Field Notes · 02

Search Stopped Sending Traffic and Started Sending Answers

I audited fifteen of the sites we maintain to see what an AI crawler actually gets. Nine have no meta description. Two have structured data. The median site ships 93 kilobytes to deliver 331 words.

Patrick Jean-Baptiste · Measured across the sites we maintain

Every few years somebody announces that SEO is dead. They have always been wrong, and they are wrong again now. But something did change, and it is bigger than the usual algorithm update.

Search used to send you people. Now it mostly sends them an answer and keeps them. Roughly six in ten searches end without a click. When Google puts an AI summary at the top, people click a normal blue link about eight percent of the time instead of fifteen. Ranking first is worth roughly half what it was.

Now you may be wondering: so why bother at all? Because the traffic did not vanish, it changed shape. The answer at the top of the page was assembled from somewhere. Somebody got quoted. The question stopped being can I rank and became am I the source it trusts enough to quote.

So I went and measured it

Opinions about this are cheap. I have a server full of real sites, so I ran the audit on fifteen live client sites, fetching each one the way a crawler would: raw HTML, no JavaScript executed, nothing rendered.

Read that last pair together, because it is the finding that stopped me.

Ninety-eight percent of what you ship is not your words

331 words is about 1.8 kilobytes of text. That text arrives inside a 93 kilobyte page. Everything else is markup, wrappers, builder scaffolding and inline styles.

For a human on a fast connection that is merely wasteful. For a machine trying to work out what your business does, it is signal buried in noise. You are asking it to find two percent.

For comparison, using the identical method: the article you are reading is 1,584 words in 22 kilobytes. Forty percent of that payload is the writing. Not because I am clever, but because there is no builder in the way.

Now go a level deeper

One site in the audit returned eleven words of text inside fifty-eight kilobytes of HTML. It is not a broken site. It looks fine in a browser. The content is assembled by JavaScript after the page loads.

Most AI crawlers do not run JavaScript. Google renders, eventually, on its own schedule. The others largely take what the server hands them. So to a machine, that homepage is close to blank. Its owner has no idea, because every human who visits sees a perfectly normal website.

That is the part I would want to know if it were my business. Not a ranking position. Whether the thing is legible at all.

The robots.txt question nobody has answered

Zero of fifteen sites say anything about AI crawlers. Not blocked, not allowed. No decision has been made, which means the decision was made by default.

There is a real choice here and it depends on what your site is for:

If you sell services locally, being quoted is the point. A plumber in Palm Beach County wants the machine to know he exists, what he does, where he works and how to reach him. Blocking the crawlers is removing yourself from the phone book.

If your site is the product, an archive, a database, original research, then being read and summarised without a visit is a straight loss, and blocking is defensible.

What is not defensible is not knowing which of those you are. Go and read your own robots.txt. Most people have never opened it.

What actually earns a citation

Having watched what gets quoted and what gets paraphrased into nothing, the pattern is not subtle.

Specific numbers beat adjectives. "Fast, affordable web design" is unquotable. It says nothing a machine can attribute. "331 words inside 93 kilobytes, measured across fifteen sites" is a fact with an owner. That is why this article is full of them.

Say who measured it and how. A claim with a method attached is one a system can stand behind. A claim without one is somebody's marketing.

Structured data is how you speak the language directly. Your address, hours, service area and prices, marked up so nothing has to be inferred. Two of fifteen sites do this. It is the cheapest fix on the list.

Answer one real question per page. Pages that answer a specific question hold up. Pages assembled from keywords get absorbed and forgotten.

And be reachable in HTML. All of the above is worthless if your words only appear after a script runs.

What this means locally

For a small business in Palm Beach County the practical picture is narrower than the panic suggests. Nobody is asking a chatbot to recommend a roofer in Lake Worth and getting a confident answer without a source. The map pack still runs local search, and it runs on your Google Business Profile, your reviews and your proximity, not on your blog.

So the local play has not changed much. Claim the profile, get real reviews, put your address and service area in structured data so the machine does not have to guess. Then make sure the site is legible in raw HTML. That is most of it.

What I would do this week

Open your own site with JavaScript turned off and see whether your words are still there. Read your robots.txt and decide on purpose. Add a meta description if you are one of the nine. Mark up your address, hours and service area. Then write one page that answers one real question with a number in it.

Alright. That is the lesson: stop optimising to be found and start writing to be quoted. The work is not that different. It is just more honest, because a machine cannot be flattered by adjectives and neither, in the end, can a customer.