When AI Gets Your Brand Wrong: Fixing Hallucinated Facts at the Source

When ChatGPT invents your founding year or still recites last year’s address, it is usually amplifying a weak, conflicting, or leftover trail of sources — or filling silence with a plausible guess. Correcting ChatGPT about your company is not a support ticket. It is source control: one canonical fact packet on your domain, cleanup of stale NAP and wrong-year residue, and patient re-testing across products.

This spoke is the brand-safety loop inside the Answer Engine Optimization playbook. The identity layer that makes those facts stick lives in entity architecture for AI search. I have been SEO certified since 2021. The AEO version of that work is still a fact sheet you can defend, not a prompt you yell at.

The short answer

• Hallucinated brand facts have URLs behind them more often than they have “the model made it up”
• Stale NAP (name, address, phone) and a wrong founding year are the two errors that keep coming back after a move or a rebrand
• Unify owned surfaces first: About, footer, press kit, Organization / LocalBusiness schema, Google Business Profile
• Search the web for the wrong claim, then correct or outcompete the leftover pages
• Re-test the same prompts. There is no universal “delete this from ChatGPT” button

Why do models invent facts about your brand?

Models invent brand facts when the public record is silent, split, or stale. They are not picking a fight with your About page. They are compressing whatever the crawl and the training mix made easy to say.

Cause	What you see in the answer	What is usually true on the web
Silence	A confident year or HQ nobody on your team recognizes	No dated About facts; schema missing foundingDate and address
Conflict	2019 on one run, 2017 on the next	Site, LinkedIn, Crunchbase, and an old PR hit disagree
Stale residue	Last office, sunset SKU, old DBA	Directories, PDFs, and partner footers were never updated
Name collision	Wrong industry, wrong city, wrong founder	Another company shares the string
Satire / scrapers	Awards, headcount, or funding you never announced	Profile farms copied a guess, then copied each other

Treat the corpus as the patient. Treating the model as the enemy wastes a week of screenshots.

ChatGPT Search is explicit that answers can include links to web sources (OpenAI: Introducing ChatGPT search; OpenAI Help: ChatGPT Search). Google’s AI features pull from crawlable, snippet-eligible pages and tell you to keep Business Profile information up to date (Google Search Central: AI features). If a leftover Yelp row or a 2018 PDF still teaches the old suite number, retrieval has something to quote.

• You can point to the exact wrong sentence, not “ChatGPT is messy about us”
• You know whether the run used search / citations
• You have already searched the web for that same wrong sentence
• You have not filed this as “the model is broken” before checking owned pages

What counts as a hallucinated brand fact?

Not every sloppy paraphrase is an incident. Log the ones that change a buyer’s next action or create a support ticket.

Claim type	Example of a real incident	Example of noise
NAP	Recites the 2019 suite or a disconnected phone	Abbreviates “Street” as “St.” when both are your format
Founding year	Says you launched in 2014 when legal formation is 2017	Omits the month when you only publish the year
Offer	Invents a $49/mo plan you never sold	Shortens a custom-quote sentence
People	Names a cofounder who left, or one who never existed	Drops a middle initial
Legal / safety	Invents a lawsuit, breach, or ban	Uses a hedge you would not write
Identity	Merges you with a same-name company in another city	Uses a common nickname you already accept

If sales or support cannot use the answer as a briefing, it is an incident. If only the brand team is annoyed, it is still worth a log row — it is not a war room.

1. Quote the wrong claim in the sheet
2. Write the correct claim with one evidence URL
3. Tag severity (critical / high / medium / low)
4. Assign an owner and a retest date
5. Do not debate tone until the fact is settled

What is stale NAP, and why do answers keep the old address?

NAP is name, address, and phone. BrightLocal’s local-citation work treats those three strings as the payload directories copy, and treats inconsistency as a trust problem for both search engines and people (BrightLocal: What is NAP; BrightLocal: NAP data accuracy). Answer engines inherit the same mess. If half the web still has the old suite, a retrieved answer will too.

Google’s Business Profile rules are stricter than “close enough.” The public name should match real-world signage and stationery, not a keyword-stuffed slogan. The address should be a real location customers can use — not a P.O. box or a remote mailbox. The phone should connect to that location and stay under the business’s control (Google Business Profile: representation guidelines). Incomplete or inaccurate profile info is also how Google describes weaker local visibility (Google: tips to improve local ranking).

NAP field	Canonical rule	Residue that keeps teaching the error
Name	One legal / public name, same punctuation and suffix	Old DBA, “LLC” on some rows, slogan glued to the name
Address	One street format, one suite notation, one city/state/ZIP	Prior office; “Ste” vs “Suite” vs ”#”; missing unit
Phone	One primary line per location, one written format	Call-tracking number on GBP, cell on the footer, 1-800 on ads
URL	One canonical host (www or apex, https)	Staging host, old domain, Facebook-as-website
Hours	One weekly grid plus holiday exceptions	Holiday hours left up in March

Humans forgive “Street” vs “St.” Models and aggregators often do not collapse those rows into one entity. Pick a format and paste it. Do not retype it from memory.

Google’s own AI-features checklist includes keeping Merchant Center and Business Profile information current (Google Search Central: AI features). That is not a local-SEO side quest. It is part of how Google describes staying eligible to be used in AI experiences.

• One written NAP card exists (name, address lines, phone, URL, hours)
• Footer, Contact, About, and location pages paste from that card
• Google Business Profile matches the card after you edit the profile
• Call-tracking numbers are labeled as tracking, not as the public primary
• Old addresses appear only on a dated “we moved” note, never as current

After a move, the leftover address is the default AI answer until you hunt it. Directories, Apple Maps, Bing Places, event pages, speaker bios, invoices turned into PDFs, and partner “locations” blocks do not update themselves.

Why does AI get your founding year wrong?

A founding year is not one event. Teams mix legal formation, first invoice, first public launch, and the year someone typed into LinkedIn at 11 p.m. Models then average the conflict into whichever year appears most often — or invent a nearby year that “sounds” right.

schema.org defines foundingDate as the date the organization was founded (schema.org/foundingDate; schema.org/Organization). Google’s Organization structured-data guide lists foundingDate as an ISO 8601 date when you publish it (Google Search Central: Organization). Wikidata’s inception property (P571) is “time when an entity begins to exist,” and it tells you not to use that field for an official opening (that is P1619) (Wikidata: P571). If your About page says 2019, your JSON-LD says 2017-03-15, LinkedIn says 2020, and a Wikidata stub says 2016, you taught four different years.

Year someone wants to publish	What it actually is	Publish it as
Articles of organization / incorporation	Legal formation	Default foundingDate unless counsel says otherwise
First paid invoice	Commercial start	“First client work in YYYY” — not the founding year
Site launch or brand rename	Public launch / DBA change	“Publicly as [name] since YYYY”
LinkedIn “founded”	Whoever last edited the company page	Align to the legal year or lock the field
First office lease	Occupancy	Do not use as founding year
Predecessor studio or side project	A different entity	Name the predecessor; do not collapse the years

Pick one year. Write the sentence that defends it. Repeat that sentence on About, the press kit, schema, and any profile you still control. If you also want the launch story, put it in a second sentence so the model has a clean pair instead of a blended year.

Google’s structured-data policy is the tripwire: do not mark up facts that are not visible on the page, and the markup must be a true representation of the page (Google: structured data guidelines; Google: intro to structured data). A foundingDate of 2017 next to visible copy that says 2019 is not “extra signal.” It is a conflict you authored.

1. Pull the formation document or the accountant’s first-year filing
2. Write one public sentence: “Founded in YYYY” plus the evidence you are willing to show
3. Put that year in visible copy and in foundingDate (YYYY or YYYY-MM-DD)
4. Search "Your Brand" founded and "Your Brand" "established"
5. Correct every owned hit before you email a journalist

How do you write one ground-truth fact packet?

The packet is a short, dated page — or a section of About — that a human and a model can both lift. If a fact is not on that page, it is not a public fact yet.

Field	Required	Notes
Public name	Yes	Matches signage and GBP
Legal name	If different	legalName in Organization markup
Founded	Yes	One year; ISO in schema
HQ / service area	Yes	Address or explicit service-area language
Primary phone	If you publish one	Same format everywhere
Leadership	If you name people	Matches Person pages and LinkedIn
Current offers	Yes	What you sell now, not what you sunset
Formerly known as	If you renamed	Dated; see the rename section below
Correction contact	Yes	A real inbox, not a void form
Last reviewed	Yes	Update when anything material changes

Use Organization markup on the home or About page so Google can disambiguate the entity — logo, sameAs, address, identifiers (Google Search Central: Organization). If you are a location business, LocalBusiness (or a more specific subtype) is the type Google documents for hours, address, and the local knowledge panel (Google Search Central: LocalBusiness; schema.org/LocalBusiness). sameAs is the official hook for “this is the same thing as that Wikipedia / Wikidata / profile URL” (schema.org/sameAs).

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Example Studio",
  "legalName": "Example Studio LLC",
  "url": "https://www.example.com",
  "foundingDate": "2017",
  "telephone": "+1-555-0100",
  "address": {
    "@type": "PostalAddress",
    "streetAddress": "100 Example Street, Suite 200",
    "addressLocality": "Nashville",
    "addressRegion": "TN",
    "postalCode": "37203",
    "addressCountry": "US"
  },
  "sameAs": [
    "https://www.wikidata.org/wiki/Q0",
    "https://www.linkedin.com/company/example-studio"
  ]
}

That block is worthless if the visible page still says Suite 80 and “est. 2014.” Markup copies the page. It does not outvote the page.

• Packet URL exists and is linked from About, press kit, and llms.txt
• Visible sentences and JSON-LD use the same name, year, and address
• sameAs only lists profiles you still control or that already match
• Sunset offers are on a dated archive page, not in the current-offer list
• One named owner can edit every field without a committee

How do you correct ChatGPT about your company?

You correct ChatGPT by correcting the record it can retrieve, then re-running the prompt. Product-UI feedback is optional extra. It is not the plan.

Step	Action	Done when
1. Capture	Save product, date, prompt, exact wrong claim, citations if shown	Another person can reproduce the miss
2. Ground truth	Write the correct fact plus one primary evidence link	Legal / founder / filing agrees
3. Owned first	About, footer, schema, GBP, press kit, redirects	Zero owned conflicts remain
4. Amplifiers	Search the wrong claim; list leftover URLs	Top 20 hits have owners and actions
5. Correct or outcompete	Edit, request correction, or publish a stronger accurate page	Each URL is closed, waiting, or accepted as residue
6. Packet	Add the fact to FAQ, fact sheet, media kit	The sentence is copy-paste identical
7. Re-test	Same prompt, same product, logged conditions	Day 3 / 14 / 30 / 90 rows exist

OpenAI documents that ChatGPT Search responses may include inline citations and a Sources panel (OpenAI Help: ChatGPT Search). When those links are present, start there. The leftover page is often in the citation list. When citations are absent, search the web for the wrong claim yourself. Do not assume the model “just knows.”

Google’s generative-AI guidance is the same idea from the other side: those features use publicly crawlable content, and the usual Search quality rules still apply (Google: optimizing for generative AI features). You do not get a private rewrite channel. You get a cleaner web.

1. Reproduce the miss in a fresh session (logged-out or a clean profile)
2. Note whether search was on
3. Fix owned conflicts the same day for high-severity errors
4. Work the amplifier list in authority order, not alphabetical order
5. Re-run before you declare victory

Which owned surfaces must agree before you chase the web?

If your own domain disagrees with itself, every correction email is theater. Stop outbound work until the owned set is one packet.

Surface	What must match	Common break
About / facts URL	Name, year, HQ, offers, formerly-known-as	Marketing rewrite drifted from legal
Footer and Contact	NAP	Designer typed a prettier address
Location pages	Per-location NAP	One schema block for three offices
Organization / LocalBusiness JSON-LD	Visible NAP + foundingDate	Plugin defaulted to an old year
Google Business Profile	Name, address, phone, URL, hours	Tracking number; unverified move
Press kit / one-pager PDF	Same sentences	Sales exported a 2023 deck
llms.txt	Pointers to the facts URL	File lists a sunset offer as current
Redirects	Old names and old paths	Soft-404 on the previous brand URL

Google’s LocalBusiness documentation is built for the location case: hours, departments, and a working location URL (Google Search Central: LocalBusiness). Multi-location brands need one NAP card per location, not one “HQ” string pasted onto every city page.

The entity architecture spoke is the identity map. This post is the incident loop when that map is already lying.

• Diff About vs footer vs schema vs GBP in one sitting
• Kill or redirect every owned URL that still publishes the old year or old suite
• Replace PDFs; do not “update the site” and leave the kit live
• Confirm foundingDate and PostalAddress match visible copy
• Only then open the leftover-source spreadsheet

How do you find leftover sources still teaching the error?

Residue is the set of pages that still publish the wrong NAP or the wrong year after you fixed the site. You will not find them by staring at ChatGPT. You find them with search operators and a spreadsheet.

Query	What it catches
"Your Brand" "2014" (wrong year)	Profiles, bios, and roundups still on the old year
"Your Brand" "Suite 80" (old unit)	Directories and footers that never moved
"Your Brand" "555-0199" (old phone)	Call-tracking leftovers and scraped listings
"Your Brand" founded OR established	Year conflicts in prose
"Former Name" studio	Rename residue
site:crunchbase.com "Your Brand"	Investor-profile drift
"Your Brand" filetype:pdf	Press kits and one-pagers you forgot

Log every hit.

Column	What you write
URL	Full URL, not a domain
Wrong claim	Quote
Authority	High / medium / low (who would a model trust?)
Owner	You / friendly publisher / aggregator / dead
Action	Edit / request / outcompete / accept
Sent	Date of the correction request
Recheck	Date you will reload the URL

Patterns show up fast. One bad Data Axle–style seed can infect five scrapers. One conference bio can freeze a cofounder who left in 2022. One PDF in a /press/ folder can outrank your new About page in a retrieval set because it is short, declarative, and full of facts — the wrong ones.

• Wrong-year query run
• Old-address and old-phone queries run
• PDF query run
• Top 20 hits classified
• High-authority leftovers assigned this week, not “sometime”

How do you correct residue you do not control?

You cannot delete the open web. You can change the mix a retriever is likely to see.

Residue type	First move	If that fails
Google Business Profile	Edit the verified profile; expect re-verify on name/address changes (GBP edit help)	Follow Google’s verification path; do not open a second listing
Directories you can claim	Claim, match the NAP card, mark closed duplicates	Request removal of the zombie listing
Friendly publishers	Email the editor with the facts URL and the evidence	Offer a corrected sentence, not a rant
Wikidata item you can edit	Fix P571 with a cited source; do not invent an item to “win”	Leave it alone if you have no reliable source
Dead blogs / expired magazines	Publish a clearer, dated facts page and earn newer citations	Accept as low-authority residue; stop refreshing it with quote-tweets
Profile farms	Ignore most of them	Spend time only if they dominate the citation panel
Name collision	Disambiguate on About (“not the [city] [industry] firm”)	Add sameAs and a unique descriptor in the first sentence

Do not flood Wikipedia with primary-source spam. Do not buy fake reviews to “correct” sentiment. Do not ship twenty thin pages that repeat “Founded in 2017” and nothing else. Those pages become more residue, not more proof.

Google’s May 2025 AI-search note repeats the structured-data rule in public: markup should match visible content (Google Search Central Blog: succeeding in AI search). That is the same rule you use on leftover schema. If a plugin still emits the old year on a template you forgot, you are the residue.

1. Claim and fix GBP and the three directories buyers actually use
2. Send evidence-backed corrections to the five highest-authority leftover URLs
3. Strengthen the facts page so a fresh crawl has something clean to lift
4. Disambiguate if another entity shares the name
5. Stop feeding rumor threads that create more text for scrapers

How should you triage severity?

Not every wrong AI sentence deserves a war room. Route by what the error costs, then publish the routing so support knows when to page marketing.

Severity	Example	Response time	Who owns it
Critical	Invented lawsuit, breach, or misconduct	Same day: counsel, facts page, publisher outreach	Founder + counsel
High	Wrong price band, coverage area, or current address that sends people to an empty suite	48 hours: owned fix + GBP + support macro	Marketing + ops
Medium	Old product name, old DBA, year off by one on a brand prompt	This week: redirects + formerly-known-as + directory claims	Marketing
Low	Minor format drift, nickname you already accept	Quarterly cleanup	Whoever owns the packet

Legal and safety falsehoods are a different sport:

1. Involve counsel before you publish a status page
2. Document primary sources
3. Request corrections from any indexer repeating the claim
4. Do not dunk in public threads that become new training text
5. Keep the packet boring and dated

Pricing hallucinations need a public posture page even when you do not list SKUs (“custom quotes; typical projects land in this band”). Support must use the same band. When a model invents a $49/mo plan, your silence is the source.

People hallucinations — wrong cofounders, invented degrees — start on Person pages and LinkedIn. Conference sites freeze bios. Send the update before the next season, not after the keynote page ranks.

How long should “formerly known as” stay live?

Keep the old name findable until the prompt panel stops retrieving it as current. Redirects stay forever. The prose line can retire later.

Change	Keep “formerly” on the facts page	Keep redirects	Remove the prose when
Product rename	6–12 months	Indefinite	Brand prompts stop using the old SKU as current
Company / DBA rename	12–24 months, often longer	Indefinite	Leftover-source search dries up
Office move	Until the new NAP is the majority hit	Old location page → new, or a dated move note	Old suite queries return the move note, not a live listing
Sunset offer	Until support tickets about it stop	Offer URL → successor or archive	Sales no longer hears the old package name

“Formerly” is not nostalgia. It is a retrieval bridge. Kill it too early and the old name becomes an orphan string the model attaches to whoever still uses it.

• Redirect map includes every old host, path, and campaign vanity URL
• Facts page states the old name and the date it stopped being current
• GBP and directories use the current name only
• Sales decks match the current name the same week
• Re-test includes one prompt that uses the old name on purpose

How do you re-test after you fix the sources?

Do not retest once. Retrieval can pick up a fixed About page in days. Training residue and uncached directory copies can linger. Plan for both horizons.

Day	What you are checking	Pass looks like
0	Baseline of the miss	Quote, product, citations saved
3	Fresh session with search on	Citations include the facts URL or GBP, not only leftovers
14	Same prompt + one paraphrase	Wrong year / old NAP gone or clearly outvoted
30	Full brand-accuracy subset	No high-severity errors on the frozen set
90	Residue search + panel	Leftover URLs declining; no new scrapers of the old suite

Log conditions. A logged-in ChatGPT memory is not the same as a fresh session. A Google AI Overview is not the same as a chat product. Write the product name on the row.

Google’s AI-features documentation is eligibility, not a promise: indexed and snippet-eligible is the floor for being used as a supporting link (Google Search Central: AI features). Your retest is whether the generated sentence is true, not whether you “rank” in a feature.

Column	Legal values
prompt_id	Stable ID (brand-year, nap-address)
product	ChatGPT Search, AI Overviews, other named surface
search_on	yes / no / unknown
wrong_claim	Quote or none
cited_url	URL or empty
accuracy	pass / fail
notes	Which leftover URL still appeared

1. Duplicate last month’s sheet; do not overwrite
2. Run the same prompts in the same order
3. Score accuracy before you read the rest of the answer
4. Capture citations when the product shows them
5. Open tickets only for fails that still have a leftover URL you have not worked

What should you refuse to do?

These moves create more conflicting text or waste the only hours you have.

Temptation	Why it fails	Do this instead
Argue with the model in a public thread	You generate more scrapable fiction	Fix sources; log the miss privately
Wikipedia spam	Editors revert you; you look desperate	Earn coverage; fix Wikidata only with sources
Fake review volume	Policy risk; does not fix a year	Correct listings and the facts page
Twenty thin “we were founded in…” posts	Duplicate residue	One facts URL, many inbound links
Second GBP listing “to replace” the old one	Duplicates split the entity	Edit and re-verify the real profile
Call-tracking number as the public primary	NAP splits across ads and the site	Keep one public phone; track behind it
Marking up a year the page does not show	Violates Google’s match-visible rule	Change the copy, then the JSON-LD

Brand safety beats clever dunks. The clean packet plus a handful of reputable corrections beats a quote-tweet war in retrieval over time.

• No public pile-on threads as the “strategy”
• No second Maps listing
• No schema that the page does not show
• No awards on the facts page you cannot prove (farms will inflate them)
• No “AI said this” blog post that repeats the false claim in the headline

What breaks if you only fix the homepage?

What breaks: About is correct, the footer still has the old phone, the plugin still emits 2014, GBP still has the tracking number, and a PDF in /press/ still says Suite 80. Retrieval samples the easy, short, repeated strings. The homepage paragraph loses.

What it costs: support tickets from people who drove to the empty suite; sales calls that start with “I thought you launched in 2014”; a week of “we already fixed that” while ChatGPT cites the PDF.

What you do instead:

1. Treat NAP and founding year as a matrix, not a hero sentence
2. Diff every owned surface in one sitting
3. Replace files, not just HTML
4. Claim the leftover directories
5. Re-test with the old address and the wrong year as prompts, on purpose

That last step is the one teams skip. If you only ask “What year was Example Studio founded?” you will miss “Example Studio Suite 80” still resolving as current.

Multilingual pages are first-class fact surfaces. A machine-translated Spanish About page with the wrong year will dominate Spanish prompts even when English is clean. Assign bilingual review to the packet, not only to campaign copy.

Synthetic company-profile farms will keep inventing headcount and awards. You will not whack every mole. Make official pages unmistakable and strip awards you cannot prove so the farms have less to distort.

What does week one look like after you catch a wrong fact?

Do not start with a 40-ticket AEO program. Start with the incident that already happened.

Day	Output	Done when
1	Incident row + ground truth	Quote, correct claim, evidence URL
1–2	Owned matrix reconciled	About, footer, schema, GBP, PDFs agree
2–3	Leftover-source list (top 20)	Each URL has an owner and an action
3	Support macro live	Tickets tagged; facts URL in the reply
3	Facts page dated	Last-reviewed stamp updated
7	First retest + two correction emails	Day-3 row filled; high-authority leftovers contacted

Capture five brand prompts across two products. Include one founding-year prompt, one address prompt, one phone prompt, one “what do they sell” prompt, and one old-name prompt if you renamed. Log every material error. If you do nothing else from this article, that kit stops silent drift from becoming customer-facing fiction.

Repeat the kit after a move, a rename, or a launch. The cost of re-baselining is small next to a quarter of unmeasured residue. Name the owner in the sheet. When someone goes on leave, transfer the ritual in writing. AEO dies in the handoff gaps.

Someone must own the fact packet with authority to make other teams update their copy. Without that owner, the old year returns through a sales one-pager. Review quarterly. Treat silence as a risk, not a steady state.

If you want a second pair of eyes, the visibility lane is built for this: accuracy sampling, leftover-source priority, and a 30/60/90 that starts with the truth layer — not another blog calendar. The system view is the AEO playbook.

FAQ

Why is AI hallucinating our brand information?

Usually because sources conflict, go silent, or still publish a leftover address or year. The model fills the gap with the most common or most plausible string it can retrieve. A clear packet on owned pages plus matching directories leaves it less room to invent.

How do I correct ChatGPT about my company?

Unify owned facts first — About, schema, Google Business Profile, press kit — then search the web for the wrong claim and correct or outcompete those URLs. Re-test the same prompts over weeks with search on when the product offers it. UI feedback can help the vendor; it is not an ops plan.

Why does AI still use our old address or phone number?

Stale NAP lives in directories, PDFs, partner footers, and call-tracking leftovers long after the site changes. BrightLocal’s citation guidance exists because those rows get copied. Claim and fix the high-authority listings, replace the files, and keep one public phone format. Then re-test with the old suite as the prompt.

How do I fix a wrong founding year in AI answers?

Pick the legal formation year unless you have a documented reason to publish a different public year, then put that year in visible copy and in foundingDate. Search for the wrong year in quotes and clean every owned hit. Wikidata P571 and leftover Crunchbase rows are common amplifiers — edit them only with a source, or they will keep teaching the error.

How long until leftover facts disappear after I correct the sources?

Retrieval-based answers can improve within days or weeks once the leftover URLs change or get outvoted. Training residue and uncached scrapers can linger for months. Plan a day-3, day-14, day-30, and day-90 retest instead of a single screenshot.

Will sending feedback in the ChatGPT UI fix it?

Feedback is worth sending on a severe miss, but it is not a reliable control loop. Fix the web evidence you own and influence, then re-run the prompt in a fresh session. If the Sources panel still lists a leftover page, that URL is the ticket — not the thumbs-down.

CTA

Hallucinations have leftover URLs. Clean the packet, then get a second pair of eyes if the residue list is longer than your week.

Start on visibility or request a visibility audit.