Evening. My Claude subscription refills its quotas. Most days they evaporate the way everyone's do — autocomplete, small refactors, a summary here and there. This is a worklog about one day with Fable. And in a way I did not expect, it was touching.
It started, as the best things lately do, with an audit. We run periodic audits on bestaiweb.ai — competitor research, content decay checks, index coverage — and I've learned that their most valuable outputs are never the things we audit for. They're the emergent finds in the margins.
The Morning a Library Burned Down Quietly
This particular audit margin contained something strange: a well-known specialist blog in our niche — hundreds of deeply technical articles, years of accumulated citations — had stopped existing. The company behind it had been acquired, and every single article now redirected to a press release announcing the deal. No archive page, no farewell post. Just a redirect.
Here is the part that matters for anyone who runs a content site: the rest of the web didn't get the memo. Thousands of pages — curated resource lists, course syllabi, personal engineering blogs, open-source project READMEs — still pointed to those articles. Every one of those links was now a small broken promise: a reader clicks expecting a technical explainer and lands on corporate PR.
By dinner that same day, we had ten pull requests open across open-source projects, each one repairing dead links in someone else's repository.
Broken-Link Building, the Honest Version
The technique is older than most SEO tools and it has a name: broken-link building. The web rots constantly — companies get acquired, blogs get shut down, domains expire. Every dead resource leaves holes in other people's websites. If you happen to have living content that covers the same ground, you can offer it as a replacement. The site owner fixes a real defect; you earn a real link. Commercial SEO suites sell exactly this as a feature.
There is a clean version of this technique and a dirty one, and the difference fits in one question: who decides about the link?
The dirty version buys the dead domain itself and redirects its accumulated authority wherever it pleases. Nobody consented to anything. Google's spam policies have explicitly named this "expired domain abuse" since early 2024, and it gets punished accordingly.
The clean version proposes and lets the other side decide. Our rules, written down before sending anything:
- Archive first. The default fix for a dead link is the Internet Archive snapshot of the original article — the reader gets exactly what the link always promised. Our own page is proposed only where it genuinely covers the same ground.
- Disclosure always. Every proposal that contains our link says plainly: we are the authors of the replacement. Dead-link fixes are welcomed; hidden self-promotion is spam, and being caught pretending costs more than a hundred links earn.
- Fix everything, not just what benefits us. If a file contains eight dead links and only two have replacements on our site, the proposal fixes all eight — six to the archive, two to us.
- One ask, no nagging. A proposal is a gift, not a campaign.
The maintainer who clicks "merge" is making an editorial decision about their own resource. That is the most legitimate form a link can take — and it is the entire ethical foundation of the technique. Everything else is procedure.
Where the AI Actually Helped
I won't publish our scripts or scoring thresholds — partly because the specifics are our edge, mostly because they wouldn't transfer anyway. The principles transfer. Four of them did the real work:
Deterministic before subjective. Everything countable was counted by code, not estimated by a model: the inventory of what the dead blog had published (public web archives keep remarkably complete records), the candidate matches against our own live content, the verification that every replacement URL we might offer actually returns a living page. An AI that is allowed to guess numbers will eventually guess wrong with confidence. Ours was only allowed to read numbers.
Judgment exactly where judgment belongs. The one step no script can do: deciding whether our article is a true replacement for a dead one — would a reader who wanted that specific article be satisfied? — or merely a thematic neighbor. That is reading comprehension at scale, which is precisely what a language model is for. It graded every candidate pair, flagged the ones where the automated match had aimed at the wrong target, and I reviewed the result. Most of our strongest replacements turned out to be glossary entries — deep, single-concept pages we'd been quietly building for a half of the year. Neither of us alone would have been both fast and right.
Preflight against reality. Search indexes lie a little. Before a single fix was proposed, the actual files were fetched from the actual repositories — and reality differed from the search results in three separate ways, including dead links that had never existed in the first place (someone had once cited an article that was never published). Every proposal was built from the file as it is, not as the index claimed. This one principle probably saved us from the embarrassing category of "helpful" PRs that don't apply cleanly.
Human gates on everything public. The AI never sent anything. Every outbound action — the wave of pull requests, their exact wording, the decision of which targets to approach at all — waited for an explicit go from a human, in my case twice over, because the plan also went through a team meeting first. Speed is worthless if it outruns consent.
The Part Where a Plan Genuinely Moved Me
At nine in the morning, this technique did not exist in my head. I did not know its name, its ethics, its failure modes, or that our site was sitting on dozens of perfect replacement pages.
By ten, there was a plan document in front of me. Not a wall of jargon — a clear map: what the opportunity was, what was already done, what the next three moves were, what it would cost, where the ethical lines ran, and a one-page explainer I could hand to non-technical colleagues at a meeting that started at ten.
I read it and it genuinely moved me. Not because the machine was fast — fast machines are ordinary now. Because understanding arrived fast. Something that two hours earlier I hadn't known existed was suddenly comprehensible, planned, ethically fenced, and waiting for my decision. The AI had compressed weeks of "read five guides, misunderstand two, ask someone senior" into a morning — and then it stopped and waited for me, at exactly the moments where stopping mattered.
That is the actual promise of this technology in a small content team, and it has nothing to do with replacing anyone. It is the compression of comprehension. The judgment stayed human. It just stopped being slow.
What We Don't Know Yet
Honesty section. As I write this, the pull requests are open, not merged. Bots have emailed me about contributor license agreements (I signed one under the wrong account first — browser sessions are treacherous — and the AI diagnosed that too). Merge rates take days to weeks; search-visibility effects take a month or two to show up in Search Console, and we will measure them rather than declare victory.
We also folded the whole process into a reusable Claude Code skill, so the next time a library in our niche burns down quietly — and in this consolidating market, there will be a next time — the morning-to-dinner pipeline is a command away. The window for these opportunities is short; the teams that move in days are the ones that already wrote down how.
So that's my answer to the question in the title. Tonight the quotas reset again, the tank is full again — and somewhere out there, another library is quietly burning. The web rots. Somebody has to bring flowers that are actually alive.
I'm an IT analyst who works with Claude Code daily on Bestaiweb.ai. Not an SEO strategist. Someone who's fascinated by how AI responds — and envious of the polymath-like breadth it has at its fingertips in a flash: it knew this decade-old technique, its ethics, and its failure modes before I finished asking. So sometimes I stop building things and let the day itself become the experiment. This is what I found. It might be wrong in places. But I love experimenting with AI about AI — and the best experiments are the ones you can't keep to yourself.
Top comments (0)