How AI Assistants Choose Their Sources (and Why It Matters)

When an assistant recommends a hotel, a tour or a destination, it is not consulting a ranked index the way a search engine does. It retrieves a small number of documents in the moment, extracts the facts it needs, and writes an answer. Which documents it retrieves decides what it says. Two large citation studies now let us see that choice in some detail.

The retrieval mechanism

Retrieval favours pages that are easy to fetch, easy to parse and consistent with other pages. A page with structured data, plain factual sentences, a stable identity and frequent updates is a low-risk source; a page built for mood and imagery, or one that disagrees with three other pages, is a high-risk one. Assistants also weight agreement: when several sources say the same thing, that version wins regardless of whether it is the most current.

None of this is about brand. It is about the properties of the page.

What the data shows

Profound’s analysis of 680 million citations across the major assistants found retrieval heavily concentrated in a small set of community, reference and review domains. On Perplexity, Reddit alone accounted for close to half of the top-ten citation share.² Each assistant has its own habits, but all of them lean on a short list of domains that are structured, frequently updated and consistently formatted.

For travel, Nicolas Sitter’s AI Hotel Landscape 2026 provides the most detailed picture. Across almost 20,000 hotel answers in 25 cities:¹

• Perplexity cited Tripadvisor in 95.5 per cent of answers.
• Gemini cited Booking.com in 63 per cent of answers.
• Perplexity routed 25.3 per cent of its links to online travel agencies and metasearch sites.
• Independent three-star hotels drew 28.8 per cent of their citations from OTAs.

The hotel’s own website sat among the least-cited sources. The organisation being recommended was, in most cases, not the source of the recommendation or of the facts attached to it.

Why the organisation’s own site loses

Three reasons recur.

It is not quotable. Marketing copy written for atmosphere — “a relaxed base for exploring the region” — gives an assistant nothing to extract. A listing that says “12 minutes on foot from the station; breakfast from 6.30” does.

It disagrees with itself. The same fact, stated differently on the homepage, the booking page and a downloadable PDF, reads to a retrieval system as uncertainty. The third-party listing that states it once, plainly, wins the tie.

It is not the freshest-looking copy. A page without dates, structured data or consistent identifiers looks no newer than a directory entry from years ago. Retrieval cannot prefer what it cannot detect.

The consequence: recommended, then handed away

The commercial effect is that an assistant can recommend an organisation and, in the same answer, send the traveller to a reseller, because the reseller’s page was the clearer statement of the facts. Marti-Ochoa and colleagues found only two of five assistants disclosing their sources at all,³ so from the traveller’s side the hand-off is invisible.

What to do about it

The order of work matters. Correct the facts in the sources assistants already cite before publishing anything new, because new content on top of conflicting sources amplifies the conflict. Then make the canonical page the clearest, freshest statement of each fact on the web — plain text, structured data, consistent names and identifiers, honest modified dates. Only then extend reach.

Organisations that do this in sequence tend to see the assistant’s citations shift towards them. Organisations that skip to step three tend to see the same third-party answers, more often.