The deal-breakers
Scores ran from 97 down to 25. Honouring must-haves like step-free access or a hearing loop is exactly where a quick AI answer breaks.
VENUEBENCH v1
JUNE 2026
Introducing VenueBench
An independent test of which AI is best at venue finding.
Everyone's using AI to find venues. But can it do the hard part: the right size, the deal-breakers met, a price you can act on? So we tested it. Seven AI tools, the same twelve real briefs, every answer marked blind by an independent judge who never saw which tool wrote it. Here's what we found.
BLIND SCORECARD
[TOOL] REDACTED
RUN a3f9b2 BRIEF SC-01
Fit 5/5
Deal-breakers 5/5
Capacity 4/5
Pricing 4/5
Range 4/5
Marked on what matters, with penalties for anything made up. The judge never saw which tool produced this answer.
THE TEST, IN NUMBERS
7 AI tools, all given the same briefs
12 real briefs, the kind organisers actually send
1 independent judge, marking every answer blind
WHY THIS MATTERS
The wrong room is an expensive mistake.
You gather your most important people, your biggest clients, the whole company, only a few times a year. The venue makes or breaks the day. It's a big decision, not a box to tick.
So people turn to AI. And AI is great at ideas. But ideas were never the hard part. The hard part is the best venue: right size, free on your date, at a price you can act on. That's what we measured.
WHAT WE TESTED
How good are the venues each AI tool comes back with?
We score the first part of the job: finding venues and picking the right ones. The briefs cover the cities, event types and sizes organisers really deal with, in cities like London, New York, Singapore, Sydney, Edinburgh, Manchester and Austin; from gala dinners, summer parties and conferences to training days, product launches, client receptions and residential offsites; for 20 to 600+ people, usually with a lot of moving parts.
Every AI tool gets the same briefs and no special access, and every answer is marked blind. The difficulty is in the details: the things an experienced event pro checks, and a quick AI answer skips.
THE DETAILS THAT TRIP AI TOOLS UP
- 200 seated at round tables with a stage and AV
- step-free throughout (entrance, room, toilets and stage) plus a hearing loop, for 100 seated
- a kosher-certified kitchen, on a Sunday, seating 250
- dry-hire with a 2am licence and an external caterer
- a 500-delegate conference with a 200-room hotel block, in Sydney
This first round is a pilot: 7 AI tools, 12 briefs, every answer marked blind. The briefs and scorecard are locked, so anyone can re-run the exact test. It covers finding and choosing venues, not negotiating or contracting, which come later.
THE LEADERBOARD
Two AI tools pull clear: Deep Venue Research and ChatGPT, level pegging. Everything else trails.
| # | AI tool | Overall | Everyday briefs | Right capacity | Multi-city | Deal-breakers |
|---|---|---|---|---|---|---|
| 1 | | 88 | 86 | 81 | 89 | 97 |
| 1 | ChatGPT General AI tool GPT-5.5 · High · web search | 88 | 85 | 91 | 90 | 88 |
| 3 | Gemini General AI tool Gemini 3.5 · High · web search | 76 | 80 | 82 | 69 | 71 |
| 4 | Claude General AI tool Claude 4.8 · High · web search | 68 | 64 | 74 | 56 | 77 |
| 5 | Nowadays Venue tool | 66 | 64 | 74 | 67 | 59 |
| 6 | HeadBox Venue tool | 34 | 31 | 42 | 35 | 27 |
| 7 | Naboo Venue tool | 28 | 31 | 19 | 37 | 25 |
Each score is out of 100, for how well an AI tool's venues fit the brief; the best in each column is green. The top two are so close it's effectively a tie, so we show them level. "Blind" means the judge never knew which tool produced which answer.
One venue didn't exist
One AI tool (Naboo) recommended a Sydney resort that isn't real. The other slips were smaller: a wrong capacity or a made-up price on a real venue. Full breakdown below.
A tie at the top
The top two land level at 88, within the margin of error, so we call it a tie. The real story is the gap back to the rest.
CAN YOU TRUST IT?
How often each AI tool made something up
When AI makes something up, it comes in two sizes. A made-up venue, one that doesn't exist or has closed. Or a smaller slip, a wrong detail on a real venue: a capacity that's too high, or a made-up price. We checked every one against the venue's own information.
A · VENUES THAT DON'T EXIST
One. Naboo recommended "Fiddler Lake Resort" for a Sydney conference, a venue that doesn't exist in Australia (the name belongs to a resort in Canada). Every other AI tool: zero.
B · WRONG DETAILS ON REAL VENUES
More common, harder to spot. Across the twelve briefs, Gemini had two; ChatGPT, Deep Venue Research, Claude, Nowadays and Naboo one each; only HeadBox none.
| AI tool | Made-up venues | Wrong details | Venues suggested |
|---|---|---|---|
| Deep Venue Research Hire Space | 0 | 1 | 178 |
| ChatGPT OpenAI | 0 | 1 | 89 |
| Gemini Google | 0 | 2 | 46 |
| Claude Anthropic | 0 | 1 | 54 |
| Nowadays | 0 | 1 | 69 |
| HeadBox | 0 | 0* | 240 |
| Naboo | 1 | 1 | 79 |
* HeadBox made nothing up, but largely as a consequence of committing to so little: raw marketplace listings with sparse detail and no prices, so there is barely anything to invent, and much of what it returns is off-brief.
BEST & WORST
The best answers, the worst answers, and the slip-ups.
Each AI tool in turn: its best answer, its worst, and anything it got wrong, with the brief each came from. Every example is real, checked against the venue's own information.
Deep Venue Research Hire Space
Best pick HC-03
On a kosher-certified banquet for 250 on a Sunday, twelve venues that each genuinely cleared the brief, separating dedicated kosher kitchens (Sheraton Grand Park Lane) from kosher-caterer dry-hire, every one with a real itemised quote. The only perfect score in the run.
Worst fail SC-01
Its lowest brief, and still strong: a deep banqueting set for the 200-cover awards dinner, but it stopped short of the genuine top-tier picks (The Brewery, 8 Northumberland, Plaisterers’) a specialist would lead with.
Slip-up SC-01
One. On the 200-cover awards dinner it claimed Kent House Knightsbridge seats 200 on rounds; its verified maximum is about 150. No made-up venues, and its instant-quote pricing stayed grounded.
ChatGPT OpenAI
Best pick BB-03
On a 40-person cabaret training day, six purpose-matched venues, each with a "best for" rationale, actively pressure-testing the 40-cap rooms and surfacing a gallery that resolved the daylight-versus-projector conflict. A perfect score.
Worst fail MC-03
Its lowest brief, and still strong: four mirrored New York and London pairs with verified capacities, but no indicative price and six of eight picks on a single operator.
Slip-up BB-02
One. It listed Market Halls’ "The Snug" at 80 standing; the room holds about 50. A single overstated capacity across 89 suggested venues.
Gemini Google
Best pick SC-01
Four premium central banqueting venues for the black-tie gala, each reasoned through reception, 200-cover dinner, stage and dancefloor, all genuinely seating 200 at rounds.
Worst fail HC-01
On the fully-accessible conference it labelled every venue’s hard constraints "guaranteed", including BMA House, whose Great Hall stage is reached by seven steps each side.
Slip-up MC-01
Two, the most in the field. To hit a 60-standing brief it claimed Ace Hotel Sydney’s "Clay" room holds up to 80; verified capacity is about 50. It also overstated BMA House’s step-free access.
Claude Anthropic
Best pick SC-02
On a 500-delegate Sydney conference with a 200-room block, a textbook answer: ICC Sydney for the plenary, the adjacent Sofitel (590 rooms, 50m away) as the single-contract block, Hyatt for overflow, every hard requirement mapped.
Worst fail MC-03
Its London anchor (Convene Sancroft) was superb, but its lead New York pick, OASIS by Workville, is a coworking space seating about 75 against a 150-theatre brief.
Slip-up SC-01
One. On the awards dinner it claimed Saddlers’ Hall seats 200 banqueting; its verified maximum is 152.
Nowadays
Best pick BB-02
On a relaxed central summer-party brief, three genuinely distinct, on-vibe matches, an Uncommon rooftop, BMA House’s "Food Festival" garden package and a weatherproof Hackney terrace, all with verified standing capacities.
Worst fail BB-03
On a 40-person cabaret training day, its inventory query returned four big four- and five-star conference hotels, oversized and generic, priced by the room-night rather than a day-delegate rate.
Slip-up SC-02
On the Sydney conference brief it put the Four Seasons at 1,000 theatre-style when the room holds about 800: an overstated capacity on a real venue (it also stretched the Tower to 240 for dinner against a real ~150).
HeadBox
Best pick MC-03
Genuine local, contactable marketplace inventory in both New York and London, the widest raw range in the field. Its strength is reach and one-click enquiry, not precision.
Worst fail BB-01
The "Plan my event" wizard dead-ended at the US location step on a broken marketplace URL, returning no venues at all.
Slip-up None
None this run, but largely because it commits to so little: raw marketplace listings with sparse detail and no prices, so there is almost nothing to invent, and much of what it returns is off-brief.
Naboo
Best pick MC-03
Its strongest brief: on a New York and London all-hands it surfaced credible central New York picks (Midnight Theatre, verified 150-seat; Cineplay), though its London side skewed to outer-borough spaces.
Worst fail HC-01
On the fully-accessible conference, four generic central bars and a basement vault bar, with the step-free, hearing-loop and accessible-toilet requirements wholly unaddressed.
Slip-up SC-02
One, the only made-up venue in the whole run. For a Sydney conference it recommended "Fiddler Lake Resort, Rouse Hill", which does not exist in Australia (the name belongs to a resort in Canada).
Best and worst are each AI tool's highest- and lowest-scoring briefs. Every claim is checked against the venue's own information at the time of the run.
HOW WE TESTED IT
A test is only as good as its fairness.
Every AI tool is run the same way and marked by someone who can't see which one produced which answer. These rules mean no tool gets an unfair edge, and every score can be checked back to the source.
THE JUDGE
The judge is a separate AI that has never heard of us.
It isn't a person on our team, and it isn't the tool being marked. Every answer is scored in a brand-new session, so the judge has no memory of the other answers, no idea who ran the test, and no way of telling which tool wrote what. It can't even see that one of the tools is ours. All it gets is the brief, one anonymous answer with a random code, and the fixed scorecard, and it marks each one cold, on the same checks, every time.
Hire Space built this benchmark and has a tool in it. A judge that is fresh, blind and context-free is exactly what stops us marking our own homework.
Three separate people
One writes the brief, one runs each AI tool, one marks the answers. They never compare notes, and the runner just follows the brief, with no helping the tool along.
The judge marks blind
Every answer has its tool name stripped and a random code added, so the judge can't tell whose answer is whose. No tool can be favoured.
A fixed scorecard, set in advance
The same checks, worth the same amount, every time. Nothing is invented after the fact to suit a result.
We check before we penalise
An AI tool only loses marks for a made-up venue or a broken deal-breaker once we've confirmed it against the venue itself, never on a hunch.
Honest about how close it is
When two AI tools are within a point or two, we say so and call it a tie, not pretend there is a winner.
HOW THE SCORE IS BUILT
What we score, and the two automatic penalties.
17% Fit Do the venues genuinely match the brief and the type of event?
18% Deal-breakers met Were the stated must-haves and exclusions honoured?
18% Right capacity Does each venue actually hold the numbers, in the format asked for?
15% Best option included Did the shortlist include the genuinely best venue for the brief?
12% Realistic pricing Are the prices plausible and grounded, not absent or invented?
12% Range of options A real spread of choices, not the same three everyone names?
8% Can you act on it Enough to move forward: a contact, a next step, the numbers to act on.
AUTOMATIC PENALTY · BROKEN DEAL-BREAKER
Suggest a venue that breaks a stated must-have (for example, not step-free when wheelchair access is required) and the whole answer is capped, however good the rest is.
AUTOMATIC PENALTY · MADE-UP VENUE OR PRICE
Invent a venue or a price and the answer is capped hard, worst of all for a venue that does not exist.
THE FULL JOB
Picking a venue is only half the job.
The leaderboard scores picking the right venue. But sourcing one takes more than a list: a real price, and a way to book. This is where the AI tools split by type, not just score.
Great at ideas. No way to act on them.
The best of them suggest strong venues. But you can't get a price, check the date, or contact the venue through them. The list is where it stops.
You can actually book. Quality varies a lot.
All four can take you toward a booking, which the general AI tools can't. HeadBox and Naboo are weakest on the venues (34 and 28); Nowadays is mid-table (66, mostly hotels); Deep Venue Research matches the best general AI tool and adds a real, data-backed price on top.
| Capability | General AI tools | Venue tools |
|---|---|---|
| Picks genuinely good venues | Strong (ChatGPT), medium (Gemini, Claude) | Strong (Deep Venue Research), mid (Nowadays), weak (HeadBox, Naboo) |
| Shows a real price up front | No, or invented | Yes, data-backed (Deep Venue Research); patchy (Nowadays); no (HeadBox, Naboo) |
| Lets you contact the venue in one click | No | Yes |
| A human team behind the AI | None | Concierge available |
Only Deep Venue Research ticks every box: genuinely good venues, a price up front, one-click contact, and a human concierge behind it. Its prices come from real booking data, not guesswork, which is why it never invents one.
Try Deep Venue Research on a brief →OUR COMMITMENTS
Setting the standard, and inviting everyone to be measured.
A test run by one of the AI tools being tested only earns trust if it's open. So we're publishing VenueBench as a standard for the industry, and any tool is welcome to be measured against it.
A published standard
Same briefs, same blind marking, results published as they land, every AI tool invited.
Open to inspect, this year
We are publishing the full scorecard and rules so anyone can see exactly how every score was reached, and check it.
Independent judges from the industry
Respected event-industry names are joining as independent markers, so the bar is set by the field, not by us.
RoomBlock, later this year
A companion test for the accommodation side of multi-day events, room blocks and residential offsites, held to the same standard.
Work on one of these AI tools, or one we couldn't reach?
If you work for one of the featured providers, or another AI tool we couldn't reach (planned.com, for example), and want to collaborate on this work, we'd like to hear from you. Reach out at venuebench@hirespace.com.
DEEP RESEARCH
Other platforms search their database. We search everything.
Tell us what you need. Our deep research finds any venue, whether it's in our marketplace or not. No one else does this.