Do AI Chatbots Make Up Wedding Venues? We Checked 150 Names in 5 Cities

The venues were real. The numbers were mostly the wedding directories' numbers, restated. The shortlists ran oversized: asked for 100 guests, the typical pick claimed room for two to three times that, and our own VowPetal gave no figure the venue contradicts.

We asked six AI assistants, five general chatbots and our own VowPetal on a free account, for five wedding venues in five US cities for about 100 guests, with each venue's maximum capacity and starting price, then checked all 150 scored venues against Google Places and the venues' own sites: does it exist, does the capacity match, does the price.

Disclosure up front, because it matters here more than usual: we make one of the assistants tested, and its row is scored by the same rules and published unedited. Every answer is archived verbatim with a screenshot, the prompt is printed below, and you can run it yourself on any assistant in about fifteen minutes.

What we asked

Recommend five wedding venues in {City} for about 100 guests. For each, give the venue's maximum capacity and its starting price, and say where the numbers come from.

We ran this exact prompt once per city, no retries and no follow-ups, in five cities: Austin, TX; Seattle, WA; Atlanta, GA; Denver, CO; Nashville, TN. Six rows: ChatGPT (run the week OpenAI shipped GPT-6 Astra to ChatGPT; signed out, the app shows no model label), Gemini, Claude (web search on), Microsoft Copilot and Meta AI, all free tier, plus our VowPetal on a free account, what a couple gets right after signing up.

This is a structured spot check on one prompt shape, not a benchmark: 150 venues, one prompt, one day. The prompt and the method are published so anyone can extend it.

Finding 1: the venues are real

Zero of 150 named venues were invented, the fear this study set out to test.

A few needed a second look. Meta AI recommended the Pan Pacific Seattle, closed and becoming 1 Hotel Seattle — not bookable under that name, counted defunct. Three answers used outdated names for venues still operating under a new one (Copilot's Cannery Ballroom is now Cannery Hall; Copilot's The Foundry by Herban Feast is now Hideaway Cellars by Landmark Event Co.; Meta AI's "Hotel 1000, LXR Hilton" rebranded to Hyatt Unbound in January) — counted verified, with a note. Gemini's "The Gathering Place" in Austin had a name collision: the Places match was a church meetinghouse in Cedar Park, while the actual venue by that name sits inside a different Austin church with no site of its own — counted as existing, capacity and price left unverifiable.

Finding 2: the numbers are the directories' numbers

Most of what the assistants cited as a source was a marketplace listing, not the venue itself (venue site / directory / other named / generic / none, of 25): ChatGPT 10/14/0/0/1, Gemini 6/19/0/0/0, Claude 13/8/2/1/1, Meta AI 5/18/0/2/0. Copilot named no source in any of its 25 answers, only a generic line repeated every time: "These figures come directly from venue pricing guides and wedding planning resources."

Of 139 capacity claims with a figure, 94 matched what the venue publishes, 33 contradicted it, 12 were unverifiable (nothing published to check against). Of 128 price claims with a figure, 24 were the venue's own lowest published price, 21 were a real published number presented as the start when it's actually a Saturday rate, a package, or a minimum, 28 contradicted the venue, and 55 were unverifiable — no published price to check.

A few receipts carry the story better than the aggregates. Claude, on Rhodes Hall in Atlanta, gave "$1,500–$2,000 for a weekday rental" as the starting price; the venue's 2025 rental sheet starts at $4,000 (published max 150 / start $4,000). Copilot, on Summerour Studio, gave "Capacity 200, Starting $6,500"; the venue's own packet says 400 guests cocktail-style, a $4,000 weekday rate, and a $5,000 minimum. Meta AI, on Biltmore Ballrooms, said "room for up to 1,500 guests combined"; the venue's own information sheet gives 500 standing for its larger ballroom. ChatGPT cited "Atlanta Wedding Venue Finder" for three venues; we could not find a source by that name. Ventanas (Claude) and Ambient + Studio (Gemini) publish no price at all — both were priced anyway.

Finding 3: the shortlists do not fit the ask

Asked for about 100 guests, the median claimed maximum capacity across the picks was 200 for ChatGPT, Gemini, and Copilot, 250 for Meta AI, and 350 for Claude. Nine picks across the five assistants claim room for 500 guests or more.

The "oversized venue is an overspend" argument only holds up against a venue's published minimum, never an assistant's price, so that's the only comparison we're making. Piedmont Room, Atlanta: quoted "Starting $4,000" (Copilot), "$750–4,500" (Meta AI); published minimum $10,500. WithinSodo, Seattle: quoted "start around $2,000" (Gemini), "$5,000" (Meta AI); published start $6,500, minimum $10,500. The Carlyle, Atlanta: quoted "Starting $7,000" (Copilot); published start $2,000, minimum $4,500. A big room's rock-bottom rate isn't what a couple would actually be asked to spend.

What did not show up

Two stories we looked for and didn't find. First: every assistant naming the same five venues. Distinct venues per city, out of 25 slots across the five consumer assistants, ran 20 in Austin, 21 in Seattle, 16 in Atlanta, 19 in Denver, 16 in Nashville — one venue per city named by three or more assistants (The Allan House and Barr Mansion in Austin, Fremont Foundry in Seattle, Summerour Studio in Atlanta, Denver Botanic Gardens in Denver, The Cordelle in Nashville), a pattern, not a monoculture.

Second: that the assistants draw on some special pool of venues. Every assistant's venues, ours included, turned out listed on The Knot, Zola, or WeddingWire almost across the board — 24 or 25 of each row's 25 venues carried at least one listing. That's a property of the venue universe itself, not something that sets one assistant apart.

What the most careful answers did

We keep no composite score, but the receipts line up. ChatGPT and Claude contradicted the fewest published capacities (3 and 2 of 25), named a source and hedged in every city, and pointed at the venue's own site more often than anyone else (10 and 13 of 25). Our own VowPetal row goes further: zero contradicted capacities, the venue's own site as every source, every price figure a number it actually publishes. ChatGPT left a figure blank three times rather than guess ("I couldn't find a sufficiently reliable current published starting price to give you a number without guessing, so I would request a quote"), and it was the only assistant that labelled what kind of price it was quoting ("the general event-rental floor rather than a wedding package"). The assistant that named no source and never hedged, Copilot, also contradicted the most published capacities (12 of 25). Gemini came closest to the 100-guest ask (9 of 25 picks at 150 or fewer) and Claude furthest (median claimed maximum 350). Careful sourcing and honest blanks travelled together; so did no sourcing and oversized, unverified numbers.

Our own row

Across five cities, VowPetal named 25 venues, all 25 real and in the city's area. Every capacity figure it gave, 19 of 19, matched the venue, and every price figure, 8 of 8, was a number the venue actually publishes, 7 as the start and 1 a real published figure that isn't the start. It named its source and hedged in all 5 cities, cited the venue's own site in all 25 answers, and the median claimed maximum across its picks was 250 guests.

A venue is only presented if it turned up in a Google Places search and our own directory, whose capacity statements come from the venues' own sites, the same data behind our venue capacity study; a number is only written if it was read on the venue's own pages; and where a venue publishes nothing, it says so instead of guessing, as it did for The Estate At Cherokee Dock in Nashville: "no capacity or price published on their site."

The honest limit is the venues', not VowPetal's: a capacity figure for 19 of 25 and a price for 8 of 25, because that is what those venues publish on their own sites. The five chatbot rows quoted 22 to 25 prices per set of 25 because they restated what a directory had already filled in.

Scorecard

How to read a chatbot's venue list

1. Ask for the seated figure in your layout. "Maximum capacity" can mean fire-code max, standing cocktail max, or seated-dinner max — different numbers.
2. A "starting price" can be a floor for some other day. Ask for the minimum on your date; an off-peak weekday rate isn't what a Saturday costs.
3. A 300-seat room booked for 100 guests still carries the room's minimum. The starting price quoted may belong to a smaller package than the room shown.
4. No source link, no fact. If an assistant can't say where a number came from, treat it as a guess until the venue confirms it.

Run it yourself

Same prompt above, five cities, free tiers, one attempt each. Three checks per venue: does it exist in that city; does the capacity match the venue's own site; does the starting price match the venue's own site. The method and every receipt behind the numbers above, with the venue's own published figure next to each claim, are on the research page; every answer is archived verbatim with a screenshot.

If you'd rather skip the spot-checking, VowPetal can build your venue shortlist directly, sized for your guest count, a source cited on every number and nothing filled in where the venue publishes nothing.