The AI accountability gap nobody is auditing for

There's a quiet fact buried in this week's Searchable data that ought to be the story, and instead it's being written up as a visibility problem. Two vendor tests, one covering 165 London businesses across 13,365 questions, one covering UK high street retailers across 72,000 questions. In the London set, 93% of businesses had at least one basic fact wrong or missing in AI answers. In the retail set, one in ten answers put a business in the wrong postcode even when the prompt named the town. Sixty-four percent of high-street retailers had at least one false fact returned about them.

The industry is reading this as a GEO problem. It isn't. It's a liability problem the SEO industry is not equipped to talk about and the AI vendors have no incentive to solve.

Here's what I mean. If Google Business Profile listed the wrong postcode for a client's shop, there is a path. You edit the profile. If a directory listed the wrong opening hours, you contact them. If a review site claimed you were closed, you dispute. The path might be tedious and the response times might be terrible, but the correction surface exists, and behind it there is a legal entity that can be held to account. Data protection regimes assume this. Consumer protection regimes assume this. Local trading standards assume this.

None of that scaffolding exists for what ChatGPT is telling a customer about your business right now.

The correction surface has vanished

When an AI system tells a prospective customer that your business is closed when it isn't, or lists services you don't offer, or places you in the wrong town, there is no dispute button. There is no support form. There is no ombudsman. There is, in most cases, no way to even know it happened.

The generative layer has skipped that step entirely.

This is a genuinely new state of affairs. Every previous discovery layer — Yellow Pages, Google Maps, Facebook, TripAdvisor, review aggregators — evolved a correction mechanism because the alternative was legally untenable. If a directory publishes false trading information about a business, the business has recourse. That recourse is why the directories eventually built the dispute forms.

The generative layer has skipped that step entirely. And unlike a directory listing, the "publication" happens once per query, uniquely, to one user, and vanishes. There's no artefact to point at. No archived record. No screenshot the business owner can wave in front of a lawyer, because the business owner wasn't there when the answer was generated.

The dispute infrastructure hasn't been built yet, because the industry hasn't been forced to build it.

Why the SEO framing understates this

The trade press is treating the Searchable data as an argument for stronger GEO practice — more structured data, more citation earning, more content, better entity consolidation. All of that is true and I've written a fair amount about it. But it frames a systemic issue as a marketing task.

A retailer whose postcode is wrong in one in ten AI answers is not primarily suffering a visibility problem. They're suffering a misrepresentation problem, and the practical consequences run through customer service, complaints, refunds, wasted deliveries, missed appointments, staff time, and — eventually — reputational damage that no amount of schema markup will unwind. If a plumber gets called out to the wrong address because ChatGPT gave a punter the wrong postcode, the plumber pays for the wasted trip. Not OpenAI.

The reason this reads as an SEO story is that SEO is the only discipline currently paying attention. Nobody in compliance is watching. Nobody in legal is watching. Nobody in customer service is watching, because the errors surface as customer confusion long after the AI answer has evaporated.

That's the loop. And we built it.

Three overlapping problems, not one

It helps to separate what's actually going on, because "AI gets things wrong" is doing a lot of work as a catch-all.

There is, first, the training-data staleness problem. Models were trained on data that reflected a moment in time and may never have contained accurate structured information about a given small business. Retraining costs money and the increments are lumpy, so errors persist for months.

There is, second, the retrieval-and-synthesis problem. Even when the model has access to fresh web data at query time, it synthesises across sources and confidently averages between them. If your Google profile says one thing and a stale Yelp scrape says another, the model may hallucinate a third answer that appears in neither.

There is, third, the confidence problem. Traditional search hedges implicitly — a map pack shows three options, a snippet quotes a source, a review shows a date. AI answers strip all of this. What used to look like data with provenance now looks like an assertion from an authority. The reader has been trained by the interface to treat the answer as fact.

Each of these has a different technical fix. What they share is that none of the fixes are the business owner's job, and yet all of the consequences land on the business owner's desk.

What businesses can actually do this week

I want to be honest about the limits here, because there is no clean answer.

The first practical step is monitoring, and monitoring is annoying but doable. Take the ten questions a real customer would ask about your business — opening hours, address, services offered, whether you're open on Sunday, whether you take walk-ins, whether you deliver — and run them through ChatGPT, Gemini, and Perplexity monthly. Do it from a fresh session. Write down what each system says. If you have more than one location, do it per location. This is the poor person's citation monitor and it's what most of the tools being sold at premium prices are quietly wrapping.

The second step is entity consolidation. The measurement problem I've written about elsewhere applies here too: you can't fix what you can't see, but you can materially reduce the surface area for hallucination by making sure the same structured data appears consistently across your website's LocalBusiness schema, your Google Business Profile, your Companies House record, your Bing Places listing, and any major directories. Models synthesise across sources. Fewer contradictory sources means fewer inventive answers.

The third step is the one nobody wants to hear. Assume the errors are happening, budget for their cost, and put a process in your customer service workflow for identifying them. When a customer arrives confused about hours, or asks about a service you don't offer, or turns up at the wrong site — log it. Not because you can bill anyone for it, but because in twelve months you will want data on how often this actually costs you, and because the industry pressure that will eventually force AI vendors to build a correction surface will only build if businesses can point at aggregated harm.

Where the counterargument lands

The strongest opposing view is that this is a transitional problem. The models will improve. Retrieval will get fresher. Provenance UI will mature — ChatGPT already shows source citations for many answers, Perplexity is built around them, Google's AI Mode links to underlying sites. In two or three years, the argument goes, error rates will drop and correction mechanisms will emerge organically because the vendors need trust to grow.

I think this is partly right and mostly wishful. The vendors' incentive is to appear authoritative, not to be correctable. The current architecture — synthesise, present, move on — is what makes the interface feel magical. Introducing a dispute surface introduces friction, doubt, and legal exposure. That's not something a fast-growth product team volunteers for. It gets legislated in, usually badly, usually late. Ask anyone who worked in adtech through GDPR.

And even if the vendors do build correction infrastructure, the training/retrieval architecture means corrections propagate slowly and inconsistently. Fixing your Google Business Profile updates Google's map pack in hours. Fixing whatever OpenAI is doing with your business data may take a retraining cycle. That's a different order of latency, and it's a durable structural feature of how these systems work, not a transitional bug.

The honest limits

I'm making a claim about liability and accountability infrastructure, not a legal argument. I'm not a lawyer and I don't know whether current UK consumer protection or data protection law extends cleanly to model outputs — that's a genuinely open question and I'd bet the first serious test cases are two to three years away.

The Searchable numbers are vendor-run and directional, not peer-reviewed. Sixty-four percent is a striking figure but it's one methodology across one moment in time, and error rates will vary wildly by business size, category, and how well-consolidated the entity data already is. A national retailer with a strong Wikipedia entry and consistent structured data across every directory is a different problem from an independent café whose Companies House record is the only public artefact.

And I'm ignoring the upside case, which is real. AI answers are also generating enquiries and traffic that the previous search paradigm might not have delivered at all. The businesses being misrepresented are, in many cases, also the businesses being discovered. The net is probably positive for many, negative for some, and impossible to measure at the individual level with today's tools.

The frame that matters

What I want people in this industry to take from the Searchable data is not a new list of GEO tactics. It's the recognition that a category of business risk has quietly emerged for which no discipline has clear ownership, no monitoring infrastructure exists at scale, and no correction mechanism has been built. That's a structural gap and it's going to sit there until either the vendors are forced to close it, the regulators legislate a close, or businesses start pricing the risk into how they work.

Meanwhile, the SEO industry is going to keep selling AI visibility audits, and most of them are going to be worth having, and none of them are going to solve the problem I've just described. Be honest with clients about what you're selling them. It's a defensive posture, not a fix.

The interesting question isn't how to earn more AI citations. It's what your business does when the citations you earn aren't accurate — and who, if anyone, is on the hook when they aren't.