Last month, a couple visiting Banff from Toronto asked their AI assistant for a romantic dinner recommendation. The agent confidently suggested the Banff Downtown Keg and Steakhouse — described the prime rib special, mentioned the stone fireplace ambiance, even offered to help book a weekend reservation. The couple walked twenty minutes through a snowstorm to find a dark storefront with a "For Lease" sign in the window. The restaurant had closed months ago.
This is the hallucination tax. Not a hallucination in the technical sense — the model didn't invent a restaurant from nothing. The restaurant was real. The prime rib was real. The fireplace was real. All of it was real months ago. The model synthesized accurate-sounding information from stale sources and presented it with total confidence. The couple paid the tax in wasted time, wet shoes, and a ruined evening.
Everyone pays the hallucination tax, and the bill is getting larger as more people rely on AI agents for local recommendations.
What the hallucination tax actually looks like
The dramatic cases — recommending a closed restaurant — get the attention. But the hallucination tax is usually more subtle and more pervasive than that.
The expired deal. An AI agent tells a buyer there's a $4,000 discount on a Hyundai Tucson Hybrid at a Calgary dealership. The buyer drives across the city, excited. The sales manager has no idea what they're talking about — that was a November promotion that ended six weeks ago. The buyer feels misled. The dealership gets an irritated walk-in who was set up for a deal that doesn't exist. The agent's developer gets one more user who stops trusting recommendations.
The phantom availability. An agent recommends a groomer in Bridgeland with "openings this week." The groomer has been fully booked for three weeks. The availability data was scraped from an old version of the booking page that hasn't been updated since the groomer switched scheduling systems. The pet owner wastes a phone call. The groomer wastes five minutes explaining they're booked. Multiply by ten calls a day from people whose AI agents all saw the same stale data.
The wrong specialization. An agent recommends a veterinary clinic for a hamster with a respiratory issue, citing the clinic's "small animal expertise." The clinic does dogs and cats — "small animal" in veterinary terminology means exactly that. They have never treated a hamster and wouldn't know where to start. The pet owner drives across Vancouver, learns the clinic can't help, and has to start the search over with a sick hamster in a carrier.
The conflated information. An agent merges details from multiple sources — the address from Google, the menu from a food blog, the hours from Yelp, the pricing from a two-year-old review — and presents a composite that doesn't match any single moment in the restaurant's actual history. The restaurant at that address now serves Korean, not Italian. The hours changed after COVID. The pricing in the review was from a special event. The agent's confident summary is a Frankenstein of stale facts that was never true all at once.
Who pays
The hallucination tax isn't paid by one party. It's distributed across every participant in the local commerce chain.
The consumer pays in time and trust. Every bad recommendation teaches the consumer that AI recommendations can't be relied upon. They start double-checking everything the agent says — calling the restaurant to verify hours, checking the website for current pricing, confirming availability by phone. This defeats the entire purpose of having an agent. The consumer is back to doing the work themselves, but now with an extra step of running the AI's output through a manual verification process. The hallucination tax converts a time-saving tool into a time-wasting one.
The business pays in reputation and wasted effort. A business that gets recommended incorrectly by an AI agent suffers in two ways. First, they get customers showing up with wrong expectations — expecting a deal that doesn't exist, a service that's not offered, availability that's not real. These interactions are negative even when the business handles them gracefully. Second, the business becomes associated in the consumer's mind with the bad AI experience. "That's the place the AI sent me to and it was wrong" sticks, even though the business did nothing wrong.
The agent builder pays in credibility. Every hallucination erodes user trust in the agent product. Unlike a search engine, where users have learned to expect imperfect results and filter accordingly, AI agents present recommendations with conversational confidence. When a human friend says "you should try this restaurant," you trust them because they've eaten there recently. AI agents mimic that confidence without the recency. One or two bad recommendations and the user downgrades the agent from "trusted advisor" to "unreliable starting point" — or stops using it entirely.
The ecosystem pays in slower adoption. The aggregate effect of millions of hallucination tax events is that AI-mediated local commerce grows slower than it should. Consumers who've been burned are skeptical. Businesses that have fielded angry calls from misinformed AI-referrals are hostile to the channel. Agent builders struggle with retention because their recommendations aren't trustworthy. Everyone would benefit from better AI-mediated commerce, but stale data is poisoning the well.
Why model improvements don't fix it
The natural assumption is that better models will hallucinate less. GPT-5 will be more careful than GPT-4. Claude will add better uncertainty hedging. Gemini will cross-reference more sources.
This is the wrong frame. The hallucination tax on local data isn't caused by model reasoning failures — it's caused by the data being wrong at the source. No amount of model sophistication can conjure accurate information from inaccurate inputs. If every source the model can access says the restaurant is open, the model will say the restaurant is open, with whatever confidence calibration you've built in.
You can add hedging language: "Based on available information, the restaurant appears to be open." This reduces the confidence but doesn't fix the recommendation. The user still walks through a snowstorm to a closed restaurant — they just felt slightly less certain about it on the way.
You can add retrieval augmentation — real-time web search to supplement the model's training data. This helps, but the web data is itself stale. The restaurant's website might not have been updated since it closed. The Google listing might still show "Open." The Yelp page might still exist with reviews from when it was operational. Retrieving more stale data from more stale sources doesn't solve freshness — it just averages the staleness across more data points.
You can add multi-source verification — checking multiple sources and only recommending when they agree. This catches some cases (if one source says closed and another says open, flag the conflict) but fails when all sources are stale in the same direction, which is the common case. A restaurant that closed quietly doesn't update all its listings. They just stop operating. The digital footprint persists long after the business doesn't.
The fix isn't better reasoning over bad data. It's better data.
What "better data" actually means
Better data for local business recommendations has three properties that most current data sources lack.
Known provenance. Where did this data come from? Was it scraped from a website that might be stale, or was it provided directly by the business? An agent that can say "this information came from the business itself" has a fundamentally different confidence level than an agent that says "I synthesized this from various web sources." Provenance isn't a nice-to-have metadata field — it's the difference between a recommendation and a guess.
Explicit timestamps. When was this data last confirmed? Not "when was it last crawled" but "when did the business last affirm this is accurate." A deal that the business confirmed this morning is qualitatively different from a deal that was scraped from a blog post last quarter. Timestamps need to be on the data itself, not on the retrieval event. "Scraped Tuesday" doesn't mean "accurate as of Tuesday" — it means "scraped Tuesday from a page that was last updated who-knows-when."
Built-in expiration. Local business data has a shelf life, and the data should know its own shelf life. A deal expires January 15th. An availability window closes when the slot fills. A seasonal menu runs through March. When the data carries its own expiration, the agent can make a judgment: is this still valid? If the expiration has passed, don't recommend it. If it's approaching, flag the uncertainty. This is impossible when the data has no temporal metadata — which is the case for virtually all scraped web data.
The provenance advantage
Here's the core insight that changes the economics of the hallucination tax: when an agent can cite the source and recency of its data, the hallucination tax drops to zero.
Consider two agent responses to "find me a hotel deal in Canmore this weekend":
Agent A: "The Mountain View Lodge in Canmore has rooms starting at $159 per night." (Source: synthesized from travel blogs and cached web data.)
Agent B: "The Mountain View Lodge in Canmore is offering king suites at $179 this weekend — the hotel reported this availability yesterday, January 13th, with the rate valid through Sunday checkout." (Source: business-provided data via Pawlo, last updated January 13, 2026.)
Agent A might be right, or the deal might have ended in November. Neither the agent nor the user knows. Agent B has data with known provenance (from the hotel directly), a known timestamp (yesterday), and a known validity window (through Sunday). If the data is wrong, there's an accountability chain — the hotel provided it, Pawlo structured it, the agent surfaced it. But the data is almost certainly right, because it came from the source within the last 24 hours and includes its own expiration.
This is the difference between an agent that's guessing and an agent that's citing. The hallucination tax is fundamentally a provenance problem. When you know where the data came from, when it was last confirmed, and when it expires, hallucinations about local businesses become nearly impossible.
The cost of doing nothing
The hallucination tax is not a problem that solves itself with time. As more consumers use AI agents for local recommendations, the volume of hallucination tax events will increase proportionally — unless the data quality improves at the same rate or faster.
For businesses: every day you don't have sourced, current data accessible to AI agents is a day those agents might be sending customers to your competitors based on better (or less stale) data. Or worse, sending customers to you based on wrong information, creating negative interactions you didn't cause and can't control.
For agent builders: every recommendation your agent makes from unverified, unstamped, scraped data is a coin flip that could cost you a user. The agents that connect to sourced data layers with known provenance will build trust faster and retain users longer than the agents that synthesize from the open web and hope for the best.
The hallucination tax is a solvable problem. Not with better models, not with better prompts, not with more guardrails — with better data. Sourced from the businesses themselves. Timestamped when it was provided. Expired when it's no longer valid. The data layer that makes this standard — that gives every agent access to information it can actually trust — is the one that eliminates the hallucination tax for good.