AI search optimization: how to get cited by ChatGPT, Perplexity and Claude

AI search optimization — also called AI SEO, or GEO in its technical form — means appearing in the answers written by assistants: ChatGPT, Perplexity, Claude, Google's AI Overviews.

This guide is the practical half: what to do, in what order, and how to verify it worked. For the underlying mechanics, see the GEO guide.

Start where the AI looks, not with your site

The first reflex is to optimise your own pages. That is the right reflex in SEO, and the wrong one here.

When an assistant answers, it consults three to eight pages. In our measurements, most are not brand sites. On Normandy campsites, five out of five sources consulted for one question were tourist boards — not a single property read directly. If you are named on none of the pages read, you cannot appear in the answer. Your site can be perfect; it is not in the loop.

What to do: identify the sources actually read in your market, then exist on them. Trade directories, comparison sites, review sites, specialist press, forums, local bodies. That is presence work, not code.

The trap: assuming those sources. They cannot be guessed — from one market to the next they are forums, institutions or merchants. It is the first figure an audit produces.

Aim at the precise question, not the broad keyword

An AI asked vaguely answers vaguely: directories, chains, generic names. Asked precisely, it names precise players.

That is the best news in AI search for a small business: on "campsite in Normandy" you will not make the cut. On "campsite with a heated indoor pool twenty minutes from Mont-Saint-Michel", the room is there — our measurements show independent properties named as soon as the constraint tightens.

What to do: list your buyers' real constraints — use, context, exclusion, proximity, budget, compatibility — and write one answer per constraint. A page that handles a precise case beats ten pages that handle the category.

Write facts, not promises

A language model reproduces what it can attribute. A figure, a date, a compatibility, an area, a duration. It does not reproduce "impeccable service quality".

Concretely, on a product or service page:

• replace every superlative with the data point that justifies it;
• put the question in the heading and answer in the sentence that follows, not three paragraphs later;
• name explicitly what you compare yourself to, and on what criterion;
• date whatever expires.

The technical basics: necessary, insufficient

They will not get you cited, but their absence can exclude you:

• content present in the served HTML, not built only in JavaScript;
• a robots.txt that does not block assistant crawlers — check, many block them unknowingly;
• structured data consistent with the visible text;
• pages that are fast and reachable without authentication.

Verify, do not assume

This is where most efforts fail. Opening ChatGPT, asking one question, looking for your name: it is the natural reflex and it proves nothing. A model is not deterministic — the same question gives three different answers, sometimes naming different brands. The detail of that problem.

A check that stands up requires several questions, several assistants, several passes, and questions that do not contain your name. Which is exactly what a GEO audit does: 40 questions, 3 assistants, 3 passes, 360 measurements, report within 48h.

The order to work in

1. Measure where you stand and which sources are read in your market.
2. Exist on those sources, starting with the two or three most cited.
3. Write the precise answers to your buyers' real constraints.
4. Fix the technical layer if it blocks access.
5. Re-measure a month later to find out whether the order was right.

Step 1 conditions the rest. It is the only one that stops you working blind.