Designing a Link Building Experiment: Control Page

Most teams judge a link building tactic by feel: "it seemed to help last time". If nobody wrote down which page received the links, when they went live and what else changed that month, there is nothing left to examine later. This article turns designing a link building experiment into a small, defensible routine: one question, a test page and a control page, a baseline reading, a recorded intervention, an observation window and rules for reading the result. The method is a way to learn. It does not promise that adding links will change rankings or traffic. Akua Adforce does not provide experiments, A/B testing or measurement services, so you would run the experiment with your own tools and your own spreadsheet.

Designing a Link Building Experiment: How to Test a Strategy Assumption

An assumption is not a strategy

Every strategy rests on an assumption, such as "if guides on our category page earn links from topical publications, that page becomes more visible". An assumption can be written in one sentence and can turn out wrong. An experiment is the smallest honest way to check it. Planning and testing sit on different layers: a document like the 90-day link building roadmap for new websites lays out what to do and when, while an experiment asks whether one step inside that roadmap is worth repeating.

When an experiment is not worth running

Not every site has enough pages, or enough traffic, to separate signal from noise. On a site with a few hundred monthly visits, ordinary week-to-week movement can be larger than anything three links might cause. Designing the experiment is still useful because it shows how much uncertainty you face, but what you can say afterwards is closer to "I have an impression" than "I have an answer".

The policy boundary

An experiment is not a licence to manipulate rankings. Google's spam policies describe link spam as links created mainly to manipulate ranking. If a paid placement is part of your test, it should carry a sponsored or nofollow label; paid links labelled that way do not breach those policies.

Writing the Experiment Question: One Variable, One Question

What a usable question looks like

"Does link building work?" is too wide to test. Compare: "If three descriptive-anchor links from topical publications point to our category guide, will that page's impressions differ from a comparable guide's over six weeks?" It is long, but it names a target page, an intervention, a period and a comparison. You can answer it with a table.

Keeping to a single change

Adding links, rewriting the title and appending a new section in the same month is not a one-variable test. You will never know which change did what. The rule is simple: on the test page, change only the links. If a fix on the page is truly necessary, either delay the experiment or make the same fix on the control page on the same day.

Fixing the output before you look

Pick the number that will answer the question in advance: impressions, clicks, average position or sessions to the target. You may watch several, but one is the primary output. Declaring another metric primary after the fact means shaping the table to look good. For background on timing expectations, read about when backlink effects tend to appear, though no duration is ever certain.

Choosing a Test Page and a Control Page

Finding a comparable twin

A control page receives no intervention but looks as much like the test page as possible: the same type (two category guides or two product pages), similar age, similar internal link count and similar seasonality. Imagine a hypothetical home textiles site: the "duvet cover guide" becomes the test page and the "bed linen guide" the control, because both serve the same audience and rise and fall in similar seasons.

Is the test page ready?

A test page must be ready to receive links. Adding links to a page whose purpose, indexability or content is unclear damages the experiment itself. A checklist for preparing the target page is in this guide to getting a landing page ready for links. The control page should be at the same level of readiness, otherwise any difference comes from preparation instead of links.

More than one control

If possible, use two or three control pages rather than one, so a freak movement on a single page does not steer the result. More controls do not make the answer certain; they only reduce dependence on one page's luck.

Recording the Baseline Reading

Which date range to use

The period before the intervention should be as long as the observation window. If you plan six weeks of observation, take the six weeks before the first link as the baseline, for both test and control pages. Exporting Search Console performance data filtered by page is enough; the article on the Search Console performance report walks through the steps.

Logging the links that already exist

Note the existing external links of both pages before you start. The Search Console links report can be a source, but it is a sample rather than an exhaustive list, its tables are capped at 1,000 rows, it can include old links, and it does not tell you whether a link is nofollow. Record what it shows as "the sample the report displayed", not as the exact link count. Counts from third-party tools are tool-specific numbers, not figures Google works from.

Writing the data limits under the table

Add known limitations beneath the baseline: Search Console data can be incomplete because of sampling and privacy filtering, and average position blends many queries even for a single page. Those notes protect you from overconfidence when you read the result later.

Defining the Intervention: When, How Many Links, Which Label

Pinning the intervention down

The intervention is the part of an experiment that drifts most easily, because execution rarely matches the plan. Record it first: how many links, to which page, with which anchor style, going live between which dates. A hypothetical plan might read: "3 links, to the test page, brand-name and descriptive anchors, 1 to 15 October." Keep a second column for what actually happened, since planned and actual often differ.

Link attributes and paid placements

If any placement is bought, it should carry rel="sponsored" or nofollow. That attribute can influence how the result reads, because how Google treats qualified links is outside your control. Keep an attribute column for each link and consider it when you interpret the outcome.

When the publisher decides the timing

With purchased placements, the publication date, the anchor wording and the exact position of the link often rest with the publisher. Your plan should reflect that dependency: ordering a placement does not mean it appears on the planned day. Do not start the observation clock until the delivery notice confirms the link is live.

Observation Window and Interim Checks

Writing the duration down first

Extending or cutting the window after seeing early numbers is among the most common mistakes. Fix the duration beforehand; in a hypothetical experiment that could be six weeks. How long is reasonable depends on your site's volatility and how often the pages are crawled, and no single duration is right. A site with large weekly swings may need a longer window than a calm one, and looking at how much the page normally moves in a week is a practical way to judge whether your window is enough.

Checks that collect, not decide

Weekly or fortnightly interim checks exist to collect data, not to decide. Abandoning an experiment at week two because "nothing is visible" is a search for an early verdict. At each interim check, ask only two things: was the intervention really carried out, and did anything unexpected happen, such as an outage or a removed publication? A third question is worth adding: is the page that links to you still reachable? A removed placement can invalidate the whole run, and an interim check is when you catch it.

When to stop early

If a serious technical error appears on the test page, a link source is taken down, or the control page changes, stop the experiment, record why, and label it "incomplete". Deleting a stopped experiment silently invites the same mistake later.

Logging Other Changes in the Same Period

The change log

Keep a separate log beside the experiment: site-wide content releases, design updates, technical fixes, campaigns, holidays and seasonal effects. The log does not remove confounding; it makes it visible. Without it, attributing a month of movement to a link becomes far too easy. Entries can be short: date, what happened, which pages it touched, and a second person who can read it so forgotten events get noticed.

Algorithm updates

If a Google update lands inside the window, both test and control pages may be affected, though not necessarily in the same way. How to read traffic movement after an update is described in diagnosing traffic loss after a Google update. Log the date of the update and add a line in the conclusion saying the window may have been influenced by it.

Changes you did not make

You cannot write the actions of competitors, publishers or search result layouts, but any change you notice goes into the log with its date. "A competitor published a new page" or "a new box appeared in the results" keeps an alternative explanation alive, and each one is a reminder that the observed difference may have another cause.

Reading the Result: A Causality Caveat

What the gap between test and control says

Suppose that after six weeks the test page's impressions grew more than the control's. That gap is a coincidence in time, not proof that the links caused it. A small sample, a single control page, a shifting query mix and seasonal differences can all produce the same picture. In a hypothetical case, most of the apparent gain might come from one seasonal query, which you would only notice by looking query by query. Write "a difference was observed", not "the links worked".

What no difference means

Seeing no difference does not prove the links were useless either. The window may have been short, the links may not have been crawled, or the change may have been smaller than the noise. The value of the experiment is that your next decision rests on better information, not that it hands down a verdict. Words like "proved" and "showed" are better left out of the write-up; "observed", "not observed" and "could not be distinguished" are more honest.

One experiment is not a decision

A single result is not enough to justify a strategy change on its own. Even repeated results in the same direction across different pages only point to a pattern. If you also want to relate links to cost, how to measure backlink ROI discusses measurement limits that matter there too.

What to Do When the Result Is Unclear

Writing the uncertainty plainly

Dressing an unclear result up as a "positive trend" creates false confidence. Say it directly: "no distinguishable difference appeared between test and control in the window; the duration and scale may have been too small to answer this question." That sentence is a finding, not a failure. Attach the next step to it: which piece of information was missing, and how will the next attempt supply it? Even an unclear result then leaves a useful trace in the team's memory.

Options for trying again

There are three paths: repeat the same experiment over a longer window, simplify the question (one link instead of three, for instance), or accept that this tactic cannot be measured cleanly on your site. Write down which path you choose and why. Simplifying is often the most useful, because a plainer intervention is easier to interpret.

A stopping rule

Set a "this far and no further" point for each tactic: if two repetitions show no clear pattern, the budget set aside for it moves to another job. Without a rule written in advance, an unclear experiment can run forever. Stopping does not mean the tactic does not work; it means that on this site, at this scale, with this measurement set-up, nothing could be learned. Say so explicitly in the note.

Tying Experiment Notes Back to the Strategy Decision

What an experiment record contains

A record holds the question, the target and control pages, the measurement window, the intervention, the change log, the result sentence and a next step. The only test of a good record is whether someone can understand what was done a year later. In practice that means no abbreviations, full page names and dates written in an unambiguous format.

From note to decision

An experiment result does not issue an order to scale up or drop a tactic. Write the decision note more carefully: "No distinguishable difference in this trial; we will still use this part of the budget to test a different tactic." For a view of how to position strategy around major updates, how Google core updates impact link building strategy provides a perspective.

Sharing the records

Collecting experiment records in one shared table stops the team from asking the same question twice. Add the line "this result is not conclusive" once beside the records, so later readers quote them with caution.

Example Template: A Link Building Experiment Plan (Hypothetical)

Everything below is invented for illustration (hypothetical): page names, dates and numbers do not belong to any real site.

Fill the plan in this order:

1. Write the question in one sentence and mark the single variable.
2. Choose the test and control pages and note their readiness.
3. Export the baseline on the same day for every page.
4. Log the intervention in planned and actual columns.
5. When the window ends, write the result as observed, not observed or unclear.

• Any tool figure you include should be labelled as tool-specific and not a Google figure.
• Share the plan before observation begins; a question rewritten afterwards weakens the record.

Frequently Asked Questions

How do you design a link building experiment?

Start with one question and one variable, then choose a test page and a comparable control page. Record the period before the intervention as the baseline, define the intervention by date, link count and attribute, fix the observation window in advance, and log other changes during the same period.

Why is a control page needed?

A control page is similar to the test page but receives no intervention. It shows the general movement of the site or search results during the same weeks. Without it you cannot tell whether a change on the test page came from the links or from a wider shift, and even with a control page causation is not proved.

How long should the experiment run?

There is no single right length. It depends on traffic patterns, how often the page is crawled and the output you measure. What matters is writing the duration down before the experiment and not stretching or trimming it after seeing numbers. Hypothetical examples often use six weeks.

If results change, is the link the cause?

No. A change on the test page in the same period does not show that the links caused it. Seasonality, content updates, algorithm changes and small samples can all explain it. Record the outcome as a difference observed or not observed, never as proof.

Can several tactics be tested at once?

It is not advised. Applying several tactics to one page makes it impossible to tell which one made a difference. Testing separate tactics on separate page pairs, or in separate time windows, produces a record that is easier to read.

What if the experiment result is unclear?

Record the uncertainty as it is. Write down which option you choose: repeat over a longer window, simplify the question, or drop the tactic as untestable on your site, and why. A stopping rule set beforehand keeps the experiment from running indefinitely.

Conclusion and Next Step

Keep the first experiment small: one question, a test page with two control pages, a six-week window. Whatever the outcome, the routine leaves a record, and the record lets you say "in this window we observed this" at the next budget meeting instead of "I think".

If the intervention involves purchased links, you can look through the listings on the backlink packages page. The purchase path runs from picking a listing and paying, through the seller accepting the order and placing the link, to a delivery notice; if what arrives differs from the order, a dispute can be opened. Timing and the final look of the placement depend on the seller, and no result is promised.