Why AI assistants get your prices, address and opening hours wrong, and how to fix it
An assistant that sends customers to your old address or quotes last year's prices is worse than one that ignores you. Here is where the wrong facts come from and a repeatable way to correct them.
By Seoptist Team · · 6 min read
Being mentioned by an AI assistant is only good if what it says is true. A dental practice that moved across town in 2024 and still gets patients turning up at the old address, a cabling firm quoted at "around £30 per point" when its price has been £45 for two years, a café listed as open on Mondays when it has never opened on a Monday: these are the complaints we hear most from businesses that start tracking their AI visibility.
The errors are rarely random. They come from a small number of causes, and each has a fix.
Where wrong facts come from
1. The model's training data is older than your change
Every assistant starts from a model trained on a snapshot of the web. If your address changed after that snapshot, the model's default belief is the old address. Live retrieval can override it, but only if retrieval happens and finds something more convincing than the model's prior.
2. A third-party listing is out of date
Assistants with retrieval lean heavily on directories, review sites and aggregators because those sources are structured and plentiful. If a directory you registered on in 2017 still shows your old number, and the assistant reads that listing before (or instead of) your site, the old number wins. This is the single most common cause we see.
3. Your own site contradicts itself
A price in the footer, a different price in the FAQ, a third in a PDF brochure that is still indexed. Opening hours in the LocalBusiness schema that were never updated after you changed them on the contact page. A model reading two of these picks one, and not necessarily the one you intended.
4. The answer was a plausible guess
When an assistant cannot find a fact, it sometimes fills the gap with something typical for the category. "Most emergency plumbers in the area charge a call-out fee of around £80" becomes "their call-out fee is around £80". The model is not lying in any deliberate sense; it is completing a pattern. The fix is to make the real number easy to find so there is no gap to fill.
5. Entity confusion
Two businesses with similar names in nearby towns, or a parent company and a trading name, get merged. Your competitor's hours become yours. This is more common than people expect and is almost always a consistency problem on your side as well as theirs.
The fix: one source of truth, propagated
The approach that works is the same one franchise and multi-site businesses have used for years, scaled down.
Step 1: Write your fact card
A fact card is a single document with every fact you want the world to agree on. Keep it short and literal.
Legal name: Example Heating & Plumbing Ltd
Trading name: Example Heating & Plumbing
Address: 12 High Street, Reading, RG1 2AB
Phone: 0118 123 4567
Email: [email protected]
Hours: Mon-Fri 08:00-18:00; Sat 09:00-13:00; Sun closed
Emergency line: 24/7 on 0118 123 4568
Service area: Reading, Wokingham, Bracknell, Henley
Registrations: Gas Safe 123456; Companies House 01234567
Core prices: Call-out £75 inc. first hour; boiler service £95; landlord CP12 £65
Founded: 2009
Seoptist calls this the fact card and uses it as the reference for accuracy checks; the same idea works in a spreadsheet.
Step 2: Make your own site match it exactly
Audit every place a fact appears: header, footer, contact page, service pages, FAQ, structured data, llms.txt, downloadable PDFs. Fix every mismatch. Where a price is shown in more than one place, generate it from one source if your CMS allows.
Pay particular attention to JSON-LD. A LocalBusiness block with stale openingHoursSpecification is a machine-readable lie, and machines read it.
Step 3: Find and correct every third-party copy
Search for your old phone number, old address and old trading name in quotes. Make a list of every site that shows them. For each, log in and correct it, or submit a correction. The usual UK list:
[ ] Google Business Profile
[ ] Apple Business Connect (Apple Maps)
[ ] Bing Places
[ ] Companies House (registered office and SIC; if the trading name differs, say so on your site)
[ ] Yell, Thomson Local, Scoot, Cylex, 192.com
[ ] Trade platforms: Checkatrade, TrustATrader, Rated People, MyBuilder, Which? Trusted Traders
[ ] Review platforms: Trustpilot, Google, Facebook, Yelp UK
[ ] Trade bodies: Gas Safe, NICEIC, NAPIT, FENSA, GDC, SRA, ICAEW, etc.
[ ] Local press articles you can ask to have updated
[ ] Your own old domains or microsites still indexed
Where you cannot edit a listing, request removal. An absent listing is better than a wrong one.
Step 4: Remove the gaps that invite guessing
If customers ask about price, publish a price or a pricing basis. If they ask about hours, publish hours, including bank holidays if you differ. If they ask whether you cover a town, list the towns. The aim is that any reasonable question about a fact has a findable, consistent answer, so the assistant never has to invent one.
Step 5: Monitor for drift
Facts drift. A directory re-imports old data, a staff member updates the Google profile but not the site, a price rises. Set a quarterly reminder to re-check the list above, and watch what assistants actually say.
Seoptist's accuracy alerts (Growth and above) compare every sampled answer against your fact card and flag a mismatch on price, address, phone or product details the week it appears, with the engine and the sources it cited so you can trace the origin.
A note on tracking phone numbers
Call-tracking numbers on directories and ads are a common and avoidable cause of NAP inconsistency. If you must use them, use them only where the platform supports a "primary" and "tracking" number distinction (Google Business Profile does), and keep the primary number identical everywhere.
Before and after
A typical pattern, offered as an illustration rather than a study: a business with one stale directory listing and inconsistent schema sees an assistant quote the wrong number in a meaningful share of answers. After the listing is corrected and the schema aligned, the wrong number disappears from retrieval-based answers within a few weeks as the sources are re-crawled, while answers that rely on training data alone correct more slowly, over months. Both move; the second just takes longer.
If you want to see what assistants currently say about your facts, the free AI Visibility Check shows the answers and their sources for five prompts.
Frequently asked questions
How quickly will an assistant stop quoting my old address?
For answers built on live retrieval, usually within a few weeks of the sources being corrected and re-crawled. For answers drawn from training data alone, not until the next model update. Fixing the sources is the only lever you have for both.
Should I report wrong answers to the AI provider?
You can, and the feedback buttons in each product exist for that, but there is no reliable correction workflow for businesses. Fixing the underlying sources is faster and more durable.
What if a directory refuses to correct my listing?
Ask for removal instead. If that is also refused, strengthen the correct signal elsewhere (your site, Google Business Profile, trade bodies) so the stale listing is outnumbered.