Single Posts

In independent UK testing, 93% of London businesses tested were described inaccurately by at least one major language model, and half of SMEs received at least one false fact compared with 32% of large companies. The fix isn't to correct ChatGPT, Gemini or Perplexity directly, because there's no edit button in those tools, it's to correct the source pages they copy from.

That distinction matters if you're spending serious money on ads. A wrong postcode in an AI answer, an old phone number in a directory, or a discontinued service repeated by Google AI Overviews can divert a buyer before your paid landing page gets a chance. Your targeting, bidding and creative can all be sound, and the campaign still looks weak because high-intent people never arrived, or arrived expecting something you don't sell.

Treat AI tools wrong business information as a feed-governance and measurement problem, not a curiosity. You need one agreed set of business facts, you need to find every public page that contradicts it, and you need evidence that a correction has actually changed the answer a customer sees.

Why AI Tools Get Your Business Details Wrong

Independent UK testing examined ChatGPT, Perplexity and Gemini more than 13,000 times across 165 London businesses, covering services, contact details, staffing and other identifying facts, and found that 93% of the businesses were described inaccurately by at least one major language model. That's not a handful of odd errors. It's the normal state of play for a business that hasn't audited its own footprint.

The risk falls unevenly. In that same testing, half of SMEs received at least one false fact compared with 32% of large companies, a rate 56% higher for smaller firms. Fabrication was also more than twice as common, at 5% of SME-related responses versus 2% for larger brands, and brand confusion appeared in 4% of SME cases against 0.7% for large companies.

Why Smaller Firms Get Hit Harder

Large brands generate volume. Press coverage, Wikipedia entries, big directories and thousands of consistent mentions give a model plenty of repeated, matching evidence to lean on, so a single stale page rarely wins.

An SME usually has a thin evidence base. If your website says one thing and three directories, an old review profile and a comparison article say something else, the stale version can look like the consensus. There's less correct material to outvote it, and fewer staff with time to notice.

• Thin source coverage: Fewer authoritative mentions mean each incorrect listing carries more weight.
• Brand confusion: A similarly named business in another town gets merged into your description, including its services or phone number.
• Fabrication to fill gaps: Where no clear fact exists, a model often produces a plausible one rather than saying it doesn't know.

The Mechanism in Plain Terms

AI tools don't ring your office to check. They assemble an answer from pages they've already read or retrieved, then phrase the result confidently, and they can't reliably tell an official page from a frequently repeated one. Frequency often beats authority.

A correct homepage doesn't cancel out five pages repeating the same outdated address. You have to fix the evidence, not argue with the answer.

That's why updating your own website alone rarely solves it, and why this work sits alongside generative engine optimisation as an AI search strategy rather than inside ordinary copywriting. You can't submit a correction to a model the way you'd edit a listing. You change what the model can find, then you test again to see whether the answer moved.

The Source Pages AI Tools Copy From

Separate UK retail testing across more than 72,000 questions about high-street retailers found that 64% of businesses received at least one false fact and one in 16 answers was incorrect. The useful detail is the error type. Postcode errors appeared in roughly one in ten location answers, about one in 15 website answers led to a dead link, a lookalike site or a different business, and wrong brand attribution occurred in one in 40 answers, with reported error rates of 4% for ChatGPT, 5% for Gemini and 10% for Perplexity.

Those errors come from somewhere specific. Four source categories account for most of what we find in an audit.

Old Directory Listings and Review Sites

Directories were often created once, at signup, years ago, with a former trading address, a mobile number belonging to someone who's left, or a service you stopped offering. Nobody owns them internally, so nobody updates them, and scrapers copy them into other directories.

Review sites and aggregators behave the same way. They repeat whatever business details were submitted originally, and they rank well, which makes them influential evidence. If your opening hours changed two years ago and three review profiles still show the old ones, an AI tool has more support for the wrong version than for yours.

Comparison Articles and Roundups

Comparison content is where price positioning and service scope go wrong. A roundup author summarises ten suppliers quickly, misreads an ex-VAT price as inc-VAT, lists a feature you withdrew, or copies a competitor's delivery promise into your entry. Because these pages are written in the comparative language customers use when they ask AI tools to choose between suppliers, they get retrieved for exactly the queries that matter commercially.

Inconsistencies on Your Own Site

Don't skip your own domain. The footer shows one address format, the contact page shows another, an old location page is still indexed, and a service page still promises something your current terms don't. Each conflict gives a model permission to pick the wrong one.

• Legacy landing pages: Campaign pages from previous promotions stay live with old prices and deadlines.
• Duplicate location pages: One per town, lightly rewritten, with hours edited in some and not others.
• Feed mismatches: Merchant Centre or marketplace data disagrees with the product page on availability or delivery.

For paid media, the sharpest consequence is diverted intent. A wrong postcode sends someone to the wrong branch, a dead URL ends the journey, and a lookalike site takes the click entirely. In your ad account that reads as falling conversion rate. In reality the ad worked and the AI answer stole the visit.

Checking What Each AI Tool Currently Says

Don't fix anything yet. Your first job is a baseline, because without a recorded starting point you can't prove a correction worked, and you'll waste time editing pages that weren't the cause.

Work through ChatGPT, Gemini, Perplexity and Google AI Overviews separately. Each one draws on different sources and produces different errors about the same business, so a clean answer in one tells you nothing about the others.

Build a Prompt Set That Mirrors Real Buyers

Ask the questions customers ask, in their wording, not yours. Cover location and postcode, opening hours, services offered, what you don't offer, price positioning, delivery or eligibility, phone number, and website address. Then add comparison prompts naming a real competitor, because those surface attribution errors that direct questions miss.

Test every branch on its own. A multi-site business is routinely correct for head office and wrong for the branch that moved, and a single blended check hides that completely.

1. Record the prompt exactly: Copy the wording into your log, including town and competitor names, so you can rerun it identically.
2. Record the tool and date: Answers change between model versions, so an undated screenshot is useless evidence six weeks later.
3. Save the answer verbatim: Paste the full text, plus any cited sources the tool shows, rather than summarising from memory.
4. Mark each claim against your canonical facts: Tag every statement as correct, wrong, outdated or invented, and note the commercial severity.

Score Severity Before You Start Editing

Not every error deserves the same urgency. A wrong founding year is embarrassing. A wrong phone number, a wrong postcode or a claim that you offer finance you don't offer costs you paid clicks, wasted calls and, in regulated categories, complaints. Rank by revenue impact first and embarrassment second.

Practical rule: Fix first whatever would make a ready buyer go somewhere else, call a dead number, or expect a price you won't honour.

Keep the log in a shared sheet rather than in someone's downloads folder, because you'll rerun it repeatedly. Our notes on how to get recommended by ChatGPT cover the visibility side of this, but recommendation without accuracy is a liability. Being named in an AI answer alongside the wrong address is worse than not being named at all.

Tracing a Wrong Answer Back to Its Source

A wrong answer usually carries a fingerprint of where it came from. Models paraphrase, but they preserve distinctive details, such as an unusual service name, a specific old postcode, an odd price format or a phrase nobody would invent independently. That detail is your search string.

Take the strangest element of the wrong answer and search for it in quotes, alongside your business name. If the AI says you offer same-day fitting in a town you left years ago, search that exact service phrase with the town. You'll often land on the directory entry, archived page or roundup paragraph the answer was built from.

Match the Error Type to the Likely Source

Different mistakes point to different places, and using that shortcut saves hours of random searching. Work through the likely origin before you widen the hunt.

• Wrong address, hours or phone number: Start with directories, review profiles and data aggregators, then check for a duplicate or unverified map listing.
• Wrong service description or positioning: Look at comparison articles, roundups and competitor pages that describe you in passing.
• Wrong price or promotion: Check old landing pages, voucher sites, marketplace listings and cached product feeds before you blame a directory.
• Wrong website URL: Search your brand name plus the domain the AI gave, because lookalike and expired domains often sit behind this.
• A competitor's facts attached to you: Compare both businesses across the same directories, since merged or scraped records are a common cause.

When the Source Is Your Own Site

This happens more often than owners expect. A legacy page is still live, still indexable and still linked from an old blog post, so from a crawler's perspective it's a current statement of fact. Search your own domain for the wrong detail using a site search, and check PDFs, old brochures and archived campaign pages, which teams forget because they don't appear in the main navigation.

Where a tool shows its citations, use them, but don't treat the citation list as complete. It shows what the tool chose to display for that answer, and the underlying claim may also exist on pages it didn't cite. Fixing only the cited page can leave three other copies intact.

Build a Source List Per Error

Finish with one row per error containing the wrong claim, the correct fact, every URL you've found repeating the wrong version, who controls each URL, and whether you can edit it directly or need to request a change. That sheet becomes your work queue.

Resist the urge to start correcting as you find things. Diagnosis first means you see patterns, such as one aggregator feeding six directories, and patterns tell you where a single fix removes multiple errors at once.

Correcting Profiles and Listings in the Right Order

Order matters because these sources feed each other. Correct a downstream copy while the upstream record is still wrong and the error comes back, which is why teams conclude the work doesn't stick when really they fixed the wrong layer.

Start with Google Business Profile. Google states that edits to Business Profile information can take up to three days to appear in search results, and verified businesses can update address, opening hours, contact information and photos. The same guidance explains that Google can add updates when it receives information from other sources reporting that a profile is incorrect or outdated, and owners can review those changes and edit, accept, resubmit or discard them.

That last point explains recurring errors. Your profile isn't a fixed record you control outright, so a stale directory entry can resurface as a suggested update to your own listing. If you only correct the profile and leave the directories untouched, you may be re-correcting it repeatedly.

A Workable Correction Sequence

1. Google Business Profile first: Find the profile, select Edit profile, review any pending updates shown beside your original information, then resubmit corrected details for each location separately.
2. Other map and directory profiles next: Work through the major directories, review platforms and industry listings from your source list, claiming unclaimed profiles rather than creating duplicates.
3. Marketplace and feed listings: Correct Merchant Centre, marketplace and comparison feeds so product price, availability and delivery data match your product pages.
4. Third-party editorial content: Contact roundup and comparison publishers with the specific line that's wrong, the correct wording, and a link to the page on your site that proves it.
5. Your own site last in sequence, first in authority: Remove or redirect legacy pages, then make sure footer, contact page and location pages all state identical facts.

Where a publisher won't update an article, ask for the claim to be removed rather than argued over. A missing claim does less damage than a confident wrong one, and removal requests are often granted faster than rewrites.

Set Expectations on Timing

A saved edit isn't a published edit, and a published edit isn't an updated AI answer. Allow for Google's stated window first, then recheck Google Search and Maps in the public view, not just inside your dashboard, because the two don't always match during propagation.

Don't declare a fix complete because you pressed save. It's complete when the same prompt, run again, returns the correct fact.

Log the date of every change against the error it relates to, then rerun your baseline prompts after the window and record the new answer beside the old one. If the answer hasn't moved, you've usually missed a source, or the tool is still using older retained information, and either way the log tells you what to investigate rather than guess.

Businesses with several locations usually need this joined up with their local presence work, which is where local SEO optimisation services and profile management overlap in practice. Controlled separately by different people, the same inconsistency keeps reappearing.

Publishing Clear Facts and Adding Schema

Corrections remove bad evidence. This part creates good evidence that's easy for machines to read and hard to misinterpret, which is what makes the fix durable.

Start with a single canonical dataset, agreed internally and written down in one place. It should cover legal trading name, address and postcode per location, telephone number, email, opening hours including exceptions, services offered, services explicitly not offered, service area, price positioning, delivery terms and returns. Every page, profile, feed and ad extension gets checked against that document, and when a fact changes, the dataset changes first.

Publish Facts Where a Human Can Verify Them

Put the facts visibly on the relevant location or service page in plain sentences. A statement like "Our Leeds branch is at this address, opens at these times, and offers these services" in ordinary readable text is worth more than the same data buried in an image, a PDF or a script that loads after interaction. Write it so each fact stands on its own, because retrieval systems lift short passages out of context.

Be explicit about negatives too. If you don't offer finance, don't install in Scotland, or don't sell to trade, say so on the page. Silence is where fabrication gets in.

Add LocalBusiness Structured Data That Matches the Page

Google's documentation identifies fields that make business facts machine-readable, including a physical location as a PostalAddress, opening hours and priceRange, and recommends defining each location as a LocalBusiness using the most specific available subtype. It also notes that providing more address properties improves result quality, and that priceRange must be under 100 characters.

Schema must reinforce what's visible, never add claims a visitor can't verify on the page. Markup that contradicts the visible text creates exactly the conflicting evidence problem you're trying to remove, and it can cost you rich result eligibility. Validate the JSON-LD, then check that the rendered page and the markup still agree after any template or plugin update, because silent drift is common on sites with multiple editors.

This is unglamorous on-page work, and it pairs naturally with an on-page SEO service review of headings, internal links and page structure. Done properly, your own location page becomes the clearest, most complete and most consistently repeated statement of your facts anywhere on the web, which is precisely what makes it the version models reach for.

Doing This In-House or With a PPC Specialist

Split the work into what you can handle yourself. The baseline audit, source tracing and most profile corrections are well within reach of a competent marketing manager with a spreadsheet and persistence. Nobody needs an agency to update a directory listing.

What tends to break in-house isn't effort, it's ownership and verification. UK government research found that 84% of businesses using AI retain at least some human input or checking, including 67% applying significant checking, while only 2% use AI outputs with none. Checking is normal practice, but in a small team it's usually nobody's named job, so it happens after a customer complains.

Where In-House Works and Where It Struggles

• In-house strengths: You know the correct facts, you can approve wording quickly, and you already hold logins for most profiles.
• In-house limits: Feed governance, landing page alignment, server-side analytics and conversion tracking sit across several people's roles, so inconsistencies creep back between audits.
• Specialist strengths: Daily work inside ad platforms makes it easier to spot where AI-referred traffic disappears and to tie a data error to wasted spend.

The reason this belongs near paid media is commercial. A wrong postcode or URL diverts high-intent users before they reach a paid landing page, so the account shows apparent underperformance while targeting and bidding are fine. Diagnosing that requires looking at branded query behaviour, landing page logs and tracking together, not treating it as a content tidy-up.

PPC Geeks is one option for that side of the work, as a Google Ads Partner running campaigns across Google Ads, Microsoft and Bing, Facebook and Amazon, with free account audits, feed optimisation, landing page creation and conversion tracking as part of its onboarding and ongoing reviews. In practice that means product feeds, location data and landing page facts get checked against each other, and AI-generated advertising copy gets human review for claims, prices and product details before it goes live.

Choose based on a simple test. If you have one location, a short source list and someone who'll own a monthly recheck, do it yourself. If you have multiple locations, a product feed, several marketplaces and no single owner, the errors will outpace a part-time effort.

Your First Step This Week

Don't try to clean the whole internet. Pick one location and one AI tool, run your baseline prompts, and find the single error that would cost you a ready buyer. For most businesses that's a wrong postcode, a dead or lookalike URL, a phone number that doesn't ring, or a service claim you can't fulfil.

Trace that one error to its source pages, correct them in order, and write down the date. Then wait for propagation and rerun the identical prompt. If the answer changes, you've proved the method on the error that matters most, and the rest becomes a queue rather than a mystery.

Build the habit around it afterwards. Monitor branded queries for incorrect destinations, validate AI-referred traffic with UTM parameters, server-side analytics and landing page logs, and treat AI answers as an untrusted referral layer rather than a source of truth for campaign or feed decisions. When a tool says something about your business that you haven't published yourself, check it before it reaches an ad, a feed or a landing page.

Visibility in AI answers is easy to chase and easy to misread. Accuracy is what protects the budget you've already committed, and it's achievable with a canonical fact sheet, a source list and a recheck you repeat on a regular basis.

If AI tools are sending buyers to the wrong postcode, quoting prices you don't charge or describing services you don't offer, PPC Geeks can audit what those tools say, trace the source pages behind each error and align your profiles, product feeds and landing pages so paid traffic lands on accurate information. Start with a free account audit and a clear correction plan by visiting PPC Geeks.