A control page is an untouched page measured over the same dates as the page you changed, so that anything affecting every page shows up in it. Choose control pages by a written rule before you read their results, because choosing after you look lets the result steer the comparison.
- Penlid's SEO significance calculator accepts 3 to 8 comparison pages and refuses more than 8, because choosing the best 8 after seeing the result would choose the answer (Penlid, SEO significance calculator, read 30 October 2026).
- Qualifying comparison pages on the "before" days added 1.08 percentage points of false "improved" results in Penlid's worst simulated case, against −0.08 when a separate stretch was used (Penlid, methodology, read 30 October 2026).
- A comparison page needs at least 10 clicks across the two 28-day windows together (Penlid, SEO significance calculator, read 30 October 2026).
- A scientific control is an element of an experiment or observation designed to minimize the influence of variables other than the one under investigation (Wikipedia, Scientific control, read 30 October 2026).
What is a control page in SEO measurement?
A control page is a page you leave untouched while you change another and measure over the same calendar days, so that whatever moved every page shows up in the control. The factor under test is your edit; the other influences are a season, a holiday or a Google update, which hit many pages in the same days. Our SEO significance calculator page puts it this way: anything that moved every page (a season, an update, a holiday) moves them in the same days
.
Our tools say comparison pages or untouched pages, and our methodology's list of limits says control pages; the names mean the same thing. A control page is not the control arm of a visitor-split A/B test: our SEO test calculator notes that generic A/B calculators assume visitors are split at random between two versions, which a page edit is not. Setting your page's change against its controls' change is the idea behind difference-in-differences.
Control pages remove what hit your page and its controls alike, and nothing else. Google's Search Console help says it can be difficult to tell whether a page change directly caused an improvement, because other events also move performance. A control page cannot remove an event that hits only your page, so the answer is timing evidence, not proof. The guide to telling if a content update worked covers the whole procedure; this guide covers the choice that decides whether it can be trusted.
Why must you choose control pages before you look at their results?
You must choose control pages before you see how they did, because the set you pick defines what your page is compared with, and a set picked afterwards can be picked to give the answer you hoped for. Our calculator page puts the rule in one sentence: Write down which pages you will compare against before you look at any of their clicks after the change.
The next one gives the reason: If you pick the pages that make the answer look the way you hoped, the interval and p-value are for a question nobody asked.
The freedom is larger than it looks. From ten candidate pages there are 957 different sets of three to eight (120 + 210 + 252 + 210 + 120 + 45, the ways to choose 3, 4, 5, 6, 7 or 8 from ten; our own arithmetic), and the calculator page warns that a different set can give a different answer. The calculator also refuses more than eight pages, because choosing the best eight afterwards would choose the answer.
We tested a related risk on simulated sites. Our methodology reports that in the worst simulated case, qualifying pages on the "before" days added 1.08 percentage points of false "improved" results, while qualifying them on a separate stretch changed the rate by −0.08. Penlid's product decides which pages qualify from a stretch that is not part of the measurement, days −56 to −29, and it freezes a pool of up to 30 untouched pages at the moment it records the change. In our words, The pool is frozen because the one fact that cannot be rebuilt afterwards is which pages you left alone.
That stretch is not a perfect firewall. The size check reads the changed page's own clicks in the 28 days before the change, and among simulated runs at the size boundary that were given a verdict, false "declined" results outnumbered false "improved" ones. Ordering qualified pages by trend overlaps the measured days too, but it matters only when more than eight qualify, which none of our simulated scenarios has.
Which pages make good control pages?
Good control pages are untouched, similar to your page and large enough to carry a signal. Our calculator page words it as Pages you did not change in either window, that earn clicks from the same kind of query and are not tiny.
Penlid's rules put numbers on that.
| Rule | Calculator (typed clicks) | Penlid's product |
|---|---|---|
| How many | 3 to 8; more than 8 is refused | At least 3; the best 8 of a frozen pool of up to 30 |
| Smallest page | 10 clicks across the two windows | 5 clicks in days −56 to −29 |
| Size match | Not enforced; closest in size is one way to cut to 8 | Its clicks in days −56 to −29 within ten times, bigger or smaller, of the changed page's clicks in days −28 to −1, when those were 10 or more |
| Changed meanwhile | You leave them out | Pages changed from 56 days before to 28 days after are left out and counted |
Each rule has a reason:
- Untouched. A control changed inside its own qualifying stretch has a "before" that is that change's "after", so the product's exclusion window is wider than either measured window.
- Alike. A survey of the method on arXiv names its key assumption: without the change, your page and its controls would have followed parallel trends.
- Big enough. A page with a handful of clicks has a before-to-after ratio that chance alone moves a long way.
- Counted the same way. Search Console credits most data to a page's canonical URL, and its Pages dimension uses the final URL after redirects, so use one URL form for every page (our advice).
How many control pages should you use?
Use at least three control pages and at most eight, and as many of those as honestly qualify. Both limits are design choices, not laws of statistics, but more pages found a real lift more often in the one scenario we tested.
| Control pages | Real lift found |
|---|---|
| 6 | 61.23% |
| 8 | 70.02% |
Few control pages also widen the interval, because with three to eight untouched pages the amount of drift between pages is itself uncertain. With fewer than three usable pages the result is "not measurable" rather than a number.
How do you choose control pages, step by step?
Choose control pages in five steps, all finished before you open any candidate's clicks for the days after your change.
- Fix the dates. Call the day the change went live day 0. The calculator's windows are days −28 to −1 and 1 to 28, with days −56 to −29 as an optional earlier window; use windows that have fully ended.
- List and strike. From the Pages tab of the Performance report (Search results), list pages in the same section that get clicks for the same kind of query. Strike any edited from day −56 to day 28, counting a redirect, a merge or a canonical change as an edit (our advice), and ask whoever else works on the site, because the calculator cannot see an edit you do not tell it about.
- Apply the floor and the cap. Drop pages with fewer than 10 clicks across the two windows. If more than eight remain, rank them by something other than their results, such as closeness in size (the calculator's example), reading the size from the earlier window when you have it (our advice).
- Test the rule. A safe rule is one that a colleague who has never seen the after-change clicks could apply to get your list (our reasoning, not a Penlid figure).
- Write it down, then look. The free calculator trusts what you type, so the discipline is yours: save the list and the rule with the date in a change log or ticket. Only then pull the numbers and type identical windows for every page into the SEO significance calculator.
If the answer is "we cannot tell", do not swap in other controls and run it again, because that reopens the choice your written list closed; see what to do when an SEO test comes back inconclusive.
Example (illustrative numbers)
These counts are invented to show the arithmetic; they are not data from any site. A guide page was rewritten on day 0. It earned 940 clicks in the 28 days before and 1,175 in the 28 days after, so 1,175 ÷ 940 = 1.25, a rise of 25.0%. Seven pages in the same section were left alone; they are the candidates, listed from the closest in size to the farthest.
| Page | Before | After | Change |
|---|---|---|---|
| Changed page | 940 | 1,175 | +25.0% |
| Untouched page A | 905 | 1,122 | +24.0% |
| Untouched page B | 980 | 1,127 | +15.0% |
| Untouched page C | 850 | 1,116 | +31.3% |
| Untouched page D | 1,060 | 1,155 | +9.0% |
| Untouched page E | 790 | 837 | +5.9% |
| Untouched page F | 1,420 | 1,477 | +4.0% |
| Untouched page G | 410 | 439 | +7.1% |
Before looking, you write down a rule: use the five untouched pages closest in size, which picks A to E (size here is the before count, to keep the example short; the product reads an earlier stretch for this). Those five earned 905 + 980 + 850 + 1,060 + 790 = 4,585 clicks before and 1,122 + 1,127 + 1,116 + 1,155 + 837 = 5,357 after, so 5,357 ÷ 4,585 = 1.1684. The changed page did 1.25 ÷ 1.1684 = 1.0698 times as well, which is +7.0% versus what the untouched pages predict.
| How the set was chosen | Pages | Combined clicks, before and after | Changed page versus them | Calculator |
|---|---|---|---|---|
| Rule fixed before looking: five closest in size | A to E | 4,585 and 5,357 | 1.25 ÷ 1.1684 = 1.0698, so +7.0% | +7.1% (−18.2% to +40.3%), p = 0.52: too close to call |
| Chosen after looking: the three slowest growers | E, F, G | 2,620 and 2,753 | 1.25 ÷ 1.0508 = 1.1896, so +19.0% | +19.0% (+6.5% to +32.7%), p = 0.010: improved |
| Chosen after looking: the three fastest growers | A, B, C | 2,735 and 3,365 | 1.25 ÷ 1.2303 = 1.0160, so +1.6% | +1.5% (−27.1% to +41.1%), p = 0.87: too close to call |
The calculator works page by page on log ratios, so its figures differ a little from the simple division. The rule fixed before looking says too close to call, and the calculator adds that the smallest lift it could have detected is about +30.9%. The two sets chosen after looking disagree with each other. The three slowest growers happen to agree among themselves, so the calculator finds no drift between them and the interval narrows, which is how the same page comes out improved.
With seven candidates there are 99 sets of three or more pages (35 + 35 + 21 + 7 + 1). We ran every one through the calculator's estimator: 93 came back too close to call and 6 came back improved. Anyone who tried sets until one said improved would have found one here, which is the reason to fix the set first.
What this cannot tell you
- Control pages are matched, not randomized. Our methodology says it plainly:
Control pages are matched, not randomly assigned.
Random assignment places participants in groups by chance, so differences between the groups are not systematic at the outset; matching on earlier behavior removes a lot of confounding but not all of it, and a different set of untouched pages can give a different answer. - A written rule does not make the controls good. It stops the result from steering the choice; it does not make the pages comparable. The earlier-window check is a warning rather than a gate, and the arXiv survey cautions that passing a pre-trend test does not necessarily mean the trends were parallel.
- An edit nobody told you about breaks a control. We only know about changes we were told about, and the free calculator trusts what you type, so a control that someone else edited looks untouched.
- Controls cannot subtract a recovery that was coming anyway. Pages picked because they were declining, then left unchanged, came back "improved" 5.50% to 9.20% of the time in our simulations, against at most 3.10% otherwise, because the controls were not chosen for a bad four weeks. See regression to the mean and choosing pages to change without bias.
Questions people also ask
What is SEO testing?
SEO testing means changing something on a site and checking whether search clicks moved by more than they would have anyway. A page edit is not split between visitors at random, so the check sets the edited page against pages you left alone over identical dates. The result is timing evidence rather than proof, because something else can still hit only the page you changed.
How do you run an SEO test on a website?
Pick one page and one change, note the day it goes live, and write down the untouched pages you will compare it with before anyone sees their numbers. Once the 28 days after the change have passed, gather clicks for the 28 days before and the 28 days after, for every page, using identical dates, and put them through an estimator that returns an interval and a p-value. Make one change at a time: Penlid refuses to measure a page changed twice within 56 days.
Are control pages the same as on-page SEO?
No. On-page SEO is work done to a page itself, such as its text, headings and title. A control page is one you deliberately leave alone, kept as a yardstick so you can see what your work added. The same page can be a control in one test and edited in another, but not within the same measurement, because controls must stay untouched for the whole period.
Is there a free SEO testing tool?
Penlid's SEO significance calculator is free, needs no account and never sends what you type out of your browser. You enter clicks for the changed page and for three to eight pages you left alone, and it returns the lift with an interval, or says the numbers cannot tell. If you have a Search Console file instead of typed figures, the did-it-work tool reads it and picks comparison pages from your own data.
Sources
- Scientific control, Wikipedia, read 30 October 2026.
- Random assignment, Wikipedia, read 30 October 2026.
- What's Trending in Difference-in-Differences? A Synthesis of the Recent Econometrics Literature, arXiv, read 30 October 2026.
- Performance report (Search results): Common tasks and use cases, Google, Search Console Help, read 30 October 2026.
- Performance report (Search results): Dimensions and data groupings, Google, Search Console Help, read 30 October 2026.
- Methodology: how we measure, and what we can't, Penlid, read 30 October 2026.
The pool is frozen because the one fact that cannot be rebuilt afterwards is which pages you left alone.
Control pages are matched, not randomly assigned.
- SEO significance calculator, Penlid, read 30 October 2026.
anything that moved every page (a season, an update, a holiday) moves them in the same days
Write down which pages you will compare against before you look at any of their clicks after the change.
If you pick the pages that make the answer look the way you hoped, the interval and p-value are for a question nobody asked.
Pages you did not change in either window, that earn clicks from the same kind of query and are not tiny.
- SEO test calculator, Penlid, read 30 October 2026.