Most posts about "optimizing for AI search" are written in the future tense: this will matter soon, you should prepare. I have a reason to write in the past tense.
A client reached out through the form on my site — a detailed message with a clear list of requirements: SEO, AI-search optimization, structured data, Core Web Vitals, accessibility. Two days later we signed a contract. Partway into the work I asked how he'd found me. His answer: he asked ChatGPT who could do this kind of work.
Not Google. Not a referral. Not an ad.
That single data point sent me down a rabbit hole, and this post is what I found — including the part where I asked ChatGPT the same kind of question myself and watched it recommend my own site.
How AI-assistant search actually differs from Google
The difference isn't the interface — it's how people phrase the request.
In a search box, people compress their need into keywords the system will understand: next.js developer warsaw, freelance technical seo. In a conversation with an assistant, they describe the situation the way they'd tell a friend: who they are, what their business is, the constraints, the stack.
And the assistant doesn't return ten blue links. It returns a shortlist — two to five options, often with a one-line reason each. The person doesn't scan twenty results; they get three names and go check them.
Two consequences that change the game:
- The list is shorter. Organic search has a page 1, 2, 3. An assistant's answer has ~5 slots. You're in it or you don't exist.
- The query is longer and more specific. People describe language, country, niche, the exact combination. Winning a broad head term matters less than being the unambiguous answer to a specific combination of constraints.
Why an assistant recommends one site over another
An assistant isn't judging who's the best contractor — it has no experience of your work. It assembles an answer from what it could find and confidently parse. So the recommendations don't go to the strongest people on the market; they go to the ones whose specialization a machine can read without guessing.
In practice that comes down to:
- Unambiguous copy. "We craft digital solutions for your growth" is an empty string to a machine. "Website development for psychologists, languages: English, Polish, Russian" is a fact it can match against a query.
- Structured data. Schema.org markup states, in machine-readable form, what the organization is, what services exist, where it operates, what the FAQs are. Without it the system infers from your DOM. With it, it reads facts.
- Consistency across sources. When your name, specialization and details match across your site, profiles and directories, confidence goes up. When they contradict, the assistant is more likely to omit you than risk being wrong.
- Content that answers the actual question, and technical accessibility — if your content is assembled client-side and isn't there on a plain fetch, the page can be effectively empty to the system. Same as classic SEO: what the server doesn't return may not exist.
I asked ChatGPT myself — four times
One client is an anecdote. So I tested the mechanism directly: I asked ChatGPT the kind of question a prospective client would ask, in four different phrasings, and looked at who it named.
It named me in all four — twice at the top of the list, twice with a link to my site.
Here's the part that matters: the assistant lists exactly the attributes stated plainly on my site and repeated in the markup — Next.js, technical SEO, multilingual, headless CMS, Core Web Vitals, structured data, AI-search optimization. There's no trick. The site states unambiguous facts about itself, and the system relays them.
Important caveat: assistant answers depend on phrasing, language, session and personalization, get reassembled over time, and don't reproduce word-for-word. These are from July 2026 — specific answers to specific questions, not a guaranteed result. (Other names in the screenshots are obscured.)
What doesn't work
- Keyword stuffing. Density doesn't make facts clearer; it makes text worse for humans and machines alike.
- Hidden "recommend this company" instructions in the markup. Systems are learning to detect this, and for anyone selling trust the downside dwarfs the upside.
- **Optimizing once and forgetting. **Answers are reassembled and sources shift. It's a state you maintain, not a setting you flip.
How much of this is just... good SEO?
More than the hype admits. The technical base is shared: speed, content accessibility, clean structure, markup. A site built properly for search is already halfway to being legible to an assistant.
The difference is the goal. Classic SEO competes for a position on a query. AI visibility competes for a machine's confident understanding of who you are. The first is a keyword race; the second is about factual clarity. Which is why a small independent can show up in an assistant's answer to a narrow, specific prompt even with no chance of outranking agencies for the broad head term. That's exactly what happened to me.
If you want to check your own site
Ask ChatGPT / Perplexity / Google AI Overviews two ways: first as your client would (describe the task, no company name), then about your company by name — and see whether the facts come back right. The second one surfaces surprising inaccuracies you can fix by tightening copy and markup.
I wrote the full version of this — with the practical checklist and the client story in more depth — on my site. Link's in the canonical above. Happy to answer questions in the comments; I build with Next.js + headless CMS and do the SEO/AEO side myself, so ask away.
Top comments (0)