Skip to content
allanninal.dev/build
Collaborate

Part 3 of 7 · Search rank reporter series ~5 min read

How queries get grouped

Grouping is where this system earns its keep and where the temptation to be clever is strongest. Automatic clustering is available, it works reasonably well, and it is the wrong choice for a reason that has nothing to do with quality.

Key takeaways

  • Rules you write, applied in order, first match wins.
  • Automatic clustering produces different groups each run, which destroys trends.
  • Brand queries are always their own theme and never mixed with anything.
  • Competitor-name queries are worth their own theme and are frequently forgotten.
  • The unmatched bucket is where next quarter’s themes come from.

Rules, not clusters

A model can look at nine hundred queries and produce sensible groups. Run it again next week on a slightly different nine hundred and it will produce sensible groups that are not quite the same, and now this week’s “emergency work” theme contains a different set of queries from last week’s.

Every trend built on that is meaningless, and the failure is invisible: the numbers look plausible and simply are not comparable. Rules written by a person produce the same groups forever, which is worth far more than the marginal improvement in grouping quality.

How a search query is assigned to a themeA vertical chain of five steps entered by a box labelled A query row, from the raw store. Step one asks whether it contains brand terms including your name and its misspellings; if so it exits to Brand, always first and always alone. Step two asks whether it contains a competitor name from a list you keep; if so it exits to Competitor, which is its own theme. Step three applies the theme rules in order with first match winning; a match exits to that theme and stops. Step four is Unmatched, which is a real bucket rather than a bin. Step five computes totals per theme: impressions, clicks and weighted position. A note says brand goes first always, because brand queries would otherwise contaminate every other theme.AWS ACCOUNTA query rowfrom the raw storeBrand terms?your name, misspellingsBrandalways first, always aloneyesCompetitor name?a list you keepCompetitorits own themeyesAny theme rule?in order, first matchTheme rulesyours, orderedThat themeand stopmatchUnmatcheda real bucket, not a binTotals per themeimpressions, clicks,weighted positionBrand first, always. Brand queries would otherwise contaminate every other theme.
Fig 1. How one query row finds its theme. Brand and competitor checks run before any content rule, because a brand query containing a service word would otherwise inflate that service.
  • App integration
  • Machine learning
  • Analytics
  • People

Why brand goes first

A query like “acme plumbing emergency” contains a service word and is a brand query: somebody already knows who you are. Counting it in the emergency theme inflates that theme with traffic that was never competitive, and worse, makes the theme look healthy while the non-brand part of it declines.

Separating brand out is the single most important grouping decision, and the difference it makes to a report is large: it is common for a business’s apparently strong performance in a service theme to be almost entirely people searching for it by name.

Competitor queries

Queries containing a competitor’s name that landed on your site are a small, interesting theme that almost nobody tracks. They are usually comparison searches, they convert unusually well, and a change in them is a genuine market signal rather than a ranking one.

Weighted position

The arithmetic detail that matters most: a theme’s position is the impression-weighted average of its queries, not the simple average. A theme containing one query at position 2 with four thousand impressions and forty queries at position 40 with two impressions each has a real position near 2, and a simple average would report 39.

The unmatched bucket

How the unmatched query bucket produces new themesA horizontal row of five boxes. Unmatched: eighteen per cent of impressions. Read the top twenty: once a quarter. A pattern: the phrase heat pump appearing forty times. New theme: added to the rules. Regrouped: with all history. A note says a growing unmatched bucket is the most useful signal in the whole system.WHERE NEW THEMES COME FROMUnmatched18% of impressionsRead the top 20once a quarterA pattern'heat pump' x40New themeadded to the rulesRegroupedwith all historyA growing unmatched bucket is the most useful signal in the whole system.
Fig 2. What the unmatched bucket is for. It is not a failure of the rules; it is the queue of things the business is starting to be found for.

An unmatched bucket that is growing means the site is being found for things the rules do not know about, and that is frequently the earliest signal a business gets that demand is shifting. Reading its top twenty queries once a quarter takes five minutes and is the highest-value five minutes in this entire system.

Next: how a movement is told apart from noise.

All posts