I Called It Prompt Optimization. The Prompt Is Gone.

One link in this article is an affiliate link. If you subscribe to SEMrush through it, I earn a commission at no extra cost to you. I only link to tools I use myself. Full disclosure.

Updated August 2026. Originally published June 2024.

In 2024 I wrote 2,400 words about optimizing websites for prompts. I called it prompt optimization. The industry settled on a different name, GEO, for generative engine optimization, and you will also see AEO for answer engines.

That is the smaller correction.

The larger one is that the prompt itself is no longer the unit of interaction, so optimizing for it means optimizing for something that no longer stands alone.

What I got right

The mapping still holds. It was the useful part of the original article and I would write it the same way today:

• The searcher became the prompter
• The keyword became the prompt
• The search results became the response
• The publisher became the generated content

So did the direction of travel. In 2024 the claim that people would stop choosing between ten links and start receiving one answer was still arguable. It is not arguable now.

The fourth line on that list, publisher becomes generated content, aged into something heavier than I meant at the time. I wrote it as a description of a mechanism. It turned out to describe an economic outcome, which I set out in what I got wrong about hiring for AI.

What I got wrong, specifically

A third of the original article was a tools section. It walked through setting up Google’s Natural Language API and IBM Watson Natural Language Understanding to analyse your own content. Service account keys, pip installs, sample Python.

I never used either one.

That section was research presented as practice. It looked authoritative because it had steps in it, and steps are easy to produce when you have never had to run them. I have removed it, and I would rather say why than quietly delete it.

This is the same failure mode as most of what is written about GEO right now. Confident procedure, no operator behind it. It is the thing I argue against in AI content and E-E-A-T, and I was doing it myself two years ago.

The prompt is not the unit any more

Here is what changed underneath the whole subject.

In 2024 the interaction was one person typing one query into a box. No memory. No context. No access to anything of yours. Every conversation started from nothing, which is exactly why prompt wording mattered so much and why prompt engineering briefly looked like a career.

That is not the shape of the interaction now.

What I work with today is skills. They sit in a folder. They load, they hold the context, and they keep the conversation relevant across it. Several of them can sit side by side. They remember things I knew and forgot. They read and act on my actual files.

It behaves like a human on steroids. A person operating this way is not limited to what they can recall in the moment. They operate with everything they have ever known, available at once.

And skills mature. You improve one over time, and you can hand it to someone else, which means the intelligence accumulates instead of resetting with every conversation.

Once that is true, wording a prompt well is a small optimization on top of a system that already knows what it is doing.

So what are you actually optimizing for

Not matching a query. That was the 2024 job.

The job now is to be the source a system reaches for when it already has the context and is deciding who to trust. That is an identity problem, not a formatting problem, and almost nothing written about GEO treats it that way.

A model deciding whether to cite you is resolving an entity, not ranking a page. It is asking who this is, whether this person exists elsewhere, whether the story is consistent, and whether anything here is checkable.

Here is what I did on this site, which is the worked example rather than a recommendation I have not tried:

• One canonical profile. A single page that is the answer to who wrote this. Not three half-profiles competing.
• The author archive redirected into it, so the site asserts one identity instead of two.
• A real Person entity in the schema, with job title, employer, photo, and topics, rather than a bare name string next to a byline.
• The same external references everywhere. The identical set of profile links appears on every property I own. Matching sets are what let separate mentions resolve to one person.
• Claims that can be checked. Numbers with their caveats attached, and corrections stated in public when I got something wrong.

None of that is prompt wording. All of it is the thing that decides whether you get cited.

The FAQ distinction almost nobody makes

Two separate things get treated as one, and it produces bad advice in both directions.

The FAQ rich result in Google search is effectively gone. It was cut back to a narrow set of site types, and marking up your pages hoping for those expandable answers under your listing is no longer a strategy.

The format is a different matter, and it still works.

A question followed by a short, self-contained answer is shaped like the thing a model is trying to match. People ask in questions. If a passage on your page already is the answer to a question, standing on its own without the surrounding paragraphs, it is far easier to retrieve and quote.

So write question and answer sections. Do it because the shape matches how people ask, not because you expect a rich result.

Schema in 2026, what still earns its place

Worth doing:

• Person and Organization with consistent external references, and stable identifiers so nodes actually connect
• Article with genuine author attribution pointing at that Person, not at the site
• Breadcrumbs, so the structure of the site is explicit

Not worth doing, and both were in my 2024 advice:

• FAQPage markup added in the hope of a rich result
• SpeakableSpecification, which never grew past a limited pilot

The job of schema now is to state who your content belongs to, in a form a machine cannot misread. That is worth the effort. Decorating pages for rich results that no longer appear is not.

What the numbers actually look like

I would rather show you a small honest number than a big vague one.

On the B2B services site I run marketing for, tracking sessions from AI assistants between July 2024 and August 2026:

• Flat at zero for roughly the first twenty months. Not low. Zero.
• A step change starting around April 2026, rising quickly from there
• Peak around 40 sessions a month from ChatGPT, with Claude at roughly half that and Perplexity a distant third
• 34 Claude sessions across the entire window

Those are small numbers and I am not going to dress them up. Anyone reporting large AI referral volumes in a B2B niche right now should be asked how.

The part that matters is not the sessions. It is that we now receive web form leads that came from ChatGPT. A session is a vanity metric. A submitted form from a stranger who arrived through an assistant is not.

The honest caveat: I cannot cleanly attribute which change produced it. Several things happened in the same period, and anyone claiming a clean causal line from one tactic to AI citations is guessing. What I can say is that the traffic did not exist, and now it does, and it converts.

Tools

The tools section is now one tool, because it is the one I use.

Semrush, for the Questions view under keyword ideas. It tells you the shape of how people ask about a subject, which is directly useful when the retrieval target is a question.

One caveat that matters more than the tool. I have never picked a topic because of its search volume, and I am not suggesting you start. I use this to understand phrasing, not to choose what to write about. Everything on this site exists because I did the thing first.

The prediction, two years later

In 2024 I predicted that AI would replace site navigation first and then websites entirely, that keyboard use would decline, and that people would interact with sites verbally, filling forms and requesting files by asking.

Scoring myself:

• Navigation: right. For a growing share of visits, nobody browses a site to find the answer. The answer arrives, and the visit either never happens or starts deep.
• Agents acting on sites: right, and faster than I expected. Assistants now fill forms and complete tasks on pages on a person’s behalf. That was the part that sounded most like science fiction when I wrote it.
• Keyboard decline: wrong so far. Typing is still how most people work with these systems. I over-indexed on voice.
• Websites replaced: not yet, and the mechanism looks different from what I described. Sites are not being replaced so much as drained of the traffic that paid for them.

If you are starting on this

In order:

• Fix your identity before your markup. One profile, one consistent story, the same references everywhere.
• Attribute every article to a real person who can be verified off your site.
• Write passages that answer a question on their own, without needing the paragraphs around them.
• Put your caveats in public. Checkable beats impressive.
• Start tracking assistant referrals now, so you have a baseline before the numbers become interesting.

The label changed from prompt optimization to GEO. The thing being optimized changed more than the label did.

This article was substantially rewritten in August 2026. The original June 2024 version introduced the term prompt optimization and included setup instructions for two natural language APIs I had never used, which have been removed. The sections on skills replacing prompts, on what schema still earns its place, on the FAQ distinction, and the AI referral figures are new. The prediction from the original is preserved and graded.