The request always arrives in the same shape: "our competitors come up in ChatGPT, we don't". The answer starts with a distinction most articles skip — there are two paths by which a name reaches a generated answer, and they are not optimised the same way.
The two paths into a generated answer
The first path is the model's memory. The model was trained on a snapshot of the web; it "knows" a company exists because enough text mentioned it. That path is slow, collective and outside your direct control: you do not enter it by editing your site, but by being mentioned elsewhere — press, directories, forums, comparison pages. Nothing you change today alters a model that has already been trained.
The second path is live retrieval. The assistant runs a search, downloads a handful of pages, reads what it finds and builds the answer out of that text — citing the URLs it actually opened. That is the path you can act on this week, and it is the subject of the rest of this article.
Why the top Google result is not the cited source
A search engine ranks documents for a human who is about to click. An answer engine assembles text for a human who probably will not. The two systems therefore reward different properties. Ranking rewards relevance and domain authority; citation additionally rewards extractability: a short, self-contained, attributable statement that answers exactly the question asked and can be quoted without distortion.
This is why a long, well-written, well-ranked page can go permanently uncited: if the answer to "what is the cancellation window?" is dissolved into a paragraph of narration, the model will prefer the competitor who wrote "Free cancellation up to 48 hours before arrival." — because that sentence can be quoted without risk.
Google's public documentation on AI features points the same way: search fundamentals still apply, and content has to bring something of its own rather than restate what everyone already says. In other words: no shortcut, plus one extra requirement.
The six conditions for a citation
1. The page must be readable without JavaScript
Many agents fetch raw HTML and run no scripts at all. If your content is injected client-side, they receive an empty shell. The test is one command, and the result is often a shock:
curl -sS -A "Mozilla/5.0 (compatible; test)" https://your-domain.com/ \
| sed -e 's/<script[^>]*>.*<\/script>//g' -e 's/<[^>]*>/ /g' \
| tr -s ' \n' ' ' | wc -w
# Fewer than ~300 words on a page that shows 1,500 in a browser:
# the content is rendered by JavaScript and the agent never sees it.2. The entity must be unambiguously identifiable
A model is reluctant to cite a company it cannot place. Exact legal name, full address, phone in international format, founding year, area served: these must appear both in visible text and in structured data, and be identical in both places. A contradiction between the JSON-LD and the footer is a reason for caution, and caution means no citation.
3. Content must be dense with facts, not adjectives
"Quality service, a passionate team" carries no extractable information. "On site within 4 hours in Casablanca, 7 days a week, fixed quote before work starts" carries three. Countable facts — numbers, deadlines, scope, conditions — are what an answer engine can reuse. The rest is filler, however well it reads.
4. Structure must match the shape of questions
People no longer type "cheap riad marrakech" at an assistant; they write a full sentence. Your H2 and H3 headings should look like those sentences, and the answer must fit in the two or three lines immediately below the heading. The pattern that works is simple: question as heading, direct answer as the first paragraph, detail afterwards. The reverse — context first, answer at the bottom — is precisely what loses the citation.
5. The information must be corroborable elsewhere
A model that cites you is staking its reliability. A claim it also finds on your business listing, in a sector directory and in a press article becomes safe to reuse. A claim that exists only on your own site stays a declaration. This is the slow part of the work, and the part your competitors neglect most.
6. Freshness must be visible and true
Time-sensitive questions — prices, opening hours, availability, regulation — favour recent sources. A displayed update date, an accurate dateModified in the JSON-LD and a sitemap whose lastmod values match reality are worth more than a redesign. One warning though: a date changed without a content change is a structured lie, and that is exactly the kind of signal that turns against you.
What does not work
| Practice | Why it is sold | Why it fails |
|---|---|---|
| Stuffing "AI" keywords | 2010s SEO reflex | The model reads meaning, not occurrences; the page reads as generic |
| Hidden text for robots only | A gain with no rewrite | Cloaked content: actionable on the search side, inert on the agent side |
| Buying an "AI visibility score" | A number is reassuring | Every tool has its own rubric; a score is not a citation |
| Publishing 40 generated articles | Volume equals coverage | Commodity content: exactly what Google's documentation waves off |
| Adding llms.txt and stopping there | Easy, visible, fashionable | The file steers readers; it does not make your page readable or factual |
Measuring honestly
This is where most engagements become unverifiable. Three measurements are real; everything else is an opinion rendered as a chart.
- Your server logs. Human-triggered agent visits (
ChatGPT-User,Perplexity-User,Claude-User) are timestamped, countable and uncontestable. Their growth is the signal closest to a citation. - Referrers. A visitor arriving from
chatgpt.comorperplexity.aishows up in your analytics: that is a click on a citation, the most direct proof available. - A frozen question panel. Write twenty real customer questions, ask the same assistants every month, record who gets cited. Answers vary between runs: this is a trend across months, never an exact measurement — and anyone who hands you that number without the caveat is selling precision they do not have.
The order to work in
By return on time invested, for a site starting from scratch:
- Week 1 — readability. Check what an agent receives without JavaScript, fix the empty pages, make sure robots.txt is not forbidding what you want cited.
- Week 2 — identity. A complete Organization or LocalBusiness JSON-LD, consistent with the visible text, same NAP everywhere.
- Weeks 3–4 — the answers. Rewrite the ten pages that matter in question/answer form, with numbers where the adjectives were.
- Ongoing — corroboration. Listings, serious directories, local press, partnerships: this is what moves the first path, the model's memory.