Reading intent from question phrasing

When teams classify website questions, they usually start with the page URL. It is the obvious feature: pricing page means pricing intent. That works, but it is weaker than it looks, because a lot of visitors ask a question on whatever page they happen to be on when the thought occurs to them.

Phrasing is the stronger predictor, and it costs nothing to use because it is already in the text.

Modal verbs carry stage

The auxiliary verb in a question is a surprisingly reliable stage marker.

• "Can I" and "do you" indicate capability checking. The person is establishing whether the product qualifies. Early stage.
• "How do I" indicates procedure seeking. The person has accepted the product does it and wants to know the mechanism. Middle or post purchase.
• "Why does" or "why is" indicates troubleshooting. Something did not behave as expected. Post purchase, usually.
• "What happens if" indicates risk assessment. Late evaluation stage, often the last question before a decision.
• "Which" or "should I" indicates comparison. The person has multiple options in play, either your plans or your competitors.

The last two are commercially the most interesting and the least likely to be prioritised, because they are lower volume than "can I" questions and higher value.

Specificity is a proxy for depth

A question that names a specific feature, plan, integration or limit came from someone who has read something. A question phrased in general terms came from someone earlier in the process.

A rough measure that works: count the proper nouns and product specific terms in the question. Zero means general browsing. Two or more means the person has done homework, and their question is worth reading individually rather than only in aggregate.

Negation is a strong signal

Questions containing a negative, such as "does it not support", "is there no way to", or "can I avoid", usually indicate the visitor has already found something they consider a problem and is checking whether it is a dealbreaker.

These are low volume and high stakes. If you only read one category by hand, read this one.

A practical classification approach

You do not need a model to start. A small set of pattern rules applied to the question text will sort the majority correctly, and the residue can be read manually. Something like:

stage = 'evaluate'   if question starts with can/do/does/is there
stage = 'procedure'  if question starts with how
stage = 'trouble'    if question starts with why
stage = 'risk'       if question contains 'what happens' or 'what if'
stage = 'compare'    if question contains 'vs', 'versus', 'or', 'better than'
stage = 'unknown'    otherwise

This is deliberately crude. The point is not accuracy in the abstract, it is that a crude classifier applied consistently reveals distribution shifts, and a distribution shift is the thing you actually want to notice. If risk stage questions double after a pricing change, that is worth knowing even if the classifier misfiles a third of them.

Where phrasing analysis breaks

Two failure modes are worth stating. Non native speakers phrase questions differently and a rule set tuned on native phrasing will misclassify them systematically. And visitors who paste a long block of context do not fit any of these patterns.

Both argue for the same thing: use the classification for aggregate trends, and read the individual conversations that matter rather than trusting a label. The classifier is a sorting aid, not a substitute for looking.