friday afternoon.
Where handing it over quietly hurts
The failures are subtler and they take a quarter or two to show up.
Generated content at volume tends to produce pages that are individually acceptable and collectively pointless. Everything is covered. Nothing has a position. You end up with a library where every book says the same thing in a different order, and there is no reason for anything to cite you over anyone else.
The second failure is source quality. What gets pulled into answers keeps skewing toward material that has some claim on being trusted, and the reading of the sources AI engines now trust that seems to hold up is that the mix keeps moving toward places with visible people, real experience and actual opinions behind them. Mass-produced coverage optimises for the wrong side of that.
Third, and least discussed, is that automated recommendations arrive without context about your business. A tool will tell you to consolidate two pages. It does not know one of them is the only thing your enterprise buyers read before a renewal call.
The split that works
The version that holds up in practice is unglamorous. Machines do the looking. People do the deciding.
Automation runs the sweep, surfaces the anomalies, tracks the descriptions, watches the crawl. A person reads that output, throws out the two-thirds that do not matter, and spends the recovered time on the parts that were always the actual work: what your company knows that nobody else does, which questions you can answer with more honesty than competitors will risk, what your positioning should be when a system has to summarise you in one sentence.
You can tell when a team has the split wrong. The tell is that they are busy and their output is interchangeable with everyone else’s.
What the argument was really about
The people insisting nothing changed are usually protecting a process they are good at. The people insisting everything changed are usually selling something.
Underneath both, the durable observation is that the cost of doing the mechanical parts fell close to zero, and when that happens to any craft, the value moves to whatever the machine still cannot do. Here that means judgement, source credibility, and having a point of view worth quoting.
Which is roughly where the smarter practitioners were spending their time anyway, back when the grind was still mandatory.
