More and more people no longer search: they ask, and get back a paragraph with two or three names inside it. Being on that list takes no new channel and no new tools, but it does take writing differently and letting in crawlers plenty of sites blocked without realising. What changes, what stays the same, and how to measure it.
More and more people no longer search: they ask. They type the whole question — "how much does a website cost in Costa Rica and what does it include?" — into ChatGPT, into Gemini or into Google itself, and get back a paragraph with the answer and two or three names inside it. If your business is not among those names, it hardly matters that you rank fifth in the list of links further down: that person already has what they came for.
The mention is the new first position
A generated answer has no ten results. It has a block of text and, beside it, two or three cited sources. That changes two things at once: fewer clicks happen overall, and the person who does click arrives far further along, because they have already read a recommendation with your name in it.
The good news is that this is not a separate channel with separate tools. Models that answer with current information do it by searching the web at that moment: if a search engine cannot find you, neither can the model. Almost everything you were already doing still counts. What changes is the shape of the content and a couple of technical decisions hardly anybody has revisited yet.
Models cite fragments, not pages
A classic search engine ranks whole pages and leaves you to find the part you need. An answering system does the opposite: it retrieves chunks of text from many sites and assembles a paragraph out of them. The winner is not the longest article but the fragment that explains itself.
In practice that turns into concrete habits when writing:
- The answer first and the reasoning after, not the other way round.
- One subheading per real question, written as the question rather than as a slogan.
- Paragraphs that make sense without the previous one: no "as we said above".
- No filler like "in this section we will look at", which sits exactly where the fact should be.
Concrete detail: numbers, dates, places
Between two texts saying the same thing, the one that ends up cited is the one that commits to a fact. "Affordable prices" cannot be quoted; a range in colones, with what it covers, can. "We serve the whole country" cannot be quoted; the list of cantons you genuinely reach can.
It is the same test a journalist applies when picking a source, and it is why sites that speak in generalities — the ones that could belong to any company — vanish from the answers even when they rank well in the list of links.
A model cannot cite what it cannot read
Before worrying about the writing, check that the writing arrives. The reasons a site gets left out are nearly always the same: the content is assembled by JavaScript in the browser and the HTML served comes back empty; the site is so slow the crawler gives up; the important information sits inside an image; or the subheadings are bold paragraphs instead of real headings.
All of that was SEO before and still is. The difference is that a search engine had the patience to come back tomorrow, and an answer being generated right now does not.
Decide which crawlers you let in
AI systems use their own crawlers, separate from the search ones, and they are controlled from the site's robots.txt. The ones that matter today are GPTBot and OAI-SearchBot from OpenAI, ClaudeBot from Anthropic, PerplexityBot, and Google-Extended for Gemini.
It is worth looking carefully, because in 2024 a lot of people blocked them outright, back when the conversation was about training and nobody was talking about customers. Today, for a business that wants to be recommended, blocking them is asking not to be named. One nuance worth knowing: Google-Extended does not decide whether you appear in Google — Googlebot still does that — only whether Gemini may use your content when answering.
Every so often the advice to publish an llms.txt file summarising the site for models comes back around. It is a proposal that no major provider has confirmed using so far: adding it takes ten minutes and does no harm, but it is not what will move the result. What moves it is the content and the crawling.
What others say about you weighs as much as what you say
For a local business, the model does not build its answer from your site alone: it uses the Google profile, the directories, the reviews, the articles that mention you. Your site says you are good; everything else is what backs that up, and without backing the model would rather name somebody else.
That is why consistency of name, phone number and address across all those sources matters more than ever. If three places give three different phone numbers, the system does not pick one: it picks another company rather than risk a wrong detail.
Answer the questions the way they are asked
People ask a model in full sentences, with the whole doubt inside, not in the two or three loose words they used to type into a search box. A site with a page — or at least a subheading — for each of those real questions has material to be quoted; one organised into "Services", "About" and "Contact" does not.
Structured data helps disambiguate who you are and what each page offers, even though FAQ rich results are barely shown these days. It does not rank you on its own: it tells the machine what each thing is, which is exactly what a model needs in order to cite you without getting it wrong.
How to tell whether it is working
There is no dashboard for this yet, so it gets measured with three simple things:
- Ask the models yourself, once a month, what a customer would ask, and write down who gets named. It is the most direct measurement available today.
- Check in your analytics for traffic arriving from chatgpt.com, perplexity.ai or gemini. It tends to be small in volume and good in quality.
- In Search Console, a drop in clicks with impressions holding steady is the classic sign that the answer is being resolved up top, with no click.
What not to do
Do not write for the robot. Filling the site with generated text nobody edited, or hiding paragraphs with instructions aimed at the model, is the 2026 edition of keyword stuffing and it will end the same way. What gets you cited is having something specific to say and saying it so it can be read on its own.
If you like, we can go through your site with this checklist: what is blocked by accident, which pages can be cited as they stand and which ones need rewriting. It is a short conversation and it costs nothing.
Want us to review your site or build a new one?
