Making your site readable by AIs, citation and agentic navigation
Being cited by a conversational assistant and being usable by a navigating agent are two distinct requirements, one plays out off the site on reliable mentions, the other plays out on the page itself.
A site can appear in a conversational assistant's answer while no agent manages to click its purchase button, and the reverse happens just as often. These are two distinct capabilities, checked by two distinct controls, and confusing one with the other puts the effort in the wrong place.
Citation plays out off the site
Being named in an answer generated by an assistant depends first on what other pages say about the site, not on what its own code displays. An independent study tracked 1885 pages already cited by AIs at the time they added JSON-LD schema markup between August 2025 and March 2026, compared with 4000 control pages that had not added any, each page measured thirty days before and thirty days after the addition, and it measured no significant rise in citations on the conversational engines observed. The markup helps an engine resolve ambiguity about an entity and makes a fact easier to extract, which remains useful, but it does not create the mention itself. The lever that matters is still built elsewhere: reliable third-party pages that cite the content with attribution, work carried out off the site and close to the search and citation optimisation covered in the content pipeline.
Agentic navigation plays out on the page
An agent browsing a page to complete a task asks a different question: can it identify what it is looking at? A button with no text and no accessible label stays invisible to an agent reading the page's structure, even when it is perfectly visible to the eye. This audit is run locally, without depending on a remote assistant.
set -o noclobber
cat > fragment-exemple.html << 'EOF'
<button class="icon-close"></button>
<button aria-label="Fermer la fenĂȘtre">X</button>
<a href="/panier">Voir le panier</a>
EOF
grep -c "aria-label" fragment-exemple.html
On this fabricated fragment, the command returns 1: only one of the two buttons carries a usable label, the first stays mute despite its visual presence. The text link names itself and needs no extra attribute.
Two scopes, two proofs
Adding schema markup in the hope of a rise in citations measures an effect that did not occur in the study cited above. Auditing only button structure, without ever checking who cites the site elsewhere, leaves the citation question entirely open. Checking that a deep page, not just the homepage, remains directly linkable, then checking that its interactive elements carry a usable name, acts on both fronts at once without confusing what each control proves.
Citation by an assistant against agentic navigation
| Requirement | What it measures | Where it plays out | Main lever | Control to verify it |
|---|---|---|---|---|
| Citation by a conversational assistant | Whether an assistant names and attributes the site in answer to a question asked by a user | Off the site, on mentions and citations found elsewhere on the web | Reputation built by reliable third-party pages that cite the content with attribution | Asking several assistants questions related to the site and noting whether it is named |
| Agentic navigation | Whether an agent browsing the page can identify its links and buttons to complete a task | On the page itself, in the structure and labels of its interactive elements | Accessible labels carried by every interactive element, even without visible text | A local audit of the page that counts interactive elements with no usable name |
Schema markup and citations, a study with no measured rise
An online shop adds complete JSON-LD schema markup to its product page. Its add-to-cart and payment buttons carry no accessible label. What can be claimed from this description alone?
A copywriter publishes a product page on her site containing three links and two buttons, then runs a script on this page that lists the interactive elements carrying an accessible label and saves the count obtained to a file.
Write in one sentence what this situation establishes, and in one sentence what it does not establish.
What this establishes: This situation establishes that the page now has a reproducible audit method that lists its interactive elements and saves a verifiable count.
What this does not establish: It does not establish how many elements actually carry an accessible label, nor whether this page is cited by a conversational assistant.
The three most common miscalibrations
- Too broad This situation establishes that the page is fully navigable by an agent and that it will be cited by conversational assistants.
- Too narrow This situation establishes nothing until the file containing the count has been opened and read.
- Beside the point This situation establishes that command-line scripts are more reliable than browser extensions for this type of control.
- Being cited in an assistant's answer depends on reliable mentions found elsewhere on the web, not on the markup placed on the page itself.
- JSON-LD schema markup makes it easier to extract and disambiguate an entity, without the study that measured it detecting a rise in citations.
- An agent navigating a page finds its interactive elements by their accessible label, a button with no text remains invisible to it.
- The agentic navigation audit is run locally on the page's structure, without depending on a remote conversational assistant.
- Every deep page of a site deserves to be directly linkable, the homepage alone is not enough for either requirement.
Open the source code of one of your own site's pages, look for every button or a tag that triggers an action, and note how many carry an aria-label attribute or visible text between the tags rather than an icon alone.
Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.
- Ahrefs, study on schema markup and citation by AIs consultée le 2026-09-02