Sort 10,000 queries by what the searcher actually wants

Open any Search Console export and you will find the same swamp: thousands of queries, most seen once, each one a tiny mystery about what the searcher wanted. Teams either ignore the long tail or hand-write regexes that rot within a quarter ("best" means commercial until it means a support article about best practices).

Intent mapping is a choice question with a fixed taxonomy. Six labels, one pick each, a confidence attached. That is exactly what the seo_intent policy does — and at batch scale you can run the whole export in minutes.

The architecture

The taxonomy stays small on purpose: informational, commercial, transactional, navigational, support, other. Every extra label is a new way for two editors to disagree. If a query genuinely fits nothing, "other" plus a low confidence is an honest answer.

Build it

curl $WF_URL/predict/batch \
  -H "authorization: Bearer $WF_KEY" -H 'content-type: application/json' -d '{
  "states": [{"text": "best crm for 20 person agency pricing"}],
  "policy": "seo_intent"
}'

rows = run("intent", queries)  # examples/seo_bulk.py
commercial = [r for r in rows if r["intent"] == "commercial"]

Then the content decisions fall out: transactional queries with no landing page are build orders, informational clusters pointing at one thin page are merge candidates, and navigational queries for competitors are someone else's problem.

Gating by confidence

A thousand queries into two dozen categories costs single-digit cents — but the win is not the auto-tagging, it is the triage. Review the few hundred rows the model flags as uncertain instead of all ten thousand. The full probability distribution tells the reviewer why it hesitated, which is usually more useful than the pick itself.

Pitfalls

• Freeze the taxonomy for a quarter. If you rename labels monthly, your trend lines are fiction.
• Locale matters. "Football boots" and "soccer cleats" carry different intents in different markets — include locale in the state.
• Re-run, don't accumulate. Intent drifts as SERPs change. Schedule the export monthly and diff the tags.