Stop your AI confusing Alphabet with alphabet
Alphabet the company, alphabet the ABC: that is the kind of mix-up I mean. The ones we found in our own news collection were stranger.
The quick way to match a name is to look up the spelling and take the best-known entry. In articles we collected in 2026 for Ins7ghts, our research platform (a knowledge graph of the data and AI news we read every day), a quick match like that said Okta was Oktay, a first name. Augury was a video game from 2021. FMR LLC, the holding company behind Fidelity Investments, was fMRI, the brain scan.
Each time, the sentence right next to the name said otherwise. “Strike Graph leans into Okta’s blueprint for a secure agentic enterprise.” “Augury is a vibration-based machine health specialist.” “FMR LLC added 7,424,130 shares … to their portfolio in Q1 2026.” A security company, a machine health company, an investor. The name alone could not tell. The sentence could, and nobody needed to know the companies to read it.
To say which one a name means, you need something real to point at. That is what Wikidata is for. Wikidata is the free, public register behind Wikipedia: every real company, product, person and place gets one entry, with a short description and its website. Point a name in an article at one entry and you have said “this is the company I mean”. Two articles that point at the same entry are about the same company, however they spell it.
How do you stop AI mixing up companies with the same name?
We measured it on our own news collection. We took 30,000 random mentions of names in news articles from 2025 and 2026. The quick name match pointed to the right Wikidata entry for 63.5 in 100 of them. Deciding each mention from its own sentences, choosing among the Wikidata candidates or none, was right for 97.9 in 100. It also found more right entries: 518 in every 1,000 mentions, against 421.
That is how we link names now. We took 300 of the links written between July and mid-October 2026 and had them decided again: 99 in 100 were right.
The check, step by step
- Never match on the name alone: take the sentence it sits in.
- List the real candidates for that name: a company, a product, a person, a common word.
- Ask which one the sentence describes: what it does, who it deals with.
- If the sentence does not decide, leave it unlinked. If the name has no entry at all, write “no link” and keep the company’s website instead. Never borrow the entry of a famous namesake.
- Decide again in every article: the same name can mean another thing elsewhere.
Once you know which company a name means, the next question is whether the article is about it at all: Is the article about them, or just naming them?
The kit has the prompts and the checklist. It comes by mail once you confirm your email.
Small print. 30,000 mentions, one from each of 30,000 random articles of 2025 and 2026, from the news on data, AI and business we collect. “Right” means: other AI models, given the same sentences and not shown our answer, came to the same decision. No person checked the 30,000 or the 300, so read the numbers as careful readers agreeing, not as a guarantee. About half of all mentions end with no link, and for most of them that is the right answer: small companies, new products and most people have no entry. The 99 in 100 covers links written for July to mid-October 2026; a link is written only when the decision is clear. These numbers are for our way of deciding, on our news collection, not a score for any prompt you run.