In 1996 I could get a page to the top of AltaVista by putting the keyword in the title, in the first paragraph, and about forty times in white text at the bottom. It worked for roughly a year. Then it didn't, and the people who'd built businesses on it spent the next year explaining to clients why the phone had stopped ringing.

I've watched that cycle repeat maybe eight times since. Directories to crawlers. Meta tags to links. Links to the Florida update, which broke everyone's Christmas in 2003. Panda, which punished thin content. Penguin, which punished bought links. Hummingbird, which stopped matching words and started matching meaning. Mobile-first. The map pack. Each one arrived with a chorus of "SEO is dead" and a fresh crop of tricks that worked for a year.

The assistants are the ninth shift and the largest. I want to write down what I think carries over, because thirty years of watching the referee change has left me with four things I believe more each time.

The trick always dies. The thing it imitated doesn't.

White text at the bottom of the page was imitating relevance. Bought links were imitating reputation. Doorway pages were imitating local presence. Every trick is a cheap counterfeit of something the search engine was actually trying to reward, and every trick dies when the engine gets better at telling the counterfeit from the real thing.

The assistants are the best counterfeit detectors yet, because they don't count signals — they read. A page stuffed with "best plumber Tucson" fourteen times doesn't fool a model that has read a million plumbers' pages and knows what a real one sounds like. What survives is the thing the trick was imitating: actually being a plumber in Tucson, and saying so plainly, in the places a reader would look.

I've already seen the first wave of assistant-era tricks — llms.txt files that promise the moon, "AI-optimised" content that's the same keyword mush with a new label, schema that claims awards nobody won. They'll work for a year. Don't build on them.

The customer moved first. The tool is always catching up.

In 2007 clients told me nobody used their phone to search. In 2012 they told me nobody looked at the map. In both cases the customers had already moved and the businesses were arguing with a trend that was over.

Owners are telling me now that their customers don't use ChatGPT. Some don't. But the customer who does isn't announcing it; she asks the assistant, gets three names, calls one, and the business never learns the referral came from a machine. The absence of evidence in your call log is the whole problem. The assistants are silent about the customers they send elsewhere. Building the meter was, in a real sense, a response to thirty years of watching clients wait for evidence that a channel had arrived until the channel had already been captured.

Legibility beats cleverness, every time.

The businesses that survived every shift were never the cleverest. They were the ones whose site said, clearly and in the right places, exactly what they did and where. That page structure — title, heading, plain description, address, phone — was right for AltaVista, right for Google, right for the map pack, and it is right for the assistants, which read the same fields and reward the same plainness.

What's new is that the assistant also wants the answer, not just the topic. It wants "a water heater replacement in Tucson runs $1,400 to $2,200," not "we offer water heater services." The FAQ page that commits to a number is the 2026 version of the title tag that said what the page was about in 1998. Same principle. The reader can't reward what you won't state.

Measure the thing, not the proxy.

Every era has a proxy that gets mistaken for the goal. Rankings were the proxy for a decade; I've built a whole practice on the observation that rankings are vanity and revenue is sanity. Impressions were a proxy. Domain authority was a proxy so seductive that agencies still sell it.

The assistant era's proxy is already forming: "AI visibility scores" with two decimals, from tools that won't show you the question they asked or the answer they got. I built Duly Noted to show you both, because the only honest measure of whether an assistant recommends you is to ask it and read what it says. That's not sophisticated. It's a meter. The sophistication is in resisting the urge to make it look like more than one.

What I'd tell my 1996 self

Stop with the white text. Write what the business is, where it is, and what it costs. Put it where a reader would look. Then check whether the reader found it, and keep checking, because the reader keeps changing.

Thirty years later, the reader is a language model, and the advice hasn't moved an inch.