SEO ·Explanation
Query-level SEO data: read beyond the aggregates
Read query-level SEO data by query and by page: split brand from non-brand, gather query variants into clusters and rank demand by the paid conversion rate.
It’s the first Monday of the month, and the report says organic search sent 8% more sessions and converted at 2.2%. Everyone nods. Nobody can answer the two questions that matter: how much of that was people searching your own name, and which queries actually did the converting? Query-level SEO data exists to answer exactly those questions, and most teams never look at it. They read organic search as one rolled-up channel row, even though it drives 60-77% of total site traffic for many businesses (Search Engine Journal). One more wrinkle: Search Console counts 134,000 clicks for the month while GA4 counts 127,000 sessions, and nobody in the room can say why. This piece is about what the aggregate hides, level by level, and where the truth lives.
Marlowe is a fictional worked example. Its figures illustrate the reading method; they are not customer results or independent benchmarks.
What is query-level SEO data, and why read organic search at that level?
Query-level SEO data is organic search performance read per search query, and per query-page pair, instead of rolled up into one channel total. The channel total is an average over things that should never be averaged: brand and non-brand demand, ten different intents, pages that compete with each other. Every decision hiding in the mix disappears the moment you roll it up.
query-level SEO data := organic search performance read per search query (and per query-page pair) rather than as one channel total, so the mix of brand and non-brand, intent, and page assignment underneath the aggregate stays visible
An aggregate is not wrong, exactly. It is unanswerable. “Organic converted at 2.2%” survives every meeting precisely because no one can interrogate it. Read at the query level, the same month’s data starts answering back.
Why your two dashboards disagree
Your two organic dashboards measure different events in different places, which is why they never reconcile. Search Console counts clicks on Google’s results page, measured on Google’s side. GA4 counts sessions that started on your site, measured by your own tag, after consent banners and blockers have taken their cut. Clicks are not sessions, and neither tool is lying; they are answering different questions. Chasing an exact match between them is a week you will never get back.
The deeper problem is what each tool withholds. GA4 knows conversions but not queries: Google stopped passing the search term along with the visit more than a decade ago, so the organic channel arrives as one undifferentiated blob. Search Console knows queries but not conversions: it watches the results page, not your checkout. Each dashboard holds exactly the half of the story the other one is missing. So the channel view can tell you organic converted at 2.2% and cannot say at what, and the query view can tell you what people searched and cannot say what it earned.
Organic search itself is awareness work: it earns the click at the wide top of the funnel and hands it to the on-site funnel below; see the pillar’s funnel section for how the stages connect, and the web-analytics sibling for the on-site half that GA4 owns.
The three levels of Search Console data
Search Console will answer at three levels of depth, and they say different things. Most teams only ever read the first.
- Property level. The site totals: clicks, impressions, average position. This is the level the channel report quotes, and it can only describe, never explain.
- Query level. Performance per search term. An impression here is counted once per search, so a query’s impressions are a real read of its demand: how many times people searched something and you were on the page.
- Query-page level. Performance per query and landing page. This is where you see which of your pages answered which search, and it carries a trap: when two of your pages appear for the same search, each page logs its own impression.
Take Iris, the marketing data analyst at Marlowe, a mid-size clothing brand with a handful of its own shops and an online store. Pulling the query “cashmere care guide”, she finds it earns 4,100 impressions at query level, but the query-page rows add up to 6,800, because Marlowe’s blog guide and its cashmere product page both rank for it (the numbers here and below are illustrative). Add up the query-page rows and you count the same search twice: read the query level for demand, the query-page level for which page answers it. The double listing is also a finding in itself. Two Marlowe pages are competing for one intent, splitting clicks and link equity between them, the failure known as keyword cannibalization. The property level cannot see it, the query level only hints at it, and the query-page level hands you the fix.
One search, two rows (illustrative). Here is a single search for “cashmere care guide”, as each level records it:
| What Google showed for one search | Position | Property logs | Query logs | Query-page logs |
|---|---|---|---|---|
| Marlowe’s blog care guide | 3 | (rolled up) | (rolled up) | 1 impression, blog page |
| Marlowe’s cashmere product page | 6 | (rolled up) | (rolled up) | 1 impression, product page |
| the search itself | 1 impression | 1 impression | 2 rows |
One person searched once. The query level counts it once, a true read of demand. The query-page level counts one impression per page shown, so the one search becomes two rows. Now the month Iris pulled, with the overlap made visible:
| “cashmere care guide”, one month (illustrative) | Searches it appeared in | Avg position |
|---|---|---|
| Marlowe blog care guide | 3,900 | 3.2 |
| Marlowe cashmere product page | 2,900 | 6.4 |
| Query-page rows, summed | 6,800 | |
| Query level, real searches | 4,100 | 3.1 |
| Searches where both pages appeared | 2,700 |
The gap is the finding: 6,800 less 4,100 is 2,700 searches, two-thirds of this query’s demand, where Marlowe answered its own search twice and split the clicks and link equity between a guide and a product page. Note too that the query-level position (3.1) is the topmost rank Marlowe held per search, not an average across its pages, so the blog page’s strong rank is quietly masking how weakly the product page shows.
The first honest cut: brand versus non-brand
The single most misleading thing about that 2.2% blended conversion rate is that it averages two different businesses. Brand queries are people who already know Marlowe: they typed the name, they were coming anyway, and they convert the way warm traffic always converts. Non-brand queries are strangers discovering the category: “best cashmere jumper”, “wool overcoat women”. They are the only part of organic search that acquires new customers, and they convert like the cold traffic they are.
Blend the two and each hides the other. Iris splits the month and the comfortable 2.2% decomposes into brand converting at 6% and non-brand at 0.7%. Against the roughly 2-3% that ecommerce sessions convert at on average (Adobe / Statista benchmark), the blend looked healthy. The split says the growth engine is weak and the aggregate was borrowing warmth from the brand traffic to hide it. A blended organic conversion rate is brand demand lending its numbers to a non-brand problem. The same split keeps the traffic story honest too: non-brand clicks are discovery, brand clicks are demand your other marketing already created, and the SEO performance sibling walks that read stage by stage.
One demand, many spellings
Search Console reports queries as literal strings, so one intent arrives shattered into dozens of rows. The person who wants a cashmere jumper types “cashmere jumper”, “cashmere sweater”, “cashmire sweater”, “kashmir cardigan”, singular, plural, misspelled, and in whatever language they think in. Every variant is its own row in the export. Sorted by clicks, each shard looks too small to matter, and the flat file buries the pattern.
That fragmentation is why query-level SEO data read keyword by keyword understates your real demand. Iris thought she knew the cashmere line’s search demand from its top query; when she gathers the variants, the demand roughly doubles. Nothing about the market changed. Only the level of reading did. The unit that makes demand legible is not the keyword but the query cluster: every variant of one intent, read together, and, one level up, the cluster hierarchy that lets her read knitwear as a whole, then cashmere, then one product’s queries, without rebuilding anything. How Maya builds those clusters, and keeps them from going stale, is what the next section shows.
The tip and the iceberg (illustrative). Iris’s top-query read of the care intent, against what its variants actually total:
| Variant of one intent | Impressions | Share |
|---|---|---|
| cashmere care guide (the top query she knew) | 4,100 | 47% |
| how to care for cashmere | 1,350 | 15% |
| how to wash cashmere | 980 | 11% |
| cashmere care | 640 | 7% |
| washing a cashmere jumper | 410 | 5% |
| cashmere maintenance | 260 | 3% |
| misspellings (casmere care, cashmere care guied, and so on) | 150 | 2% |
| roughly 120 long-tail variants of three words and up | 890 | 10% |
| The intent, read as one cluster | 8,780 | 100% |
The top query is 47% of its own intent: read keyword by keyword and you size this demand at 4,100, read the cluster and it is 8,780. The market did not double, the reading did. The long tail alone is over a hundred variants that churn monthly, which is why no one maintains this by hand. The same rules that gather the variants also roll up, so Iris reads at whichever altitude the question lives:
| Altitude | Cluster | Impressions | The question it answers |
|---|---|---|---|
| Level 1 | Knitwear | 61,200 | how is the category doing? |
| Level 2 | Cashmere | 24,400 | is the line growing? |
| Level 3 | Cashmere, care intent | 8,780 | should the guide get investment? |
| Level 3 | Cashmere, product queries | 9,950 | is purchase intent there? |
One hierarchy, no rebuilding: the care intent is the same 8,780, now in context, and a third of the cashmere line’s demand turns out to be people asking how to look after the product, not how to buy it.
The paid query is your organic crystal ball
Here is the join that neither dashboard can make alone. Search Console has no conversions and GA4 has no queries, so an organic conversion rate per query does not exist anywhere in your stack. But Marlowe bids on many of those same queries in paid search, and on the paid side every query has a measured conversion rate: same searcher, same intent, same moment of demand, landing on your site with the meter running. Cross-industry, search clicks convert around 7% (WordStream), but the average is not the point. The per-query spread is.
That paid conversion rate is the best available estimate of what an organic click on the same query would be worth: call it the query’s theoretical CVR. Your paid search program has been running, at full price, the conversion experiment your SEO program could never afford. The read reorders Iris’s priorities by conversion quality instead of raw volume:
- High paid CVR, weak organic rank. The queries where page-one organic positions are worth real money, since the top three results take 68.7% of clicks (Sistrix). This is where the next quarter’s content and links belong.
- High paid CVR, strong organic rank. Demand you already own; the question becomes whether you should keep paying for it too, the brand-leakage check the paid-search sibling covers.
- Low paid CVR, any rank. Informational demand. Worth owning for awareness, but never the priority for a conversion goal, however large its volume looks.
Ranked by that estimate, raw volume stops leading (illustrative):
| Query cluster | Monthly searches | Paid CVR | Organic position | The read |
|---|---|---|---|---|
| womens cashmere sale | 2,100 | 11.4% | 14 | Build now: converts hard, stuck on page two |
| cashmere crewneck | 3,600 | 8.7% | 8 | The leverage zone, bottom of page one |
| best cashmere jumper | 2,900 | 5.0% | 5 | Worth a nudge |
| cashmere care | 8,780 | 1.2% | 3 | Informational: awareness only, never the priority |
| marlowe cashmere (brand) | 2,400 | 14.8% | 1 | Already owned: the brand-leakage question |
Sort that by volume and the care intent (8,780) tops the list, the very intent the cluster section flagged as mostly “how to look after cashmere”. Sort by the theoretical CVR and it drops to the bottom for a conversion goal, while “womens cashmere sale”, on a fraction of the volume, becomes the first content bet.
State the caveat plainly: this is an estimate, not a measurement. Ad clicks land on tuned landing pages, an ad’s audience skews from an organic snippet’s, and position changes trust. The theoretical CVR ranks your opportunities; it does not promise the outcome.
One scoping note before the tooling: this exercise ranks by conversion rate because a store’s goal is the sale. A business whose model is attention rather than purchase, an ad-supported publisher earning on every pageview, runs the identical paid-to-organic read but ranks by click-through and click volume instead, because a visit is the outcome it sells. Which metric sits on top is a business-model choice, and the full-funnel pillar covers it.
How Maya turns scattered data into answers
Most SEO reporting stops at description: here are your clicks, here is your average position. The work that matters lives further up the value pyramid: diagnosis (why the number moved), decision (what to do), outcome (did it work). Maya is one implementation of the principle that the analytics layer should climb that pyramid for you, on definitions it cannot quietly reinterpret. The foundation comes first, in order:
- A marketing data warehouse unifies every platform’s raw exports in one cloud database: Search Console next to GA4, Google Ads and the store’s own orders, reconciled so the numbers become comparable.
- A governed Power BI semantic layer defines every metric once on top of it, so a click, a session and a conversion each mean exactly one thing in every answer, and the two-dashboards disagreement dissolves into two well-defined columns instead of one confused argument.
- A conversational layer, a Model Context Protocol (MCP) server, lets an AI assistant (Claude, ChatGPT, a Slack bot) query that governed layer in plain language. Prompts steer; the semantic layer enforces. That split is what separates a governed MCP from a chatbot wired to a raw database.
On that foundation, the query-level SEO data reads this piece has walked through stop being spreadsheet projects:
- The brand-versus-non-brand split comes packaged. Maya separates the queries that carry Marlowe’s name from the ones that don’t, so the honest cut is the default view, not a filter someone has to remember to maintain.
- Query and query-page levels are both first-class. Demand read at query level, page assignment at query-page level, so the double-counting trap and the cannibalization finding fall out of the same question.
- Paid and organic sit in one model at the query level. The theoretical-CVR read is a question, not a project: organic clicks and rank next to the paid conversion rate on the same queries, with the honest framing that the paid number is an estimate of the organic one, not its measurement.
Clusters: every variant, any altitude
The cut that pays off most is the cluster. Iris doesn’t just want the cashmere variants gathered; she wants every product category Marlowe sells, at every level of depth, misspellings included. In Maya’s UI she builds hierarchical query clusters from simple text-matching rules, no code, up to ten levels deep: level one splits brand from non-brand, level two groups by category (knitwear, shirts, outerwear), level three by line (cashmere, linen, wool coats), and on down. A rule that matches “cashmer” alongside “kashmir” catches the misspellings and variant spellings in one definition. Each level previews live as she tweaks a rule and refreshes on its own as new data loads into the model, so a cluster is defined once and never goes stale. Now she reads demand at any altitude, all non-brand, then knitwear, then just cashmere, as one consistent hierarchy instead of a flat export thousands of rows long.
flowchart TD
Brand["Brand · level 1"]
NonBrand["Non-brand · level 1"] --> Knitwear["Knitwear · level 2"]
NonBrand --> Outerwear["Outerwear · level 2"]
Knitwear --> Cashmere["Cashmere · level 3"]
Knitwear --> Merino["Merino · level 3"]
Outerwear --> Wool["Wool coats · level 3"]
Here is the read it gives her, with the theoretical-CVR column doing the prioritizing (illustrative numbers):
| Query cluster | Demand (impressions) | Your organic clicks | Paid conversion rate, same queries | What Iris reads |
|---|---|---|---|---|
| Non-brand › Knitwear › Cashmere | 41,000 | 1,900 | 5.8% | high-value demand, weak capture: next quarter’s content budget |
| Non-brand › Knitwear › Merino | 18,000 | 2,400 | 2.1% | decent capture, modest value: hold |
| Non-brand › Outerwear › Wool coats | 26,000 | 3,100 | 0.6% | volume without conversion intent: awareness, not priority |
Cashmere jumps the queue and wool coats drop down it, on conversion quality the flat export could never see.
Without Maya: Iris exports Search Console at query level, exports it again at query-page level, dedups the overlap by hand, builds the brand filter, pulls the paid search terms report, joins the two on query in a spreadsheet, and a day later has one month’s answer that is stale by the next report. With Maya: she asks “Which non-brand clusters have high paid conversion but weak organic rank?” and the governed model reads organic demand, rank and clicks next to the paid conversion rate on the same queries, ranked and diagnosed. Minutes, not a day.
The questions a marketer actually asks Maya:
- “How much of my organic traffic is brand, and how does each half convert?” The packaged split, read next to the conversion data in the model.
- “Which queries rank on two of my pages at once?” The cannibalization shortlist from the query-page level.
- “Which non-brand clusters have high paid conversion but weak organic rank?” The theoretical-CVR priority list.
- “How did the cashmere cluster’s demand move this quarter, misspellings included?” One cluster, every variant, any altitude.
The honest caveats are about the source data, not the tool: Search Console’s numbers arrive with a two-to-three-day lag and long-tail sampling, and they are queries, not people. Conversion reads come from the GA4 and store data unified in the governed model, queryable alongside the organic side, and paid conversion figures depend on the conversion actions being configured upstream. The theoretical conversion rate stays an estimate. Maya does the assembly and the guardrails; you make the calls.
The bottom line
An aggregate organic number is an average over brand and non-brand, over ten intents, over pages competing with each other, and averages of unlike things answer nothing. Read query-level SEO data in that order: demand from query-level impressions, page assignment from the query-page level, the brand split before any conversion claim, variants gathered into clusters, and paid conversion rates on the same queries as the prioritizer your organic data cannot supply. Or let one governed model hold all of it, so each of those reads is a plain-language question instead of a week of exports.
Next first-Monday, the report will say organic converted at 2.2% again. What will you ask it?
Glossary
- Query-level SEO data: organic search performance read per search query (and per query-page pair) rather than as one channel total, so the mix underneath the aggregate stays visible.
- Property-level data: the site-total view of organic performance (all clicks, all impressions, one average position). Describes, never explains.
- Query-page level data: performance per query and landing page pair. Reveals which page answers which search; impressions here are logged per page, so two of your pages on one results page each count one.
- Brand vs non-brand queries: brand queries contain your name or a close variant; non-brand queries do not. Brand is demand your other marketing created; non-brand is new-customer discovery. Split them before reading any rate.
- Conversion rate: conversions divided by clicks or sessions, depending on the side measuring it. Ecommerce sessions average roughly 2-3%.
- Theoretical CVR: the paid-search conversion rate on a query, used as the best available estimate of what an organic click on the same query would convert at. An estimate for prioritization, not a measurement.
- Keyword cannibalization: two or more of your own pages ranking for the same query and splitting its clicks and authority between them. Visible only at the query-page level.
- Query cluster: a group of closely related search queries that share one intent (variants, plurals, misspellings included), read together rather than string by string. In Maya you build these from simple text-matching rules in its UI.
- Landing page cluster: a set of landing pages grouped by a shared text-matching rule (a section, a template, a product line), read together rather than page by page, likewise built in Maya.
- Cluster hierarchy: the nested levels of a cluster, up to ten deep, from brand/non-brand at the top down through category and line, so you read demand at any altitude as one consistent tree.
Related
- the full-funnel pillar: how every discipline maps to one model
- SEO performance: the stage-by-stage read, impression to click to new demand
- web analytics: the on-site half of the story that GA4 owns
- search ads: the paid side of the query, and the brand-leakage check
- competitive research: who captures the demand your clusters reveal
- generative engine optimization: the emerging surface where queries become AI answers
Ask this about your own account
Connect your sources, then ask Maya the question this piece just answered, on your own numbers.