What it is NOT good for
Let us start with the most important one, because this is the most common and the most expensive mistake. Do not ask a language model for search volume. Not because it is imprecise, but because the number is not produced by measurement: the model follows patterns, and "2,400 searches a month" is exactly the kind of pattern that looks credible.
This is not a theoretical worry. Research into how models work internally shows that in many cases the system "knows" it has no grounded answer to the question — there is a measurable internal signal for it — and answers anyway. The reasoning variants are not better at this; according to one study they are worse: they hold back from answering less often than the simpler models. So the very kind of model most people would expect the best answer from is the one that invents the number with more confidence.
But the "real" number is not real either
Here comes the other half of the truth, the half that keyword tool marketing does not emphasise. The numbers in the keyword planner and in the analytics suites are estimates too, not measurements.
According to one large tool vendor's own public study, their estimate gives a realistic value for roughly 60 percent of keywords, and the search engine's own planner for 45 percent. So even the "official" number misleads you on every second phrase. The lesson is not to throw the tools away — it is to treat search volume as an order of magnitude, not to the decimal. The difference between "lots of people search for this" and "hardly anyone does" matters; the difference between 2,400 and 2,900 does not.
Where it genuinely helps
Four jobs where a language model really does beat the traditional tools.
- Generating questions in real language. "What does a five-person accounting firm ask when it first starts thinking about AI?" — that produces twenty questions that are in no keyword tool, because nobody types them in exactly that form.
- The customer's vocabulary. The trade term and the layman's term are rarely the same. The model knows both and can translate between them.
- Clustering. Sorting five hundred phrases into themes is a job you would spend hours on by hand, and the model does it to an acceptable standard in minutes.
- Finding the gaps. "What other questions come up in this subject that I have not written about?" — the model is good at this because it sees what is missing instead of ranking what is already there.
Intent classification, where it loses
There is one job where the measurement produced a surprising result. Classifying search intent — whether someone wants to learn, to buy or to navigate — is a classic machine learning task, and plenty of AI tools advertise that a language model does it better.
According to a peer-reviewed study on a standard data set, it does not. A traditional, supervised baseline model scored around 0.74; the best setting of the large language model reached 0.69, and the smaller variant 0.58. In other words, the old, cheap method beat the expensive one. That is a useful reminder that not every task got better because we put a language model on it.
The recipe
1. START WITH THE BUYER, not the keyword.
Write in one sentence who the content is for and in what situation.
2. ASK FOR 30 QUESTIONS, the way they would say them.
"Not keywords: full questions, the way a customer would
type them or say them, in plain English, no jargon."
3. TAKE THEM INTO A REAL TOOL and look at the volume there.
NEVER ask the model for a number. The order of magnitude is the point.
4. HAVE THE REST CLUSTERED into themes, and name the clusters.
That becomes the skeleton of your content plan.
5. DECIDE THE INTENT yourself, cluster by cluster.
One sentence: "someone searching this wants to do what in five minutes?"
Common misconceptions
- "The model knows the search data." It has no access to it. It produces the number from patterns.
- "A reasoning model is more careful." One measurement says the opposite: it less often says that it does not know.
- "The keyword tool's number is exact." That is an estimate too, and it misleads on half the phrases. Use it as an order of magnitude.
- "Every NLP task got better with a language model." In intent classification a cheap, old method beat it.
What to take away
In keyword research a language model manufactures and organises questions; it does not supply data. Always take the number from a tool — and take that as an order of magnitude too. The biggest gain is not in the list, it is that the model helps you think in the customer's words. The next step is turning that into publishable material: the Content Briefs That Work post covers it. And if the goal is showing up in AI answers, the question list is the starting point there as well — How to Get Cited by AI Answer Engines sets out what counts.