A pub fact needs a receipt
A five-pub check of sunset claims found a map pin roughly three kilometres wrong. Before AI judges a pub, the underlying facts need to stand up.
Last weekend we tested a simple question across five pubs: can the customer actually get a sunset view?
It sounds like the sort of thing a listing can answer with a tick. It is not. We used the pub coordinates, the direction of the building, the path of the sun and the land around it. Then we held that against the claim already sitting in the listing.
The check did a useful job before it gave us an answer. One pub’s map pin was roughly three kilometres from the building. The listing said one thing, the public map and the estate record said another. If we had trusted the first coordinate, the calculation would have been neat and wrong.
That is the part people miss when they talk about AI judging a business from its data. The model is not the first risk. The fact underneath it is.
A tick is not evidence
“Dog friendly”, “Sunday roast”, “real fire” and “sunset view” are not bits of decoration. They shape where a customer clicks and what they expect when they arrive. Get them wrong and the pub has paid to disappoint somebody before they have crossed the threshold.
The answer is not to turn every facility into a research project. It is to give each claim a route back to something real. A current menu can support Sunday roast. A photograph and a site visit can support a fire. Reviews, the pub’s own pages and a proper location can support a view. When the sources agree, leave the claim alone. When they disagree, flag it. When there is no evidence, do not pretend a blank is a yes.
Joe’s planning note for Project George puts it simply: confirm what is already supported, flag the conflict, then interrogate the gap. It is still planning, which is where it should be. The moment a system starts making or changing claims at scale is exactly the wrong time to discover that its map pins, menus or old listing fields are loose.
Start with the ground beneath the answer
This matters beyond pub listings. Hospitality businesses are filling their systems with useful-sounding facts, then asking AI to recommend, rank and write from them. A model will make a clean answer from a dirty record. It cannot tell you that a coordinate points down the road unless you have built that challenge into the job.
The commercial test is blunt. Could you show a customer, operator or landlord why this claim is there? If the answer is no, it is not a fact yet. It is a guess wearing a badge.
That five-pub sunset check has not proved an estate-wide system. It has done something more useful at this stage. It has shown the order of work. First make sure the pub is where the record says it is. Then decide what the evidence really supports. Only then let a model make a judgement from it.
What leaders should do next
- Pick five customer-facing claims that affect choice, then ask what exact evidence supports each one today.
- Check your location, opening-hours and menu data against an independent source before using it to train, prompt or automate anything.
- Decide which conflicts should stop a claim going live and who has the authority to settle them.
- Keep blanks as blanks until the evidence turns up. A missing fact is cheaper than a confident wrong one.