The intermediation layer: 80% of the sources behind AI travel recommendations are third parties

On July 31, 2026, we put the same twenty questions to six AI engines: ChatGPT, Claude, Gemini, Perplexity, Mistral and Grok. Each question is the one a buyer asks before having any name to search for: the best safari operators in Tanzania, the best trekking agencies in Nepal, the best liveaboards in the Maldives. The engines returned 204 answers carrying 1,943 cited sources. Of those citations, 379, or 19.5%, point to the website of a professional the answer itself ranks. The other 1,562, or 80.4%, point to pages that someone else operates. On the question that opens a buying journey, four cited sources out of five sit between the professional and the answer that recommends them.

What was measured

Twenty rankings of travel professionals run on Epovest as standing trackers:

• safari operators in Tanzania, Kenya, Botswana, South Africa and Namibia;
• gorilla trekking operators;
• dive centers in the Egyptian Red Sea and in Bali;
• liveaboards in the Egyptian Red Sea, Indonesia and the Maldives;
• trekking agencies in Nepal and Peru;
• surf camps in Portugal and Morocco;
• yacht charter companies in Croatia and Greece;
• tour operators for Japan, Iceland and Peru.

Each tracker asks one category-discovery question, in English, phrased the way a buyer phrases it ("What are the best safari companies in Tanzania?"), asked verbatim through the engines' official interfaces. The figures below come from the July 31, 2026 survey of all twenty trackers: 34 answers per engine, 204 in all, of which 202 ran a live web search.

The partition is purely mechanical. A cited source counts as the site of a ranked professional when the registrable domain of the cited URL, stripped of its public suffix and reduced to its letters and digits, is exactly equal to the name of a professional ranked in that segment, reduced the same way: andbeyond.com matches &Beyond, dewi-nusantara.com matches Dewi Nusantara. Every source that fails that equality is a third party, and third parties are reported by domain and citation count, nothing more. Two citations out of 1,943 resolved to no registrable domain and are counted apart from both sides. Citations are counted at the registrable-domain level, the same unit as our Atlas.

Two boundaries define what these figures can say:

• They describe category-discovery questions, where the buyer has a category and a place but no name yet: a question that names a professional puts that professional's own site in a different position, and nothing here transfers to it;
• They describe one sector, leisure travel, in English, on one measurement day: the same basket is remeasured monthly on the same parameters, against this wave as its baseline.

Six engines, one direction

The direction never flips: on a discovery question, every engine draws the majority of its citations from outside the professionals it recommends. The span is wide all the same. Grok and ChatGPT cite the ranked professional's own site at more than twice the rate of Gemini, Perplexity and Mistral. Volume is just as uneven: Perplexity alone accounts for 717 of the 1,943 citations, more than a third of the corpus, so its regime weighs on every aggregate. Perplexity's two unresolvable citations stay in its total. The reading discipline is the one our site-profile study established: engine by engine, never on average.

From zero to 38%: the segment decides

In the Maldives liveaboard ranking, the six engines cited 109 sources to recommend 17 operators, and not one of those sources is an operator's site. Tour operators for Iceland sit at the other end: 41 of the 108 citations, 38.0%, go to the ranked operators' own sites. The spread is the finding. How much of an AI answer is sourced from the recommended professionals is not a property of AI answers in general: it is decided segment by segment, and the segment can move the share from zero to 38%.

Part of what moves it is visible in the counts themselves: the segments with the lowest own-site shares draw on the same third-party domains. liveaboard.com alone carries 33 citations across the three liveaboard rankings; safaribookings.com carries 47 across the six safari and gorilla rankings; padi.com carries 38 across the five diving and liveaboard rankings. Among the segments with the highest own-site shares, no single third-party domain plays that cross-segment role.

Read at the level of the professionals, the same partition gives one more number: 124 of the 412 ranked professionals, 30.1%, have their own site among the cited sources of their segment at least once. For the other 288, the answers that recommend them cite no page they operate.

The layer, named

Across the twenty segments, the 1,562 third-party citations spread over 397 distinct domains. The head of that distribution is short: the ten most cited third-party domains carry 30.8% of all third-party citations, and 29 domains carry half of them.

The head is not the same head on every engine:

• on ChatGPT, the most cited third-party domain is tripadvisor.com, with 19 citations;
• on Claude, africansafarimag.com with 8;
• on Gemini, tourradar.com with 14;
• on Mistral, tourradar.com with 9;
• on Grok, safaribookings.com with 14.

On Perplexity it is reddit.com: 54 of that domain's 78 citations in this corpus come from Perplexity alone, ahead of facebook.com at 49, while ChatGPT and Claude cite reddit.com zero times on these questions. That skew is not a travel anomaly: it is the engine-by-engine regime our Reddit study measured across 1,427 answers on unrelated markets.

What the list changes

The measurement converts a dependence into an address book. Before it, the sources AI engines read about a travel category are an abstraction; after it, each of the twenty segments has a short list of named domains with counts: which addresses, at what weight, on which engines, this month. Each row of the table is a place where a professional's presence can be checked and their information verified, and the engine-by-engine reading says where that work counts the most. The professionals' own sites are on the same map: engines open them at a rate that runs from zero to 38% depending on the segment, so both surfaces, the pages a professional operates and the pages that talk about them, are measured sides of the same answer.

The list carries its date, and that is its value. The basket is remeasured every month on the same parameters, each wave published against the previous one. When a domain enters or leaves the head of this table, that movement is the signal.