22 years of SEO, and the one rule I still use every week
I launched a freelance marketplace in Japan in 2004 and it's still running. In 22 years I have never bought an ad. I went all the way into SEO instead, and organic search has carried the whole business ever since.
In that time I've watched every kind of SEO advice arrive and leave. One method has never stopped working, and there is nothing clever about it.
Take the keyword you want, open all ten results on the first page, and read them properly rather than skimming. Then write the piece that covers everything those ten cover between them, and add whatever you know from your own experience that none of them could have written. That's the whole rule.
It's plain enough that people keep looking past it, but look at what it does. If your page genuinely covers more than any of the ten currently ranking, then on the one question Google claims to be answering, yours is the best result. You don't have to guess at the algorithm, you've built the thing it says it's looking for.
It's slow. It took two years before I hit number one for SOHO, which is short for small office home office and was the word people searched here at the time. Crowdsourcing didn't exist as a term yet. Most weeks inside those two years looked like nothing was happening.
I use AI heavily now. On one of my sites AI writes 200 articles a day. That site did 14,000 sessions in March and 90,000 last month. The number gets misread, so let me be exact about what I'm doing. I never ask a model to write an article about a keyword, because that gets you the same page everyone else's model produces and Google has no reason to prefer it. I ask for the rule above: read the ten pages currently ranking for this term, in detail, then write something that beats all ten. Same tool, different instruction, and what comes out is not the same thing at all.
It's also why the argument about AI content and SEO keeps going in circles. Google has never said you can't use AI. What gets caught under scaled content abuse is pages that exist mainly to rank, and the policy doesn't ask who typed them. A page that beats the top ten isn't in that category. A page that restates them is, however many you publish.
22 years, and that's the method. What are you running into when you have AI write your articles?
90k sessions is the interesting signal, but traffic alone doesn't validate the AI publishing model. Is the newer site seeing comparable growth in leads or revenue, or has the traffic-to-business-value relationship changed as output scaled?
It went the other way, if anything. Sessions are about 6 times what they were in March and ad revenue over the same months is more than 20 times, so the value per session went up rather than down as output scaled.
The number that convinced me is partnership enquiries, which have gone up a lot in the same period. Those come from people who arrived, read enough to decide the site is worth dealing with, and then wrote to us, which is harder to fake than sessions.
The size of it surprised me too, to be honest.
That partnership signal is much stronger than traffic alone. If you’re open to it, what’s the best email to reach you on?
Hello Aryan, I hope you don't mind the message. Could you kindly let me know the best email address to contact you?
You can reach me at hello@beryxa.com.
Hard-earned agreement here. We picked a semi-generic name for our product and it taught me the other half of this rule: even a great page can't win a search that doesn't map to your category. People search for the problem, not your made-up category. Took us months to figure out why "the best page" wasn't ranking — nobody typed our words. Now before writing anything I check whether the exact words in people's heads match the words on the page. That one check beats most on-page SEO advice.
Here's the flow we run for every product, in case it's useful.
It starts with one sentence saying what the thing does as an action, not what features it has. The name comes out of that sentence. One of ours stands in for you while you're away, which gave us Kagemusha, the body double a warlord sends in his place.
Then a 30 minute gate on the day we name it. Search the exact name for existing software in the same category and drop the candidate if anything comes back at all. Check it doesn't start with the same two letters as our other products, so it stays unique when you type it into a launcher. One trademark screen, class 9 word mark. The domain has to be free, .app for Mac apps and .com for SaaS. Then say it out loud and check what it sounds like in Japanese as well as English.
The category search kills more candidates than trademarks and domains put together. The heavier legal work waits until just before the first public release, because renaming costs nothing before that and is close to impossible after, since changing the bundle ID resets every permission your users granted.
We stop when a round turns up no new reason to reject anything, rather than after a set number of rounds.
Your point is the half that naming can't solve though. The name is for people who already know you. The pages are for the words in someone's head before they've heard of you. Trying to make one word do both is where it goes wrong.