Is the CTR-by-Position Curve Real? We Tested 1,000 Rankings

You have seen the curve. Position 1 takes the lion’s share of clicks, position 2 roughly half of that, and the line slides down a smooth ramp until everything past page 1 rounds to zero. It is in every pitch deck, every “why you need to rank higher” slide, every forecast that promises “if we move you from position 8 to position 4 you’ll get X% more clicks.” The CTR-by-position curve is the most-cited chart in SEO. So we did the obvious thing and checked whether it describes a real site.

The short answer: at the very top it roughly holds, and almost everywhere else it falls apart. On ~1,000 ranking queries from one of our own sites, the top three positions tracked the canonical curve well enough. Below that it went non-monotonic inside page 1, and then it did something the curve says is impossible — positions 11–20 out-clicked positions 4–10, with a median organic CTR of 15.6% versus 8.7%. The published curve is real the way the average human has 1.97 legs is real: a population statistic that describes almost no individual.

This is for the founder being shown a CTR forecast and the SEO building one. It is a companion to our Keyword Difficulty validity study — same house style, real data, stated method — and an applied case of measuring SEO like a psychometrician: before you trust a number, ask what it actually measures.

Method (reproducible)

• Data: first-party Google Search Console CTR-by-position, pulled via Ahrefs’ GSC integration, for one site we run. ~1,000 ranking queries with click data, over roughly the last three months. Snapshot 2026-06-28.
• What “CTR-by-position” means here: for each integer position, the average click-through rate across every query that ranked there in the window, weighted by impressions. Standard GSC math — the same math the public curves are built from, just on one property instead of millions.
• Per-position sample size (n): the number of distinct ranking queries behind each position’s CTR. It ranges from 2 to 80. Hold that number — the thinness is the finding, not a defect to apologize for.
• Scope and limits, up front: one site, one ~3-month window, one tool’s averaging on top of GSC’s own. Deep-position CTRs rest on tiny samples (n = 2–6 in places) and are therefore high-variance. CTR is not a clean function of position anyway — it also depends on query intent, brand, and SERP features. We are not claiming this site’s curve is the curve. We are claiming the opposite: that no single site has a stable curve to forecast from, and ours is the proof of concept.

We are not naming the property. Everything below is aggregate CTR-by-position — no keywords, no traffic counts.

Finding 1 — At the top, the curve roughly holds

Give the instrument its due. The top three positions on our site behaved about how the textbook says:

A clean monotonic step down, 1 > 2 > 3, with a big gap between first and the rest — qualitatively the canonical shape. The absolute numbers run hotter than the famous public curves (position 1 at ~49%, well above the share those population curves usually report), which is exactly what you would expect from a single niche site whose top rankings skew branded. The point stands: at the very top, position dominates, and the curve is a usable rough sketch. If the story ended at position 3, the curve would be fine.

It does not end at position 3.

Finding 2 — Inside page 1, it stops being monotonic

The entire forecasting value of the curve rests on one assumption: every step down costs you clicks, monotonically. Watch it break above the fold of page 1:

Position 7 (8.8%) edges out position 6 (8.7%). Position 9 collapses to 3.8% — below position 10 at 7.1%, below position 8 at 6.0%. The “every position down loses CTR” rule is already violated on page 1, where samples are largest (n up to 80) and the curve is supposed to be most reliable. If position 9 under-clicks position 10 on your own data, what exactly is a “move you from 8 to 4” projection forecasting?

Finding 3 — The smoking gun: positions 11–20 out-click 4–10

Here is the result that should retire the forecast. The canonical curve says page 2 is a CTR graveyard — low single digits, then zero. On our site, page 2 out-performed the bottom of page 1.

Compute the medians and the inversion is unambiguous:

• Median CTR, positions 4–10: 8.7%.
• Median CTR, positions 11–20: 15.6%.

Positions 11 through 20 clicked through at nearly twice the rate of positions 4 through 10. Under the monotonic curve this cannot happen — a result on page 2 is supposed to be strictly worse than a result on page 1. On a real property it is routine. (The deep tail also contains many positions sitting at flat 0% with n = 2–6; we are not hiding those, they are the other face of the same small-sample coin.)

Finding 4 — Why your curve inverts

This is not magic, and it is not a measurement bug. It is what the smooth public curve is built to hide.

The published CTR curve is an average over millions of keywords. At that scale, query mix washes out: every position contains a representative blend of branded and unbranded, navigational and informational, high-intent and idle. What is left is the pure effect of position, with error bars so tight they vanish. That is a genuine, useful population fact.

Your site is not a population. The deep-ranking queries that still get clicks are a selected set: branded, navigational, or niche-intent terms where the searcher already wants you specifically and will scroll to page 2, page 3, position 83 to find you. Generic queries that rank deep get no clicks and quietly drop out of the data. So each deep position’s CTR is the average of a handful of unusually determined searches — n of 2 to 80, skewed by whatever queries happened to land there. Position 83 reads 50% CTR because two people searched something only you answer and both clicked. That is not a ranking triumph. It is a sample of size two.

The curve did not lie about the population. You applied a population statistic to an individual and got noise wearing the shape of structure.

The measurement lesson

A population average can have tiny error bars because it pools millions of observations — the noise cancels and what survives is signal. Take that same curve to your ~1,000 keywords, spread across positions at roughly tens-per-bucket, and the error bars swallow everything past the top three. You are not reading a curve. You are reading sampling variance and calling it a law.

This is the measurement-error story in one chart. Classical test theory says every observed number is a true value plus error, X = T + E. The public CTR curve is nearly all T because millions of observations drown out E. Your per-position CTR, at n = 2–80, is mostly E. Same metric, opposite reliability — and reliability does not transfer when you change the sample.

So when a deck promises “moving from position 8 to position 4 gets you +X% clicks,” look at what it is doing: it is reading a difference off a population curve and projecting it onto your handful of position-8 queries, whose own CTR is dominated by which specific queries they are and how many people searched them — not by the position. That projection is measurement error dressed as a forecast. It will be wrong in a direction nobody can predict, because the thing it is built on does not exist at your scale.

None of this means position is irrelevant. At the very top, on enough queries, it clearly drives clicks. It means the curve as a forecasting tool — a fixed CTR you can multiply by projected impressions — is a population artifact that does not survive contact with one site’s data.

What to do instead

You do not need the generic curve. You have something better and you are sitting on it.

1. Build your own CTR baseline from your own GSC. It is the only curve that describes your queries, your brand pull, your SERP features. Treat the public curve as folklore, not a forecast.
2. Segment before you average. Split branded from unbranded, informational from transactional, queries that trigger SERP features from clean ten-blue-links. A single blended CTR-by-position curve is a fiction stitched from incompatible query types — the same sin as GSC’s blended “average position.”
3. Treat any single-position CTR as an estimate with error, and read the n. A position backed by 3 queries tells you nothing; a position backed by 60 means something. Put the sample size next to the rate, always, and refuse to forecast off a cell whose n you would not bet on.
4. Stop forecasting clicks from position deltas. “Rank higher → this exact CTR → this much traffic” chains three shaky estimates into a confident number. Track search engine visibility and actual clicks for the terms that matter, with rolling windows, instead of multiplying yourself a projection off a curve that is not yours.

A curve is a tool, not a prophecy. Read what it measures, on whose data, at what sample size — not the smooth line someone drew through a million other sites.

Frequently Asked Questions

Is the CTR-by-position curve real?

As a population average, yes — pooled across millions of keywords it shows a smooth, reliable decline because the noise cancels out. As a description of any one site it is mostly fiction. On our ~1,000 real rankings the curve held only for the top three positions, then went non-monotonic and even inverted, because a single property has too few queries per position for the average to be stable.

Why did page-2 rankings get a higher CTR than page-1 rankings on your site?

Because the deep-ranking queries that still earn clicks are a selected set — branded, navigational, or niche terms where the searcher wants you specifically and scrolls to find you. Those few determined clicks, averaged over tiny samples (n = 2–80 per position), produce CTRs that beat mid-page-1 positions. Query mix and sample size, not position, drive the organic CTR curve down there.

Can I use the organic CTR curve to forecast traffic from a ranking improvement?

Not reliably. “Move from position 8 to 4 for +X% clicks” multiplies a population-curve delta against a handful of your own queries whose CTR is dominated by their intent and sample size, not their position. That is measurement error dressed as a projection. Forecast from your own segmented GSC data, with error bars and sample sizes attached, instead.

What sample size do I need before a per-position CTR means anything?

There is no magic threshold, but read the n on every position before you trust its CTR. In our data, positions backed by 60–80 queries were stable enough to reason about; positions backed by 2–6 were pure variance — one position read 50% CTR on a sample of two. Always display the query count beside the rate, and discount thin cells hard.

Should I ignore my own CTR data because it is noisy?

No — yours is the only CTR data that describes your site. The fix is not to discard it but to handle it correctly: segment by intent and brand, attach sample sizes, use rolling windows, and treat each figure as an estimate with error rather than a fact. A noisy instrument you understand beats a smooth curve built from someone else’s million keywords.

Method note: first-party Google Search Console CTR-by-position via Ahrefs, one site we operate, ~1,000 ranking queries over a ~3-month window, per-position n = 2–80, snapshot 2026-06-28. Figures are impression-weighted average CTR per integer position; medians computed across the stated position ranges. If you want this kind of measurement rigor — your own CTR baseline instead of a borrowed curve — applied to your SEO, our fractional SEO practice is built around exactly that.