Cleaning your keyword core: what to throw out and why

A raw core straight after collection is always a mix. Out of 500 queries, only 200-300 are truly yours. The rest is junk that, if left in, will spoil your clustering, your tracking, and your pages. Cleaning is the stage that separates an amateur approach from a professional one.

Let's break it down by category.

Category 1: informational queries on a commercial site

The most common junk. If you run a commercial site (a store, services) but your core is full of "how," "what is," "why" queries, they aren't yours. Their intent is not a purchase but learning.

What to do: either throw them out entirely (if you don't need them at all), or set them aside in a separate "blog list" — that's where informational articles go, the ones that pull traffic into the funnel but not onto your main commercial pages.

Markers in the query: "how," "what is," "why," "what for," "difference," "history," "reasons," "types."

Category 2: geo mismatches

If you're in Moscow, every query mentioning other cities gets thrown out. It's simple: "cake spb," "cake yekaterinburg" — no.

Tip: in Wordstat you can set the region to "Moscow" right away, and most geo mismatches won't show up in the results. But the right-hand column ("queries similar to") still tends to have junk from other cities. Throw those out.

Marker: a city or region name that differs from yours.

Category 3: not your product/service subtype

If you only sell kids' cakes but your core has "wedding cakes," that query isn't yours. Simply because Wordstat includes everything with the word "cake" in its results, and some of that isn't for you.

Markers: adjectives before the main word that mark out a subcategory. "kids' / adult," "men's / women's," "budget / premium," "for sole proprietors / for companies."

The decision depends on your strategy:

• Narrow the core: keep only your subtype. Minimal junk, cleaner clusters.
• Expand the business: add new subtypes to your product. Sometimes collecting a core reveals demand for something adjacent you hadn't considered.

Category 4: competitor brands

"palych cakes," "cake volshebnitsa," "u pushkina cake" — these are competitors. You won't win on their brands, and you shouldn't. Ranking for someone else's brand means either misleading the user (they were looking for a specific store) or catching a tiny CTR.

What to do: negative keywords with competitor names. As a bonus, cleaning up tells you who your main competitors are — a list you'll want later for analysis.

Category 5: wholesale

If you're a retail store but your core has "cakes wholesale," "bulk moscow," "cake procurement," those are B2B queries, not your audience. And the reverse: if you're a wholesaler, throw out the retail ones.

Markers: "wholesale," "bulk," "procurement," "price list," "for business," "b2b," "for resale," "dealers," "franchise."

Category 6: video / images / music

These are queries where the user is looking for content, not buying: "cake video," "cake photo," "cake coloring page," "cake cartoon." Not your customer.

Markers: "photo," "picture," "video," "wallpaper," "coloring page," "song," "cartoon," "game."

Category 7: DIY / by hand

Queries like "cake by hand," "how to make a cake" — again an informational intent, but an especially dangerous one, because it's often phrased similarly to a commercial query. The user wants to make it themselves, not buy it.

Markers: "by hand," "on your own," "how to make," "instructions," "life hack," "masterclass."

If your business can sell DIY kits (ready ingredients + a recipe), this junk turns into gold for you. If not, throw it out.

How to clean, technically

In Excel/Google Sheets:

1. Load the list into a single column (say, A).
2. In column B, write the junk category if a query fits one of the seven.
3. Filter for blanks in column B — that's your working core.

Or via AI: you can paste the whole list into an AI assistant with the prompt "sort these queries by junk category and keep only the commercial Moscow queries about kids' cakes." Claude/GPT handle this task quickly, especially on 200-500 queries.

Our service has a free intent breakdown — it labels queries by type (commercial / informational / navigational / transactional / mixed), and you spot the informational ones right away.

Negative keywords for Wordstat and ads

An extra bonus of cleaning: you build a stop list for the whole niche.

A stop list is a list of words you do NOT want to show up for. It's used both in Wordstat (so these queries don't creep into future research) and in Yandex.Direct (so your ads don't show for irrelevant queries).

A typical stop list for kids' cakes in Moscow:

wholesale
bulk
procurement
recipe
by hand
how to make
masterclass
video
photo
picture
wedding
adult
spb
yekaterinburg
krasnodar
... (all irrelevant cities)
... (all irrelevant subtypes)

You write this stop list once and then reuse it every time you expand the core.

Duplicates — a separate pain

After cleaning, near-duplicates often remain:

• "buy cake moscow" and "cake moscow buy" (the same thing in essence)
• "cake to order" and "order a cake" (the same thing)
• "kids' cake" and "cake for kids" (the same thing)

Wordstat treats these as different phrases and shows different frequencies. Search engines have long treated them as one and the same query (this is about normalization — why "buy" and "will buy" are one thing).

Solution: when cleaning, keep only one phrase from each pair of near-duplicates in the core. This removes 10-30% of "phantom" queries.

How long cleaning takes

A raw core of 500 queries takes 60-90 minutes to clean by hand, or 15-20 minutes with AI help. It's worth it: after cleaning, clustering yields 2-3 times fewer clusters (no duplicates), and every cluster is definitely yours.

Next article: search suggestions — a free source of 1000+ long-tail ideas.