- An answer engine only cites pages it has first found in an index, then managed to read.
- What matters must be written in the HTML: a price or opening hours displayed by a script may escape them.
- You can refuse model training without closing the door to the robots that search and cite.
An answer rather than a list of links
For a long time, searching the internet meant scanning a list of links, then choosing which one to open. With ChatGPT, Perplexity, Copilot, or the generated answers Google sometimes shows above its results, the question receives a written answer, in a few sentences. For many questions, these engines consult the web at the moment the question is asked, read a few pages, then write their answer, citing their sources with links. This is sometimes called ‘GEO’, for Generative Engine Optimisation: the art of becoming one of those sources.
For a shop, a practice or a craftsman, the change is real, but it erases nothing. The contest now takes place on two levels: appearing in the classic results, and being among the few pages an answer cites. An answer cites only a handful of sources, and the reader does not always need to click, since the essentials have already been given. Being cited therefore means being named at the very moment someone takes an interest in your trade, even if they never visit your site.
How an answer engine chooses its sources
The exact workings of each engine are not public, and they change often. The general principle, however, is known, and it is enough to tell you where to direct your efforts. It usually unfolds in four stages.
- The question becomes searches
The engine rephrases the question as one or more queries, often more precise than the user's own. ‘A good florist near Lille station’ becomes searches about the trade, the neighbourhood and the reviews.
- It queries an index
These queries go to an index, that is, the vast catalogue of pages a robot has already crawled. Google relies on its own; Copilot relies on Bing's, which also serves, in part, other engines.
- It reads a few pages
Among the results, it opens a handful of pages and extracts what answers the question. A slow page, or one unreadable without scripts, risks being skipped; a page that turns its robot away cannot be read at all.
- It writes and cites
It composes its answer and links all or some of its statements to the pages they came from.
Two consequences follow. First, a page that is not in the index cannot be cited, however good it may be: the first task remains search optimisation, which we describe in our search discipline. Second, being found is not enough: you must be readable by a hurried robot, which takes what it understands at first glance and moves on to the next page.
What they actually read on a page
A web page arrives in two stages. The server first sends the HTML, that is, the text of the page with its structure; then the browser runs the JavaScript, the programs that complete, animate or display certain content. Many of these engines' robots, however, read the HTML as it arrives, without running the JavaScript: do not count on them to wait for the end of the show.
The difference can be considerable. A price loaded by a script, opening hours displayed by a widget, the name of a dish in a menu that only opens on a click: all of this can exist for the visitor and be missing for the robot. The extreme case is a site built entirely in JavaScript, whose original HTML is little more than an empty placeholder. For an engine that reads only that HTML, such a page has nothing to say.
What to check
- The key information (name, address, telephone, opening hours, prices, area served) is written as text in the HTML, not only in an image or by a script.
- The content of tabs and collapsed questions is present in the page as soon as it arrives, even if it only shows on a click.
- The page stays fast and light: what speeds up display for a visitor also makes reading easier for a robot (see our article on speed).
The simplest test is within your reach. In your browser, display the page's source code (often with a right click, ‘View page source’) and look for your telephone number or opening hours. If they are not there, a robot that reads only the HTML will not see them either.
Writing to be cited
An answer engine looks for a sentence that answers. It therefore finds what it needs in pages that ask a question and answer it directly, in one or two clear sentences, before going into detail. This is the opposite of a home page that speaks of excellence and passion without ever saying what you do, where, for whom and at what price. Precision is citation's best ally: a verifiable fact can be quoted, a vague phrase cannot.
- Answer your customers' real questions, the ones you are asked on the telephone or at the counter, in a questions-and-answers page or throughout your pages.
- Begin every answer with the answer: ‘Yes, we deliver from Tuesday to Saturday, to the neighbouring towns’, then the details.
- Give precise, dated facts: your opening hours, your ‘from’ prices, your lead times, the date a page was last updated.
- Devote a page to each important service, rather than one long page that mixes everything: a page that deals with a single subject is easier to cite.
The page title and its description matter here as much as anywhere: they are often the first lines an engine reads to decide whether a page deserves opening. We devote a whole article to them, title, description, H1. None of this is written ‘for robots’: a page that clearly answers a customer's questions is, first of all, a better page for that customer.
A precise fact gets cited; a vague promise never does.
The same information everywhere, and mentions elsewhere
An answer engine does not rely on a single page. It cross-checks what your site, your Google business profile, the directories you appear in, and sometimes articles about you, all say. If your site announces you are closed on Mondays and your profile says you are open, it will hesitate, or keep the wrong information. Consistency is not a detail: it is what allows a machine to describe you without getting it wrong.
Structured data helps remove ambiguity. It consists of a few lines of code, invisible to the visitor and written in the JSON-LD format, which explicitly tell a robot: here is a business (Organization), here is a local establishment with its address and opening hours (LocalBusiness), here are questions and their answers (FAQPage). It does not replace the text of the page, it confirms it, and must therefore say exactly the same thing.
Finally, a business described only by its own site is rarely cited. Mentions on other sites, professional directories, the local press, partners' or suppliers' pages, give the engine a second voice that confirms the first. The simplest one to maintain is your Google profile: see our article on the Google business profile.
Opening the door to the right robots
Every engine sends out its robots, and every site can tell them, in a small file placed at its root, robots.txt, what they are allowed to crawl. Two families must be told apart there, often bearing the same brand name. Training robots collect pages to teach models how to write; search or answer robots crawl the web to find the sources an answer will cite.
- Training:
GPTBotfor OpenAI,ClaudeBotfor Anthropic, andGoogle-Extended, which governs the use of your pages by Google's Gemini models, not your ranking in Search. - Search and answers:
OAI-SearchBotfor ChatGPT search,PerplexityBotfor Perplexity, andBingbot, whose index feeds Copilot. - Google: its generated answers rely on the
Googlebotindex, the same one used for classic search; blocking it means disappearing from both.
You can therefore refuse training while remaining findable: two lines such as User-agent: GPTBot then Disallow: / keep OpenAI's training robot away without touching its search robot. The choice is yours; it simply deserves to be made with full knowledge. Bear in mind, too, that this file is an instruction, which reputable robots respect, not a lock.
It remains to be present in the indexes. On Google's side, Search Console lets you check that your pages are known and crawled. For Bing, whose index feeds Copilot and partly serves other engines, the equivalent is called Bing Webmaster Tools. The IndexNow protocol also lets you notify Bing, and the engines that adopt it, that a page has just been created or changed, rather than waiting for its robot to come by.
What the llms.txt file is worth
A newcomer is much talked about: llms.txt. It is a proposal published in 2024: a summary of the site, written in Markdown (a very simple text format), placed at its root, to help language models grasp the essentials without crawling every page. But to date, in October 2026, no major engine has publicly committed to reading it.
Should you write one? It costs almost nothing, on one condition: that it says exactly the same thing as the site. We have one, generated automatically from the same sources as our pages, so that it cannot contradict them. A forgotten file, announcing old prices or a discontinued service, would do more harm than having none.
Measuring, and what cannot be promised
Knowing whether these engines send you visitors is possible, in part. Links cited by ChatGPT often add the parameter utm_source=chatgpt.com to the address, which your audience measurement tool files among your traffic sources (with your visitors' consent, as explained in our article on cookies). Other visits show up as coming from perplexity.ai or from Copilot. None of these traces is exhaustive: many readers read the answer without clicking, and some visits arrive without saying where they come from.
The most telling measurement remains manual. Ask these engines yourself, regularly, the questions your customers would ask (‘caterer for a wedding in Arras’, ‘employment lawyer in Brussels’) and note which sources are cited. Do it in several tools, a few days apart. You will soon see whether your name appears, and which pages are preferred to yours, which often reveals what yours is missing.
Finally, the limits must be stated honestly. Nothing guarantees being cited: answers vary from one question to another, from one day to the next, and sometimes from one user to another. Anyone who promises you a place in ChatGPT's answers is promising what nobody controls. What one can do is bring the conditions together, as in a moss garden: you do not order the moss to grow, you prepare the stone and the shade where it will take hold.
What you can do this week
Tick as you go: nothing is saved, it is simply a reminder.
Founder and director of Digital Air Media, a digital creation house opened in Paris in 2017. He writes here what the house applies to its own pages before offering it to others.