Tone Is a Lever
Fixing the writing meant getting close enough to the pages to notice the h1 tag was on the wrong element. That's not a coincidence. That's the whole argument for reading your own output.
I said the pages were vanilla slop.
Not the nicest feedback I've ever given, but accurate. Nine vendor pages on BigTechReseller.com, all technically correct, all completely forgettable. No sales psychology. No reason for a reader to remember one vendor over another.
I asked for it fixed.
What I got back was a lot more than a rewrite.
The Seed Was Fine. The Voice Wasn't.
The original brief was straightforward: two new vendors, three new categories, five comparison articles, closing every gap in a competitor-seeding handoff doc.
That part shipped clean. Structurally correct, on time, nothing broken.
And it read like nothing. Generic B2B copy that could describe any vendor in any category. I've seen a thousand pages like that. I wasn't going to publish nine more of them.
Rebuilding the Voice
Fixing tone properly meant more than swapping adjectives.
All nine vendor descriptions got rewritten with actual sales-psychology specificity, not just the two new ones. Opening lines got punched up with real energy instead of throat-clearing. A boxed excerpt treatment went in, reusing the site's existing orange brand color instead of inventing a new visual pattern for the occasion.
Then came the part that made the writing actually functional: a real autolinking function that connects vendor jargon to glossary definitions site-wide, ten new glossary terms to support it, and that same autolinking wired into full article bodies, not just the vendor pages.
The Bugs Were Hiding in the Voice Pass
This is the part worth actually paying attention to.
Wiring autolinking into every article body meant every glossary term suddenly became a live, clickable link everywhere it appeared. That's when a real accuracy problem surfaced: a glossary entry had IBM Power Systems and AS400 treated as synonyms. They're related. They're not the same thing. Nobody noticed while it was inert text. It became impossible to ignore the moment it turned into a link readers would actually click.
Same session, different bug: recoloring every genuine page-title H1 orange, a purely visual request, meant looking closely at which element on the page actually was the H1. Turned out, on vendor.php, it wasn't the vendor name. It was the small muted subtitle underneath it. The visually prominent vendor name sat in a different tag entirely. Backwards from what a screen reader or a search engine would read as the page's real subject.
Neither of those bugs has anything to do with tone. Both got caught because fixing the tone meant somebody had to actually look at the live pages closely enough to notice something else was wrong.
Making It Stick
Fixing nine pages once doesn't prevent page ten from shipping generic again next month.
So the standard got written down as a real document, docs/BigTechReseller-Editorial-Voice-Guide.md, and then made enforceable: a mandatory read-first rule went into the site's CLAUDE.md and AGENTS.md files, the instructions every AI author on this site reads before doing anything. Claude Code picks it up automatically. Codex picks it up automatically. Nobody has to remember to paste the guide into a fresh session and hope it gets followed.
Why Tone Is a Lever
I called this session important because it creates the tone as a lever. One two-day correction, and every future page on that site inherits the standard instead of needing the same 'this is vanilla slop' conversation repeated indefinitely.
But the standard isn't really just about sounding better. Automated verification, JSON checks, PHP lint, a curl sweep confirming every URL loads, caught zero of the three real defects in this campaign. A person reading the actual output for tone caught all three, because catching tone requires the same close attention that catches a backwards h1 tag or a wrong synonym.
Content quality review and technical QA aren't two different jobs. They're the same job, done by someone who's actually looking.
Explore Related Research
Browse our documented case studies, experiments, and systems.