Cvent says 94% of hotels never appear in AI answers. The company publishing the number also sells the listing tier that fixes it.

Cvent, the venue sourcing platform, reported this month that 94% of hotels fail to appear in AI search results, drawn against a survey of 1,650 event planners that found 75% already using AI during sourcing, most often to find and select venues. Alongside it the company published an analysis of nearly 6,000 listings on its own supplier network showing that properties on its highest tier are 50% more likely to be cited by AI assistants than properties on its basic tier.

This piece is for independent hoteliers and commercial managers deciding whether to spend money on AI visibility this quarter. The August version of this article treated the invisibility gap as settled and the cause as thin websites. The measurements of that gap disagree with each other by a factor of nearly three, and the explanation circulating for why hotels lose is not supported by any published documentation.

The two published measurements of the same gap are three times apart

Our own research in April put the share of hotels appearing when an AI assistant is asked where to stay at 16%. Cvent's figure for venue listings cited in corporate event queries is 6%.

Both cannot be describing the same thing, and neither is wrong. The queries differ, since a corporate event brief carries capacity, layout and audiovisual requirements that a leisure stay query does not. The samples differ. The dates differ by five months, which is long enough for retrieval behaviour to change. What the spread tells an operator is that no reliable industry number for AI visibility exists yet, and any vendor quoting one to three significant figures is quoting a single test.

Cvent has not published the method behind the 94%, which matters more than the usual methodology complaint because of who benefits from it.

The company that measured the gap also sells the tier that correlates with closing it

The 50% citation advantage Cvent reports for its top listing tier is the commercially interesting finding, and it is a correlation inside a product the company charges for.

There are at least two readings. Properties buy the higher tier and the richer listing then gets retrieved more often, which is the reading the finding is presented to support. Or the properties that buy the highest tier are the ones with active commercial teams, complete data and more third-party coverage everywhere else, and the listing tier is a marker of that rather than a cause. Cvent's published analysis does not separate them, and separating them would require holding property quality constant across tiers.

None of that makes the underlying advice wrong. It makes the number a sales figure until somebody outside the company reproduces it.

The cost-cap explanation circulating for why chains win is not documented anywhere

The mechanism now being repeated in trade commentary is that assistants are billed for each web search, that developers cap each answer at five to ten searches, and that assistants therefore fall back on chains and online travel agencies because those need fewer searches to verify.

The first part is documented. Anthropic's published pricing for web search on the Claude API is $10 per 1,000 searches on top of token costs, each search counts as one use regardless of how many results come back, and the retrieved results themselves are billed as input tokens. Search genuinely costs money and long research answers genuinely cost more.

The second part is not. The same documentation shows the number of searches is set by whoever builds the application, through a parameter called max_uses, rather than fixed by the model provider. No published figure for a five-to-ten cap in any consumer assistant exists, and no published analysis links a search budget to which hotels get named. The cost is real and the causal chain from cost to chain preference is somebody's hypothesis presented in reporting voice.

That distinction decides where a hotelier should spend. If the cause were a search budget, the fix would be ranking higher on Google so an assistant reaches you inside the budget. If the cause is what the retrieval layer can parse, the fix is the detail itself, and those are different pieces of work with different owners.

The detail that gets retrieved is operational, not promotional

Where the published guidance converges, and where it matches what we saw in our own testing, is on specificity. An events page saying the property offers flexible spaces gives a retrieval system nothing to match against a brief. A page giving the room name, the floor area, capacity by layout for theatre, boardroom, U-shape and cabaret, the audiovisual equipment included, the accessible entrance and the walking time from the main station gives it something to match against every constraint in the query.

The same holds outside meetings. Assistants summarise the corpus of text that exists about a property, and for a small hotel the single largest contributor to that corpus is the hotel. That cuts both ways, since errors in a property's own description travel into the answers alongside the useful detail, and so do the omissions.

What the evidence does not support is a deadline

Nothing published establishes how much business has actually moved. Cvent's survey measures planners using AI during sourcing, which is a different rung from bookings created, and neither figure has been tied to a change in request-for-proposal volume at any named property.

Some operators are treating that gap as a reason to act cheaply and immediately, rewriting the meetings and property pages with operational specifics because the work costs nothing but time and improves the pages for humans regardless of what any assistant does with them. Others are declining to buy listing upgrades or visibility audits until somebody outside the vendors publishes a replicable measurement, on the reasoning that a market where the two available numbers differ by a factor of three is not one to spend into yet.

The test that would settle it for a single property takes an afternoon and costs nothing. Put a brief through three assistants with the constraints a real planner would use, record which properties come back and what each assistant says about yours, and repeat it a month later. That produces a measurement for the one property whose visibility the operator actually controls, which is more than any published figure currently offers.