Every crawler that visits your site does roughly the same thing. It loads the page, scrapes the HTML for text, and runs natural language processing over it to work out what the article is about. That’s true for Google, and it’s true for the AI services reading the web now: Perplexity, ChatGPT, Claude, and everything else in that category. Running NLP is cheap, and cheap is what wins when you’re indexing the whole internet.
What that process finds becomes what the machine believes your article is about. What it misses never enters the picture at all.
So it’s worth looking closely at what it misses. We built a tool that runs both analyses side by side on any URL: what Google’s natural language processing detects, and what TopicalBoost’s topic map detects. Across roughly three-quarters of a million articles we’ve analyzed, about 30% of the topics we surface come from the AI layer we’ve added on top of Google NLP. Here’s what that 30% actually looks like in practice.
Cato at Liberty, on New Jersey’s voter registration problem
Google NLP finds five things: The New York Times, the United States Congress, New Jersey, John Samples, and Steven Richer. Samples and Richer are the Cato scholars quoted in the piece.
Read that list back and try to guess the subject. You get “New Jersey,” and that’s about as far as it takes you.
TopicalBoost adds the National Voter Registration Act of 1993, voter registration, and electoral integrity. The interesting one is the Act, because the article never names it. It refers to “motor voter,” the nickname the bill picked up because it tied vehicle registration to voter registration. Google NLP is literal. If the phrase isn’t in the text, it isn’t in the analysis.
AEI, on the pushback against data centers
James Pethokoukis wrote a piece about local opposition to data center construction. Google NLP finds The New York Times, NIMBY, amazon.com, artificial intelligence, OpenAI, and Meta.
It does not find “data center.”
That’s the whole subject of the article, and the most-used language model in the world read the piece and didn’t register it as a topic. TopicalBoost picks it up, along with capital expenditure, the productivity paradox, and a reference to The Information: A History, a Theory, a Flood.
The Urban Institute’s 2025 impact report
Google NLP finds the United States Senate, “house,” and Tim Scott. The “house” is ambiguous. It may have matched the building, not the House of Representatives.
TopicalBoost finds affordable housing, the Road to Housing Act, the United States House of Representatives, mortgage loans, the legislative process, manufactured housing, and housing policy in the United States as an abstract subject.
That last one is the part worth sitting with. Housing policy is not a phrase you can point to in the document. It’s the category the document belongs to, inferred from everything else in it. Google NLP doesn’t make that leap, because it isn’t built to.
Illinois Policy, on a career and technical education bill
The article covers a bill on Governor Pritzker’s desk that would let high schoolers take more career classes. Google NLP finds Illinois, the Bureau of Labor Statistics, and the American Institutes for Research.
From those three, this could be a story about unemployment numbers. It could be a story about anything happening in Illinois.
TopicalBoost finds vocational education, the Illinois General Assembly, workforce development, J.B. Pritzker, unemployment, and career and technical education.
The resolution problem
Working with Illinois Policy, we noticed something specific about the names Google NLP won’t resolve. It reads “J.B. Pritzker” in the text and declines to connect it to the Knowledge Graph entry for the governor of Illinois. Same with Stacy Davis Gates, who runs the Chicago Teachers Union.
Even “Chicago Teachers Union” gives it trouble, and that seems unambiguous until you notice what happens on second reference. Articles shorten it to CTU. CTU is also Colorado Technical University. Faced with that, the model often declines to commit, so the words appear in the article and the entity never gets attached.
Seeing a name is not the same as knowing who it refers to. Only the second one is worth anything to you.
What changes when you declare it explicitly
None of this is academic. When you feed the full topic set back to Google as structured data and internal links, the machine stops guessing.
Illinois Policy implemented TopicalBoost and analyzed its full archive. Google Discover clicks went from roughly 49,000 a month before activation to roughly 146,000 a month after, a tripling. All those articles where the entity never resolved are now declaring, in markup, that they are about J.B. Pritzker and Brandon Johnson and the Chicago Teachers Union. The pool of articles Google is willing to feature grew accordingly.
The Commonwealth Foundation went from essentially no presence in Google Top Stories to twelve appearances in the eleven months after activation, nine of them at position one or two, as major legislative fights moved through the Pennsylvania General Assembly.
The pattern repeats across clients. Identify the topics correctly, declare them in markup, link them internally, and traffic rises across Search, Google News, and Discover together.
Try it on your own content
We’re making this comparison tool available on topicalboost.com shortly. Drop in a URL from your own site and it will show you both analyses: what Google’s NLP detects, and what it’s leaving on the table.
If Google’s NLP can’t tell what your article is about, it’s a fair bet Google can’t either.