A traveller opens an AI assistant and types: “Long weekend within three hours of Melbourne, toddler and a dog, somewhere with a good bakery and a short walk.” Thirty seconds later they have a shortlist of three towns, two places to stay and a route. They have not visited a destination website. They may never visit one. Whether your region made that shortlist depended on what the assistant could find, read and trust about it at the moment the question was asked.

That scene is no longer unusual, and the evidence behind it is worth a board paper.

What the numbers say

Google’s AI Overviews, the generated summaries that now sit above many results, were reaching two billion people a month by mid-2025, according to figures Alphabet’s chief executive gave on the company’s July 2025 earnings call (TechCrunch, July 2025⁠ (opens in a new tab)). Google’s conversational AI Mode had passed 100 million people a month in the United States and India by the same date.

The effect on clicks is measurable. Pew Research Center tracked the real browsing of 900 US adults in March 2025 and found that when an AI summary appeared, people clicked a traditional result on 8% of visits, against 15% when there was no summary. Roughly one in five searches in the study produced a summary, and people who saw one were more likely to end their session there (Pew Research Center, July 2025⁠ (opens in a new tab)).

Zero-click behaviour predates generative answers. SparkToro and Datos found that about 58% of US Google searches in 2024 ended without a click to the open web (SparkToro, 2024 zero-click study⁠ (opens in a new tab)). Gartner forecast in early 2024 that traditional search engine volume would decline significantly as AI chatbots and virtual agents absorb queries (Gartner, February 2024⁠ (opens in a new tab)). Analysts’ figures vary, and the direction matters more than any single number.

Australia is at an earlier point on the curve. A Tourism & Transport Forum survey of 2,011 Australians in late 2024 found that 7% would use AI to plan and book their next holiday, rising to 12% among under-35s (TTF, January 2025⁠ (opens in a new tab)). Tourism Australia’s Tourism 2035 strategy, released in August 2026, lists “marketing to humans and machines” as a priority: making sure its marketing reaches the algorithms and AI systems that now recommend destinations, not only the people reading them (Tourism Australia, August 2026⁠ (opens in a new tab)).

What changes for a destination

For twenty years the destination website’s job was to rank, attract a visit and turn that visit into a referral to an Operator. Each of those steps still matters. But a new step now sits in front of them: an assistant reads your content and decides whether, and how, to use it in an answer the traveller may act on without ever seeing your page.

That creates three new ways to lose.

  • Absence. The assistant can’t reach or parse your pages, so your region drops out of the shortlist and a neighbour’s takes the slot.
  • Error. The assistant finds you but reads stale opening hours, an old price or a tour that stopped running. It presents that as fact, in your destination’s name.
  • Flattening. Your content is generic enough that the answer about your region sounds like the answer about any region. The traveller picks on distance or price.

The response to all three is the same: the destination has to be legible to machines as well as persuasive to people. The two goals agree: clear structure, specific facts and current data are what a good editor wants anyway.

How an assistant builds an answer

It helps to separate two things that are often confused.

Training is what happens when a model is built. Crawlers such as OpenAI’s GPTBot and Anthropic’s ClaudeBot gather web content for future models, and Google’s Google-Extended control governs whether content Google already crawls may be used to train its Gemini models. You can allow or block each of these in robots.txt. Google states that Google-Extended has no effect on a site’s inclusion in Google Search and is not a ranking signal (Google Search Central⁠ (opens in a new tab)).

Retrieval is what happens when a traveller asks a question. Search-oriented agents fetch pages at that moment to ground the answer: OAI-SearchBot for ChatGPT search (OpenAI⁠ (opens in a new tab)), Claude-SearchBot and Claude-User (Anthropic⁠ (opens in a new tab)), and PerplexityBot (Perplexity⁠ (opens in a new tab)). Block these and you remove your destination from answers, not from training.

A DMO can reasonably take a position on training. It should think hard before blocking retrieval, and it should check that an over-cautious security setting hasn’t done so already.

What the research says works

A Princeton-led study presented at KDD 2024 tested nine ways of rewriting source content and measured how often a generative engine cited it. The strongest gains came when sources were cited and quotations and statistics were added, with visibility improvements of up to about 40%. Old-fashioned keyword stuffing did not help (Aggarwal et al., “GEO: Generative Engine Optimization”⁠ (opens in a new tab)).

Translated for a destination: write content that states things. “Ninety minutes from the airport, 14 cellar doors, two with restaurants open on Mondays” is more citable than “a wine region waiting to be discovered”. Attribute facts to a source. Keep Operator details current. Answer the questions people actually ask, in the shape they ask them.

Five things a DMO can do this quarter

  1. See what the machines see. Fetch a key page as plain text, and look through your server logs for the agent names above. If the agents never arrive, or arrive and leave within a second, you have your first finding.
  2. Fix the facts first. Opening hours, prices, booking links and seasonal closures carry your destination’s credibility into every answer. If your Operator Products are synced from ATDW, make sure the sync runs often and any content your team has added alongside them is current.
  3. Structure for questions. Rewrite the pages that answer common planning questions as question-and-answer content with FAQ structured data, so both people and assistants can parse them.
  4. Publish a map for assistants. An llms.txt file and clean markdown versions of key pages cost little and signal that your content is meant to be used. We explain what llms.txt is, and who reads it today, in a companion article.
  5. Measure referrals from assistants. Add chatgpt.com, perplexity.ai, claude.ai, copilot.microsoft.com and gemini.google.com as a channel in your analytics, and count requests for /llms.txt. Report it to the board alongside organic search, so the trend is visible before it is urgent.

Where a platform helps

Much of this is infrastructure more than marketing, which is why it suits a shared platform better than a one-off project. On Roam, AI Discovery is live on every site at no extra cost: each site automatically generates and maintains an llms.txt file and serves clean, structured markdown when an AI agent visits. Improvements roll out platform-wide without developer work. How AI Discovery works.

A word of caution

Nobody can promise a destination a place in an AI answer, and anyone who does is selling something other than results. The systems change monthly, they weigh sources differently, and they are not transparent about why. The shift affects every destination site at once.

The honest position is the one that has always applied to search: make your content the most useful, accurate and specific account of your region that exists, make it easy for machines to read, and measure what happens.

If you would like to see how your destination’s site reads to an AI assistant today, book a demo and we will check it with you.