Day ???/100. Still not counting.
Hey—It's Tim.
Yesterday I asked you to open ChatGPT and Claude and ask them a question about your own company. "Who is [your company]? What do they sell, who runs it, what does it cost?"
If you missed that one, go and do it now. It will take you 2 mins.
If you did it, you've got a list. Wrong pricing. A founder who left in 2023. A product you sunsetted. A description of your company that was true two funding rounds ago.
Here's the shitty part: there is no support ticket for this.
You cannot email OpenAI or Anthropic…
There's no form.
No "request a correction" button.
Nobody is coming to fix it for you.
The model learned something wrong about you and it will keep telling everyone who asks until the facts it learned from change.
So you overwrite it.
Why the wrong answer wins.
We need to understand a few things first as to why this happens to know how to fix it.
When a model has conflicting facts about you, it doesn't default to your website. It weighs sources. And your own site is not the heaviest one in the pile.
Your site says pricing starts at $10. A TechCrunch piece from 2022 says $50. Your LinkedIn company page still lists the old headcount. Crunchbase has the founder who left. Wikidata has nothing at all.
Five sources, four of them wrong, and the only one you control is the one carrying the least weight.
Which is why "we updated the website" does fuck all. You updated one vote in an election you were losing 4-1. More like 12-1 if you take into site trust…
The repair, in four moves.
1. Make your About page aggressively factual.
Every entity in a knowledge graph needs a home. One canonical URL that acts as the source of truth. For you that's your homepage or your About page.
And most About pages are useless for this, because they're a design-led manifesto. "We believe in building better futures." Beautiful. Contains not one extractable fact.
Add a hard fact block. A table or a list, plainly visible in the DOM:
→ Year founded
→ Founders, by name
→ Legal entity name
→ Headquarters address
→ Key executives, with standard job titles
That last one matters more than you'd think. "Chief Happiness Ninja" is unparseable bollocks. Call them Head of People and let the NLP map the relationship.
If you're on a custom stack, mark it up properly:
<h2>Company Facts</h2>
<dl>
<dt>Founded</dt>
<dd>2015</dd>
<dt>Headquarters</dt>
<dd>San Francisco, CA</dd>
<dt>CEO</dt>
<dd>Jane Doe</dd>
</dl>And it has to be server-rendered. If those facts only exist after JavaScript runs, we're back to yesterday's problem.
One specific thing worth doing: if the model got your CEO wrong, address the exact hallucination on the page.
"Jane Doe is the current CEO, taking over from John Smith in 2021."
Naming the conflict directly helps the parser resolve it. The point here is that you’re addressing the problem and being the only place with the solution.
2. Assert your identity with sameAs.
Identity resolution is genuinely hard for a machine. Is the Acme Corp on your site the same Acme Corp on LinkedIn? On Crunchbase? If it can't tell, it won't pass the trust from those domains back to you.
sameAs is you asserting they're the same entity.
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "Organization",
"name": "Your Brand",
"url": "https://yourdomain.com",
"sameAs": [
"https://www.linkedin.com/company/yourbrand",
"https://twitter.com/yourbrand",
"https://www.crunchbase.com/organization/yourbrand",
"https://www.wikidata.org/wiki/Q123456"
]
}
</script>Exact URLs. Not twitter.com, the actual profile. Crunchbase and Wikidata carry the most weight here, so get those in.
Homepage only. Don't put Organization schema on every blog post, that confuses things. Blog posts get Article schema with a publisher property.
On WordPress, Yoast or RankMath will do this for you under Social Profiles. Webflow and Framer, paste it into head custom code. Shopify, theme.liquid wrapped in a homepage conditional.
Before you write a line of it, paste your homepage into the Schema Markup Validator and look at what's actually there. Half of you will find an Organization block already sitting on the page with an empty sameAs, which means this is a ten-minute job rather than a new build.
3. Get a Wikidata item.
This is the one almost nobody does and it's the highest-leverage thing on the list.
Wikidata is the structured backbone of Wikipedia, and effectively every major model has ingested it. Creating an item mints you a Q-ID: a permanent, unambiguous identifier for your company across the entire AI ecosystem.
Without it you're a string of text that might mean you. With it you're a concept.
Go and search it for your company name right now. If nothing comes back, that's the gap.
It's free, it's admin, it takes an afternoon:
→ Make an account. Fix a few typos elsewhere first, new accounts creating new items get scrutinised
→ Search hard to check you don't already exist. Duplicates get punished
→ Create the item. Label and a plain description ("software company based in California")
→ Add statements: instance of, official website, country, inception, founder
→ Reference every statement. A news article, a press release, a Crunchbase link. Your own site alone won't cut it
→ Take the Q-ID back to your sameAs array and close the loop
4. Fix the other sources, then wait.
Crunchbase, your LinkedIn company page, any directory listing you can still log into. These are the ones outvoting you.
Crunchbase first. It's the one that turns up in the sameAs of every company that's got this right, and it's usually the one carrying a founder who left three years ago.
Then wait. This is the honest bit: model weights don't update because you edited a page. Correction arrives either at the next training run, or through real-time retrieval if your domain has the authority to get pulled live.
Which is exactly why you start now rather than the week you notice it's costing you deals.
The actual order.
If you only do one thing today, do the About page fact block. It's an hour and it's the foundation the other three lean on.
If you've got an afternoon: fact block, then sameAs.
If you're going properly: all four, Wikidata last, because it needs the references the other steps generate.
Send me what you found.
Hit reply and tell me the worst thing ChatGPT or Claude said about your company. Wrong price, a founder who left years ago, a product you killed, whatever it turned up.
I'm collecting them. The best one goes in Thursday's email, anonymised unless you want the credit.
No form. No signup. Just reply to this email.
Tomorrow I'll show you why your content can be genuinely excellent and still be invisible to retrieval. It's a structural thing and it's fixable in an afternoon once you can see it.
Thursday is the big one. I'm giving away the 100 checks I run on a site, plus five recipes so you're not staring at a hundred rows wondering where to start. Free, no signup, no gate, nothing to fill in. That's the email to keep.
See you tomorrow.
✌️ Tim "no support ticket for reality" Hanson
CMO @Penfriend.ai
Same brain, different platforms: X, Threads, LinkedIn.
P.S. Seriously, reply. I read every one and I answer every one. It's the best part of doing this.
Penfriend.ai
Made by content marketers. Used by better ones.
