A traveller in Montreal asks her AI agent to find a pet-friendly hotel in Banff for next weekend. The agent searches, finds a property listed as "pet-friendly," and recommends it. The traveller books three nights, drives nine hours with her seventy-pound golden retriever, and arrives to discover the hotel's pet policy allows dogs under twenty-five pounds only. The booking is non-refundable. The nearest hotel with availability that accepts large dogs is forty minutes away in Canmore.
The data wasn't wrong. The hotel is pet-friendly. The data was incomplete — it lacked the context that would have turned a technically accurate fact into an actually useful recommendation. And the absence of context didn't just produce a bad recommendation. It produced a nine-hour drive, a wasted booking, a furious customer, and a one-star review for a hotel that did nothing wrong except fail to specify a weight limit.
This is the context problem, and it's the single biggest source of bad AI recommendations in local commerce.
Data without context is noise
Consider how much local business data is technically accurate but practically useless without context:
"Open until 10 PM." But it's a restaurant where the kitchen closes at 9:15, last seating is 8:45, and the bar stays open until 10. An agent that recommends this restaurant for a 9:30 dinner is technically within the stated hours and practically wrong. The diner arrives, can't order food, and blames the agent.
"Private dining available." But the private room is a curtained-off corner that seats six, requires a $1,500 minimum spend, and needs to be booked three weeks in advance. An agent that recommends this restaurant for a private dinner for twelve people next Tuesday has matched on "private dining" without any of the context that would make the match useful.
"Free parking." But it's four spots behind the building, first-come-first-served, and they're always taken after 5 PM. An agent that factors in "free parking" as a convenience for an evening visit has used accurate data to make a misleading recommendation.
"Accepts new patients." But the next available appointment is in eight weeks, and the doctor only sees new patients on Wednesday mornings. An agent that recommends this clinic to someone who needs care within a week has matched on a boolean — accepting new patients: yes — that tells half the story.
In every case, the raw data is correct. And in every case, the raw data without context produces a recommendation that fails the consumer, frustrates the business, and undermines trust in the agent.
Why scraped data strips context
The context problem isn't accidental. It's structural, and it's rooted in how most AI agents acquire their data.
Web scraping extracts fragments from pages designed for human browsing. A human reading a hotel's website sees "pet-friendly" on the amenities page and then reads the full pet policy linked at the bottom: dogs under 25 lbs, $50 per night fee, one pet per room, must be crated when unattended. The human processes the boolean and the context together, as the page designer intended.
A scraper extracts "pet-friendly" from the amenities list and moves on. The policy page might be on a separate URL, behind a click, in a PDF, or rendered dynamically in JavaScript. Even sophisticated scraping pipelines that follow links and extract text struggle to associate the constraint ("under 25 lbs") with the attribute ("pet-friendly") in a structured way. The scraper gets the fact. It loses the context. And without the context, the fact is dangerous.
The same pattern repeats across every data point. A restaurant's hours are on one page; the kitchen closing time is mentioned in a blog post from 2023. The parking information is on the "Directions" page; the "first-come-first-served after 5 PM" detail is in a Yelp review response from the owner. The doctor's "accepting new patients" status is on the practice website; the eight-week wait and Wednesday-only intake schedule is mentioned nowhere online — it exists only in the receptionist's knowledge.
Scraping is architecturally incapable of preserving context because the context is scattered, implicit, and often unwritten. The more important the context — the constraints, the qualifiers, the conditions that determine whether a recommendation will actually work — the less likely it is to be on a single scrapable page.
Context transforms data into intelligence
The difference between data and intelligence is context. Data tells you what exists. Intelligence tells you what it means for this specific situation.
Here's what raw data looks like for a veterinary clinic in Calgary:
name: Bow River Veterinary
type: veterinary clinic
hours: Mon-Fri 8am-6pm, Sat 9am-1pm
services: wellness exams, dental, surgery
rating: 4.7 stars (312 reviews)An agent working with this data can tell you the clinic exists, is highly rated, and is open on Saturdays. It cannot tell you whether this clinic is right for your specific animal with your specific need.
Here's what the same clinic looks like with context:
name: Bow River Veterinary
type: veterinary clinic
hours: Mon-Fri 8am-6pm (last appt 5:15pm)
Sat 9am-1pm (wellness only, no surgery)
Emergency: partners with Calgary Animal
Referral & Emergency after hours
species: dogs, cats (no exotics)
specialization: senior pet care, geriatric wellness
protocols, pain management
dental: full dental suite, Dr. Chen handles
complex extractions (15+ years experience)
new_patients: accepting, next available wellness
appointment in 6 business days
(urgent/sick visits: same-day if called
before 10am)
fear_free: certified practice, all staff trained,
separate cat/dog waiting areas
parking: dedicated lot, 12 spots, accessible
entrance on north side
seller_intent: "looking for more senior dog wellness
clients — we have capacity for 8 more
geriatric patients this quarter"Now the agent can make a genuinely intelligent recommendation. A senior dog owner looking for a new vet gets matched to a practice that specializes in exactly what their dog needs, with the specific note about geriatric capacity. A cat owner who's anxious about vet visits gets matched on the fear-free certification and separate waiting areas. Someone with an iguana gets correctly filtered out — no exotics. Someone hoping for a Saturday dental procedure gets correctly told it's wellness only on Saturdays.
Every piece of context prevents a bad match or enables a good one. The Saturday wellness-only note prevents the agent from booking a dental cleaning on Saturday. The 6-day wait for wellness appointments prevents the agent from promising next-day availability. The emergency partner note gives the agent a fallback recommendation for after-hours. The seller intent signal tells the agent exactly who the clinic wants to see.
The cost of missing context
Bad recommendations aren't just inconvenient. They have real costs that cascade through the system.
For the consumer: A wasted trip, a non-refundable booking, a missed appointment, an experience that doesn't match expectations. The traveller who drove nine hours to Banff with a dog that's too heavy. The diner who showed up at 9:30 to a closed kitchen. The patient who waited eight weeks for an appointment they could have gotten elsewhere in three days. Each of these experiences erodes trust — not just in the specific agent, but in AI recommendations generally.
For the business: A frustrated customer who leaves a negative review for something that wasn't the business's fault. The hotel didn't lie about being pet-friendly — the data just lacked the weight limit context. The restaurant didn't misrepresent its hours — the distinction between bar hours and kitchen hours wasn't captured. The business takes the reputational hit for a data quality problem it didn't create and may not even know exists.
For the agent ecosystem: Every bad recommendation makes consumers less likely to trust agent recommendations in the future. Trust is the currency of the agent economy, and bad context is counterfeit. If agents consistently make recommendations that fail in practice — technically accurate but contextually wrong — consumers will revert to doing their own research, and the entire promise of agent-mediated commerce stalls.
The cost isn't theoretical. It's happening right now, in every market, every day. AI agents are recommending businesses based on decontextualized data, and the resulting mismatches are accumulating as bad reviews, wasted trips, and declining trust.
Why context can only come from the source
There's a tempting belief that better AI — more sophisticated models, better reasoning, smarter extraction — can solve the context problem. If the model is smart enough, it can infer that "open until 10 PM" for a restaurant probably means the kitchen closes earlier. If the extraction is sophisticated enough, it can piece together pet policies from multiple pages and reviews.
This is partially true and fundamentally inadequate. A sufficiently smart model can guess that a restaurant's kitchen closes before the bar. It can't know that it closes at 9:15 specifically, that last seating is 8:45, and that the bar menu after 9 includes only cocktails and a cheese board. A model can infer that "pet-friendly" probably has some constraints. It can't know the weight limit is 25 pounds, that there's a $50 nightly fee, and that dogs must be crated.
Context isn't something that can be inferred or extracted. It has to be provided. The business knows its kitchen closes at 9:15 because that's when the line cooks leave. The hotel knows the weight limit is 25 pounds because that's what the building's insurance policy requires. The clinic knows the next available wellness appointment is in six days because that's what's in the scheduling system. This information exists inside the business and nowhere else.
The only reliable way to get context-rich data is to source it directly from the business, in a structured format that preserves the qualifiers, constraints, and conditions that make the data actionable. Not scraped fragments reassembled by inference. Not booleans stripped of their fine print. Sourced intelligence — data provided by the people who know, in a format that preserves what they know.
In the data hierarchy, raw facts are the foundation. Context is what makes them useful. And the only place context lives, reliably and completely, is inside the business itself. The agents that serve their users best will be the ones that query data layers rich in context — not just what a business is, but what it means for this specific person, at this specific moment, with these specific needs.