How to get cited by AI answers, and why French is the easier win

Search has quietly split into two products. One still hands you ten links. The other hands you an answer and names three or four sources. Everything you know about ranking applies to the first. The second rewards something else, and it is worth understanding before your competitors do.

The good news: unlike most things sold under an acronym, this is concrete work with a countable outcome. You can measure whether it worked this afternoon.

Step 1: check that the crawlers can reach you at all

This sounds too basic to be the first step. It is the first step because it is where most sites silently fail, and because no amount of content work matters until it passes.

Open your live robots.txt in a browser: yourdomain.com/robots.txt. Read what is actually served, not what is in your repository. Look for GPTBot (OpenAI), ClaudeBot (Anthropic), PerplexityBot, Google-Extended and CCBot. If they are disallowed, you are invisible to AI answers by configuration, and nothing else on this page applies yet.

Step 2: write things that can actually be quoted

Models lift sentences that stand on their own. Marketing copy is built to do the opposite: it qualifies, it hedges, it flows. That makes most of your site unquotable, no matter how well written.

The pattern in the right column never changes. Three ingredients, every time:

• A number, not a qualifier.
• A named source, so the model can attribute it and a reader can verify it.
• A date, so the claim stays checkable as it ages.

There is a side effect worth naming: this discipline makes you honest. You cannot write a sentence in that shape without having measured something. Most pages fail the test not because they are badly written but because nobody ran the measurement.

Step 3: structure it so a machine can lift it

One idea per section. Headings that describe the answer rather than tease it. Real FAQ blocks answering questions people actually ask, in their words. Tables for anything comparative, because tables survive extraction intact where prose does not.

Then schema.org, and here is the part that trips up competent teams: valid means reparsed by a script, not eyeballed. Several types have rules that are enforced silently.

• ItemList: every itemListElement needs a complete item object with a concrete @type, a name and a url. Entries without one are ignored, with no error surfaced to you.
• BreadcrumbList: every link needs item except the last, which is the current page.
• FAQPage: the answer text must match what a visitor actually sees on the page.
• Escaping: escape < as \u003c inside JSON-LD, or a stray </script> in your content silently breaks the whole block.

Step 4: publish llms.txt

A plain-text summary at the root of your domain describing what you do and what you can be quoted on. It is not a formal standard, it costs almost nothing, and the sites doing it properly are noticeably over-represented among those getting cited.

Generate it from your content at build time rather than maintaining it by hand, otherwise it goes stale within a quarter and starts contradicting your pages. This site publishes llms.txt and a full version at llms-full.txt; open them and copy the shape.

Step 5: measure it, do not infer it

This is where most AI visibility work stops being honest, so it deserves plain language.

The measurement is unglamorous and takes ten minutes:

1. Write down the ten questions a buyer asks before choosing you. Their words, not your positioning.
2. Ask all ten in ChatGPT, Perplexity, Gemini and Google AI answers. In French if France matters, since the answers differ substantially by language.
3. Count how many name you, and note which competitors appear instead.
4. Record the date next to the numbers.
5. Repeat monthly. Individual answers fluctuate, so the trend is the signal and any single run is noise.

Why French is the cheaper win right now

A model answering in English chooses from an enormous pile of credible English sources. Answering the same question in French, it chooses from a much smaller one. Fewer French pages state a given fact cleanly, with a figure and a date, in a structure worth lifting.

That asymmetry is a real, temporary advantage. Publishing genuinely citable French content is cheap today and will not be once every French competitor has worked this out. If you sell into France, this is the rare case where being early beats being big.

It also compounds with ordinary French SEO, because both need the same foundation: content that is genuinely French rather than translated. If your French pages were translated, start there instead, with why translation never ranks.

What nobody can promise

AI answers are non-deterministic. The same question can return different sources on two consecutive runs, and every model update reshuffles preferences. Anyone selling guaranteed ChatGPT placement is selling something they do not control, and you should treat the guarantee as information about them.

What is controllable: being fetchable, being quotable, being structured, and being measured over time. That is the honest scope, and it is enough to move the trend.