Last Updated on November 13, 2026 by Statnzee Team
Ask “if X changes a little, how much does Y change?” and get a useful answer, using real numbers from a public analytics report.
Start with a roller coaster
A rider starts at rest at height h. All the potential energy turns into kinetic energy, so the energy balance and the resulting velocity are:
Now raise the starting height by 2 percent. How much faster is the rider at the bottom? Because v is proportional to h raised to the power 0.5, a small percentage change in h produces about half that percentage change in v:
That is about 1 percent. The exact value is:
So the shortcut is off by less than 0.01 percentage points. This is linear approximation: near a point, a curve behaves like a straight line, so a square root can be replaced by simple multiplication.
The same idea works in reverse. If you want 1 percent more velocity, you need about 2 percent more height. And it is the same trick that lets you reason about a dashboard without a spreadsheet.
The one rule to remember
If Y is proportional to X raised to the power k, then for small changes, the forward rule is:
and the inverse rule is:
The shortcut is reliable for changes up to roughly 20 percent. Beyond that, check the exact value. A 50 percent increase in height gives exactly:
while the shortcut says 0.25. Analytics is full of relationships with this shape: products (k = 1), ratios (k = −1) and square roots (k = 0.5 or −0.5). The rest of this article shows where each one appears.
About the data used here
The examples come from the public Cumulative Analytics Overview on dsplendid.com, which I read on November 13, 2026. The detailed report on that page covered calnzee.com for Aug 15 to Nov 12, 2026 (Google Analytics 4 and Search Console).
The portfolio-wide tables on the same page cover 30 days. Numbers move every day, so treat them as a worked example. The method is what matters.
Relationship 1: Metrics that multiply
Many metrics are two other metrics multiplied together. When that is true, small percentage changes simply add. The most familiar example is clicks in search:
The Calnzee report shows 17,500 impressions and 84 clicks, so the click-through rate (CTR) is:
That is 0.48 percent. For a product, the percentage changes add:
- Forward: impressions up 10 percent and CTR down 5 percent gives clicks of about +5 percent (exactly +4.5 percent).
- Inverse: to reach 100 clicks, you need this much growth:
That is about +19 percent. You can get it with about 20,800 impressions at the same CTR, or a CTR near 0.57 percent at the same impressions, or roughly +9 percent on both.
The same structure shows up elsewhere: sessions equal users times sessions per user, and total engaged time equals sessions times average time per session.
A realistic complication. The two factors are rarely independent. The top queries in the report hint at why. “pythagorean theorem” earned 117 impressions and 1 click (about 0.9 percent). “best calculus textbook” earned 45 impressions and 1 click (about 2.2 percent). “best calculus books” earned 21 impressions and 1 click (about 4.8 percent). Broad, informational queries inflate impressions but dilute CTR, while queries with buying intent turn impressions into clicks at a higher rate.
Each of those is a single click, so this is a hypothesis to test, not a conclusion. But it suggests a useful habit: measure your effective exponent. If you grow impressions by 30 percent through broad content and clicks grow only 10 percent, your effective k is about 0.33, and more broad content has sharply diminishing returns compared with targeted pages.
Relationship 2: Why more visitors can mean less time on site
Average time per session is a weighted blend of groups of visitors:
So when traffic grows, the average can fall without anyone reading less. The new visitors are simply a different group.
A clean example. Take 100 sessions averaging 150 seconds. Add 10 percent more sessions (10 visits) that last about 4 seconds each. The new average is:
That is about 137 seconds, which is 8.9 percent lower. A handy rule of thumb: every 1 percent of sessions that are near-instant pulls the average down by about 1 percent.
The inverse question. If your average falls 10 percent after a traffic surge, roughly 10 percent of your sessions are probably near-zero engagement. That is a diagnosis, not just a number.
Real numbers from the report. The homepage has 1,060 sessions averaging 2m 32s. Several articles show averages of 2 to 14 seconds. A “Page not found” row has just 8 sessions but an average of 61m 12s, which is most likely a tab left open rather than real reading. Using the ten rows in the top-content table (1,596 sessions in total), the blended average works out to about 2m 28s. Remove those 8 sessions and it drops to about 2m 11s. In other words, 0.5 percent of sessions moved the average by roughly 13 percent. This is an approximation, because the rows overlap and cover only the top ten pages, but the lesson holds: check for outliers before trusting an average.
Traffic quality matters even more. The report shows 2,364 visitors, up 1,228 percent on the previous period, with Singapore (38.9 percent) and China (26.4 percent) at the top of the location list. On the portfolio overview, the site’s own “verified human estimate” tells a different story: for the portfolio over 30 days, the United States (106 users) and India (95 users) lead the verified list, while Singapore and China dominate the “not counted” list. The page itself notes these are estimates, not verified bot detection. The takeaway for any site: before you ask why average time fell, segment by channel, country and device, using the filters most analytics tools provide.
Relationship 3: The square-root law of trust
This is the closest cousin of the roller coaster, because it has the same 0.5 exponent. The random error in an average shrinks with the number of sessions n like this:
- Forward: 10 percent more sessions reduces your error by about 5 percent, because the factor is:
- Inverse: to cut your error by 10 percent, you need this many times the sessions:
That is about 23 percent more sessions (the linear shortcut says 20 percent, which is close). To halve your error, you need four times the sessions.
A practical use. The article “Why Does x² + y² = 1 Give a Circle?” has 33 sessions. To halve the uncertainty in its numbers you would need:
That is about 132 sessions. Until then, small differences between pages are mostly noise.
Try it on engagement rates. For a percentage p measured on n sessions, the standard error is:
The ellipse article shows a 22.22 percent engagement rate on 36 sessions, which is 8 engaged sessions. Its standard error is:
That is about 6.9 points, so its true rate could plausibly sit anywhere from roughly 9 percent to 36 percent. The circle article shows 6.06 percent on 33 sessions, which is 2 engaged sessions. The 16-point gap between the two articles is only about two standard errors, which is suggestive but not conclusive. Collect more sessions before rewriting anything.
Relationship 4: Do more pages mean more search impressions?
Sometimes you have to estimate the exponent from data. The portfolio page lists three sites with both search-visible pages and Google impressions over 30 days:
- Calnzee: 161 pages, 1,019 impressions
- Flipnzee: 41 pages, 216 impressions
- Dsplendid: 10 pages, 28 impressions
Fitting a power curve (a straight line on log-log axes) gives an exponent of about 1.3:
If that held, then 10 percent more pages goes with about 13 percent more impressions:
and for the inverse, 20 percent more impressions would need about 15 percent more pages:
An exponent above 1 is plausible: each new page can rank for several queries and also lifts the pages around it. Calnzee has about 2.5 ranking keywords per page (401 across 161 pages), compared with about 0.5 for Flipnzee (19 across 41) and 0.6 for Dsplendid (6 across 10).
Be honest about the limits. Three sites of different ages and topics is a thin base. Flipnzee is the oldest of the three yet has fewer impressions than its page count alone would predict, so age, topic and content quality clearly matter too. Treat 1.3 as a hypothesis you test on your own site by recording pages and impressions every month.
Relationship 5: The conversion chain
A chain of steps multiplies in the same way:
The Calnzee report shows 0 key events in this period, so there is no honest conversion estimate to make yet. But the structure is worth setting up in advance. If engagement rate and conversion rate each rise 10 percent, events rise about 21 percent:
To double your events with traffic held flat, you need about 41 percent more on both levers:
or a bigger push on one. This tells you which lever is cheaper to move.
Cheat sheet
| Relationship | Exponent k | Forward | Inverse |
|---|---|---|---|
| Clicks vs impressions (CTR fixed) | 1 | +10% impressions gives +10% clicks | +10% clicks needs +10% impressions |
| Clicks vs CTR (impressions fixed) | 1 | +10% CTR gives +10% clicks | +10% clicks needs +10% CTR |
| Error of an average vs sessions | −0.5 | +10% sessions cuts error by about 5% | −10% error needs about +23% sessions |
| Impressions vs search-visible pages (three-site estimate) | about 1.3 | +10% pages gives about +13% impressions | +20% impressions needs about +15% pages |
| Events vs engagement rate and conversion rate | 1 each | +10% on both gives +21% events | Doubling events needs about +41% on both |
How to find your own exponent
You do not have to guess k. Take two snapshots of your data and compute it:
For small changes, simply divide the percentage change in Y by the percentage change in X. As a hypothetical example, if pages grow from 100 to 110 and impressions grow from 1,000 to 1,150, then:
So k is about 1.5. Once you have k, you can run the forward question (“what happens if I add 10 pages?”) and the inverse question (“how many pages for 30 percent more impressions?”).
Pitfalls to avoid
- Small samples. One click or 33 sessions cannot support a strong conclusion. Check the square-root law first.
- Mix changes. If the audience changes, an average can move with no change in behavior. Segment before you interpret.
- Big jumps. Linear approximation works for small changes. For anything above about 20 percent, calculate the exact value.
- Correlation is not causation. A fitted exponent describes a pattern in past data. It does not prove that adding pages causes impressions.
- Dependent factors. In a product like clicks = impressions × CTR, pushing one factor can move the other. Measure the effective exponent.
The takeaway
The roller coaster teaches a habit of mind: before reaching for a complex model, ask what shape the relationship has. If it is a product, percentages add. If it is a square root, percentage changes are halved. If it is an average, ask who is in the mix. With a few real numbers and one forward-and-inverse rule, you can answer both “what happens if I grow this?” and “how much do I need to grow to hit my goal?” in a few seconds.
Discover more from Statnzee
Subscribe to get the latest posts sent to your email.

Leave a Reply