키워드 vs 구글 분류: 시끄러운 수직 뉴스 알림 수정

Your industry alert is still a string match

A “banking desk” built on q = bank will happily ingest river banks, blood banks, data banks, and robbery blotter. A “rock feed” on q = rock will mix geology, wrestling, and cocktail recipes with actual Rock Music coverage. The channel looks productive; the product is not on-vertical.

This page is a bakeoff diagnostic for noisy keyword desks. It is not the full Google taxonomy filter cookbook (exact path vs prefixTaxonomy, Stories, classify). That lives on Perigon’s News API filter by Google taxonomy guide – link out when you are ready to wire the params.

1. Why do vertical keywords inflate alerts?
2. How large are the gaps for bank, fund, and rock?
3. When should you use taxonomy vs Perigon category vs topic?
4. What bakeoff should you run before shipping a desk?

Why vertical keywords inflate alerts

Keywords match tokens. Vertical desks need aboutness. Google Content Category paths (V2) on Perigon articles describe industry and content family – Banking under Finance, Rock Music under Arts & Entertainment – so a filter can mean “this kind of story,” not “this substring appeared.”

Habit What you get q = bank / banking Central banks, rivers, features named “bank,” crime, metaphors q = fund Mutual funds plus “fund a project,” fundraisers, verb uses q = rock Music plus geology, sports nicknames, food, idioms Taxonomy path for the vertical Articles classified into that Google hierarchy

You still stack dates, sources, and companies. The fix is choosing the right aboutness filter first.

Live bakeoff: bank, fund, rock (30 days)

Articles counts for pubDate 2026-06-25 through 2026-07-25 (round in product copy; huge open queries are order-of-magnitude):

Desk intent Keyword filter About how many Taxonomy / structured filter About how many Banking vertical q = bank ~988,000 taxonomy = /Finance/Banking/Other ~363,000 Banking (wordier keyword) q = banking ~259,000 prefixTaxonomy = /Finance/Banking ~363,000 Funds vertical q = fund ~695,000 taxonomy = /Finance/Investing/Funds ~398,000 Rock music vertical q = rock ~341,000 taxonomy = /Arts & Entertainment/Music & Audio/Rock Music ~210,000

Read the table carefully. Taxonomy is not always the smallest number – q = banking (~259k) is below Banking taxonomy (~363k) because many banking stories never use the token “banking.” The failure mode is different: keywords miss on-vertical pieces and include off-vertical tokens. Taxonomy stays on the Google Banking branch either way.

Keyword bank (~988k) is still ~2.7× the Banking taxonomy slice – that gap is mostly noise and adjacent senses. Keyword fund (~695k) vs Funds path (~398k) shows the same pattern. Keyword rock (~341k) vs Rock Music (~210k) cuts a large geology/idiom/sports tail while keeping tour and album coverage the music desk wants.

For a wider finance firehose, prefixTaxonomy = /Finance was about 3.4M in the same window – useful as a branch monitor, not a tight desk. Pair a narrower leaf path when alerts must stay readable.

Taxonomy vs category vs topic (pick one job)

Perigon exposes three related dimensions. Mixing them up recreates keyword chaos with nicer names.

Filter What it is Use when taxonomy / prefixTaxonomy Google Content Category paths Industry / content-family desks (Banking, Rock Music, Health Conditions) category Perigon broad themes (Finance, Tech, Health, …) Fast dashboard buckets; e.g. category = Finance was ~3.5M in this window topic Granular Perigon topics Storyline types – e.g. topic = Cryptocurrency ~187k vs q = cryptocurrency ~89k (topic can out-recall a thin keyword)

Rule of thumb: Google path when the desk is an industry tree; Perigon category for coarse theme chrome; topic when you need a named storyline slice. Stack carefully – do not OR three overlapping definitions of “finance” and call it precision. Deep param behavior stays on the taxonomy filter guide linked above.

Bakeoff checklist before you ship a vertical alert

1. Write the desk intent in one sentence (“Rock Music coverage,” not “anything with rock”).
2. Run q = token and the matching Google taxonomy (or prefixTaxonomy) for the same from/to window.
3. Spot-check twenty titles from each list – count off-vertical hits and obvious misses.
4. If the keyword is quieter but incomplete (like q = banking vs Banking path), prefer taxonomy and accept the larger on-vertical set.
5. Only then add source packs, companies, or labels – aboutness first, polish second.

Key takeaways

• Vertical q filters are string matches; taxonomy paths are aboutness filters for industry desks.
• Live gaps matter: bank ~988k vs Banking taxonomy ~363k; rock ~341k vs Rock Music ~210k (30 days ending 2026-07-25).
• Quieter keywords are not always better – banking can under-recall vs the Banking path.
• Use taxonomy for industry trees, category for broad themes, topic for granular storylines.