WriteMySEO / Blog / Decomposing a Traffic Drop Into Rank, CTR, and Coverage
SEO data

Decomposing a Traffic Drop Into Rank, CTR, and Coverage

A click-level shift-share method for Search Console data that separates lost queries, lost impressions, and lost click-through rate — plus where the arithmetic misleads you.

"Organic clicks are down 18% month over month" is not a finding. It is four different things happening at once, summed into a single number: you stopped appearing for some queries, you appear less often for others, you rank lower where you still appear, and the SERP takes a bigger cut of the clicks where your position didn't move at all.

Those four causes have completely different fixes. Collapsing them into one percentage guarantees you will investigate the wrong one. The good news is that Search Console's data model makes the split mechanical — clicks are a product of two measured factors, and products decompose cleanly.

The identity that makes a drop decomposable

For any slice of Search Console data, clicks is an exact product:

clicks = impressions × CTR

That is not a model. CTR in the Performance report is defined as clicks divided by impressions, so the identity holds by construction at every level of aggregation. Which means the change in clicks between two periods can be split into a volume term and a rate term with no estimation:

Δclicks = (i₁ − i₀)·c₀ + i₀·(c₁ − c₀) + (i₁ − i₀)·(c₁ − c₀)

The three terms are the impression effect (at the old CTR), the CTR effect (at the old impression count), and the interaction. They sum to the total change exactly. This is standard shift-share arithmetic and it is the only part of this analysis that is beyond dispute.

The trick is applying it at the right grain. Do it on property totals and you get a meaningless average. Do it per query, then aggregate the terms, and you get a ranked list of where each effect came from.

Split the query set before you compute anything

Aggregate CTR change is contaminated by set composition. If you lose 2,000 head queries with 1% CTR and gain 2,000 long-tail queries with 9% CTR, your "CTR improved" while your business got worse.

So partition first. Full outer join the two periods on query (or query × device × country, if you have the volume to support it) and bucket:

The three buckets plus the three shift-share terms inside retained sum to your total click delta. Now a drop has an anatomy:

ComponentWhat it meansTypical cause
Lost queriesYou stopped surfacing entirelyDeindexing, canonical consolidation, falling off page 2+
Gained queriesNew surface areaNew pages, new rankings, query fan-out shifts
Impression effectSame queries, shown lessPosition drops, demand decline, fewer page-2 loads
CTR effectSame exposure, fewer clicksPosition drops within page 1, SERP features, title changes
InteractionBoth moved togetherUsually small; report it, don't interpret it

If lost queries dominate, you have an indexing or consolidation problem. If the CTR effect dominates with flat position, you have a SERP layout or snippet problem. Those investigations share nothing.

Impressions are not demand, and this is where it lies to you

The decomposition is arithmetically exact but causally ambiguous in one specific place: the impression term conflates search demand with your visibility.

Google counts an impression when your result appears on the results page the user loaded. Results on page two are only counted as impressions when the user navigates to page two. So a drop from position 9 to position 12 shows up as an impression collapse, not a position change — the query looks like demand died when in fact you fell off the first page. Average position for that query may even improve, because the only impressions still counted are the rare page-two loads where you happened to rank 11th.

Two defenses, neither perfect:

  1. Check the position distribution, not the average. Pull average position for each query in both periods alongside the impression delta. A large impression drop with position crossing the ~10 boundary is a ranking event, not a demand event.
  2. Use a control cohort. Pick queries where your average position moved less than, say, 0.3 across the two periods, and measure their aggregate impression change. That is your best within-property estimate of category-level demand movement. Subtract it from the impression effect on the rest. This is a difference-in-differences argument and it holds only if the control queries face the same seasonality as the affected ones — which you should check, not assume.

Where the CTR term comes from

CTR change has two sources: you moved within the page, or the page changed around you. Separate them by building a position→CTR curve from your own data (median CTR by rounded average position, segmented by at least brand versus non-brand, since the curves differ sharply). Then for each retained query compute an expected CTR at the new position and split:

The residual is where title rewrites, lost rich results, new ad load, and AI Overviews live. Google has stated that clicks and impressions from links in AI Overviews are included in the overall Performance report totals rather than broken out, so you cannot attribute the residual to AI Overviews from Search Console data alone. You can only observe that clicks fell at constant position and go check the live SERP.

Treat the residual as a flag, not a measurement. Your own CTR curve is noisy, intent-dependent, and fitted on the same data you are testing.

Two data artifacts that will skew the buckets

Anonymized queries. Search Console withholds rare queries for privacy, so the sum of query rows is less than the property total. The gap is not random: it is concentrated in low-volume, long-tail queries — exactly the ones that flip between the lost and gained buckets. In the BigQuery export, filter on is_anonymized_query = FALSE and report the covered share of total clicks alongside your decomposition. If query-level rows cover 60% of clicks, say so.

Period selection. Use whole weeks, equal length, aligned by weekday, and prefer a year-over-year comparison for anything seasonal. The Search Console UI retains roughly 16 months, which is just barely enough for a YoY comparison and not enough to do it twice — another reason to set up the BigQuery export before you need it.

Doing it on the BigQuery export

Remember that the export stores position as a sum, so average position is SUM(sum_top_position)/SUM(impressions) + 1:

WITH per_period AS (
  SELECT
    IF(data_date >= '2025-04-01', 'p1', 'p0') AS period,
    query,
    SUM(clicks) AS clicks,
    SUM(impressions) AS impressions,
    SAFE_DIVIDE(SUM(sum_top_position), SUM(impressions)) + 1 AS avg_pos
  FROM `proj.searchconsole.searchdata_site_impression`
  WHERE data_date BETWEEN '2025-03-04' AND '2025-04-28'
    AND search_type = 'WEB'
    AND is_anonymized_query = FALSE
  GROUP BY period, query
)
SELECT
  COALESCE(a.query, b.query) AS query,
  IFNULL(b.clicks,0) - IFNULL(a.clicks,0) AS d_clicks,
  (IFNULL(b.impressions,0) - IFNULL(a.impressions,0))
    * SAFE_DIVIDE(a.clicks, a.impressions) AS impression_effect,
  IFNULL(a.impressions,0)
    * (SAFE_DIVIDE(b.clicks,b.impressions) - SAFE_DIVIDE(a.clicks,a.impressions))
      AS ctr_effect,
  a.avg_pos AS pos_before, b.avg_pos AS pos_after
FROM (SELECT * FROM per_period WHERE period='p0') a
FULL OUTER JOIN (SELECT * FROM per_period WHERE period='p1') b
  USING (query)
ORDER BY d_clicks;

Rows where a is null are gained queries; where b is null, lost. Sum the effect columns to reconcile against the total delta before you interpret anything — if they don't add up, your period filter or join key is wrong.

What to do with it

  1. Run the decomposition at query grain, reconcile the terms to the total click delta, and report the anonymized-query coverage percentage next to it.
  2. Rank the four buckets by absolute click contribution and investigate only the largest one first. Most drops are dominated by a single term.
  3. For lost queries, pull the URLs that served them in the prior period and run them through the URL Inspection API before assuming a ranking cause.
  4. For a dominant impression effect, check whether affected queries crossed the page-one boundary before you call it demand.
  5. For a dominant residual CTR effect at stable position, go look at the live SERP for the top ten affected queries. No dataset you own will tell you what changed above your result.
Search ConsoleanalyticsBigQueryCTR

WriteMySEO produces marketing content, not legal, medical, financial, or compliance advice. Figures cited reflect publicly reported industry data at time of writing and shift over time.

Get started

We write this well about your industry, every month.

AI-drafted, human-reviewed SEO content on a flat subscription. Blog posts, metadata, schema, and internal links, shipped on a monthly rhythm.

See plans

More from the blog