A handful of museums have started letting visitors chat with an AI persona before they ever stand in front of the work. The bot answers questions, sets expectations, and frames the mood a visitor brings into the room. Those pilots are a preview of what is already happening to every gallery and project space that does not run its own chatbot: the first interpretation of a show is written by a model the gallery does not control, and the visitor reads it before deciding whether to come at all.

A collector asking ChatGPT about a mid-career painter's new exhibition gets a confident paragraph of synthesis. A critic doing background on a program gets a tidy summary of the roster. A tourist with two free hours gets a shortlist of three shows worth seeing. The gallery wrote none of those answers, yet each one speaks for it, and most land before a human has opened the press release.

The First Review Is Now a Machine Summary

The old sequence was legible. A gallery sent a release, a critic filed a notice, a listings editor built a weekend guide, and a visitor arrived with a rough idea of the show. Every step had a human in it. Every step was traceable.

That sequence has thinned out. A Harvard Business Review piece on answer engines puts it plainly: large language models now synthesize whatever they can find into a single response, collapsing the research step the visitor used to spread across a dozen tabs. For an exhibition, that means the gallery's bio, a few old reviews, an artist's past wall text, and whatever Reddit says about the roster all get blended into one paragraph the visitor reads before the opening.

The visitor is not consulting the gallery. The visitor is consulting a model that has already consulted everyone else about the gallery. The show begins, in a real sense, in that paragraph.

Auditing what those chatbots currently say about a show is the kind of monthly check a specialist like Digital Marketing can run, so the gallery knows what visitors are reading before they arrive.

The Numbers Behind the Shift Are Not Small

This is still a minority behavior, but it is heading somewhere else fast. Chatbot referral traffic to the open web climbed sharply through 2025, and the pool of destinations those tools send people to has broadened from a narrow set of reference sites to a much wider range of specialist and independent sources. The audience being handed to websites by a chatbot is no longer a rounding error, and the range of sites it hands them to widens every month.

For a gallery, that cuts two ways. More people are arriving pre-briefed by an AI, and the pool of sources the AI pulled from is wider and messier than any press officer would choose. Writing on the shift to answer engines describes a search layer that answers rather than lists, which means the summary a visitor sees is already an argument about what the show is.

What Actually Changes the Answer

Models weight corroboration. A claim that appears on the gallery's own site and nowhere else reads as marketing. A claim echoed by an independent review, a museum catalog, an auction house note, and a university syllabus reads as fact. The work that moves the summary happens off the gallery's own domain, in forms a model can lift cleanly.

• Independent coverage with specifics. A review in a trade publication that names the artist, the medium, the dates, and one sharp line about the work is worth more to the model than a long gallery essay. Specific, attributable sentences in third-party outlets are what the system treats as evidence.


• Clean, machine-readable facts. Exhibition dates, artist biographies, past shows, and a plain-language description of the work should live somewhere a crawler can parse without guessing. If the gallery does not state the dates clearly, the model will infer them, and sometimes it will infer wrong.


• A consistent name graph. The gallery, the artists, the curators, and the works should be spelled and formatted the same way across every surface, from the site to the museum catalog to the auction result. Inconsistency is how a model ends up blending two artists with similar names into one person.


• Audits of what the models currently say. The way to know what visitors are reading before they arrive is to ask the chatbots directly, in the plain phrasing a visitor would use, and record the answers. If the summary is wrong, the fix is upstream, not on the gallery's homepage.

The Exhibition Now Has Two Audiences

Every show is now read by two audiences in sequence. The second is the visitor who walks through the door. The first is a language model deciding what to tell that visitor the show is about.

That first reader does not care about the invitation design or the vinyl on the wall. It cares about what independent sources say the show is, how consistently the gallery's own facts are stated, and whether the artist's history can be reconstructed without guesswork. Galleries that feed the model the way they would feed a careful critic get a first review they can live with. Galleries that ignore it get whatever the internet averages out to, and they get it before the opening.