A/B testing an affiliate site: the setup in seven steps
An A/B test shows half your readers the current page (version A) and half a changed copy (version B), then compares what each group did. On an affiliate site you're paid when a reader clicks your link and buys from the merchant, so the result that counts is commission, not clicks.
| Step | What to do | Why |
|---|---|---|
| 1. Pick the page | Your busiest "best X" list, review or comparison page | A quiet page never reaches an answer |
| 2. Write one change and the reason | "Putting the top pick above the intro will lift clicks, because phone readers never scroll to it" | One change tells you what caused the result |
| 3. Check the visitors you need | Use the table below before you start | An undersized test wastes weeks |
| 4. Give each version its own tracking ID | A second Amazon tracking ID or a network sub-ID on version B's links | Your site sees clicks; only the network sees sales |
| 5. Split readers with a testing tool | Each reader keeps seeing the same version | A fair split |
| 6. Set the end date first | Whole weeks, never stopped early | Early peeks pick false winners |
| 7. Decide by commission per visitor | Then publish the winner and remove the test | Google asks for prompt cleanup |
Three things change how you apply this:
- Traffic. Many pages are too small to test; see the visitor table below.
- Clicks vs commission. A version can win clicks and lose money by steering readers to a cheaper product.
- Rules. Disclosure law, Amazon's policies and Google's testing rules bind every version.
What should I A/B test on an affiliate site?
Start with what sits between the reader and the first click. Bigger changes first; button colors last.
| What to test | Version A vs version B | Watch for |
|---|---|---|
| Top pick box | Below the intro vs right under the headline | More clicks, same commission per click |
| Links and buttons | Text links vs buttons; "Check price on Amazon" vs "See it on Amazon" | Name the store: Amazon bans links that hide it |
| Comparison table | Three products vs six; or price column vs no price column | Readers who scan vs readers who read |
| Number of sellers | One merchant vs the same product at two or three stores | Clicks spread out; check total commission |
| Product order | Best overall first vs best value first | Which product gets the clicks, and what it pays |
| Proof near the pick | Your own test photos beside the pick vs lower in the review | Trust before the click |
| Phone layout | No sticky bar vs a slim bar with the top pick | Clicks that are accidental taps |
For review pages, Google suggests "links to multiple sellers to give the reader the option to purchase from their merchant of choice" and first-hand evidence such as your own photos (Google Search Central: Write high quality reviews).
Button tests rarely touch rankings. Google says changes to a button's size, color or placement, or to the call-to-action wording, "often have little or no impact on that page's search result snippet or ranking" (Google Search Central: Minimize A/B testing impact). Rewriting the review itself is different: that's the content Google ranks, so test it carefully and watch Google's free tool that shows which searches bring people to your site and flags pages Google can't index. More.
What should I never A/B test?
Some versions break the law or program rules, whatever they do for clicks.
| Don't test | The rule |
|---|---|
| Moving the affiliate disclosure away from the links | FTC: "The closer the disclosure is to your recommendation, the better." |
| A "Buy now" button as the only disclosure | FTC: a "buy now" button "would not be adequate"; "Paid link" right next to the link can be enough |
| Letting the affiliate link speak for itself (Canada) | Competition Bureau: commissions are a connection to disclose, and "linking to an affiliate webpage is unlikely to be enough" |
| Removing Amazon's statement | Amazon requires "As an Amazon Associate I earn from qualifying purchases." on your site (US and Canada agreements) |
| Typed-in Amazon prices | Amazon allows prices only in links it serves or from its API, with a date/time stamp if refreshed less often than hourly |
| Short links or buttons that hide the store | Amazon bans any link "in a manner that makes it unclear that you are linking to an Amazon Site" |
| Made-up stock counts, buyer counts or past sale prices | Competition Bureau: "Be accurate when you make claims about the number of items left or about demand"; Amazon bans "inaccurate, overbroad, deceptive or otherwise misleading claims" about products or prices |
| Countdown timers that reset | FTC's 2022 dark patterns report: a "countdown clock that just goes away or resets when it times out" |
| A different page for Googlebot | Cloaking breaks Google's spam policies, test or not |
The FTC's own example of a disclosure that works:
"I get commissions for purchases made through links in this post." (FTC: FTC's Endorsement Guides, What People Are Asking)
On deadlines, Canada's Competition Bureau tells businesses:
"Ensure that countdown timers do not run out or automatically reset while the sale price remains in place." (Competition Bureau Canada: Fake urgency cues)
So an "evergreen" timer that restarts for every new reader is out. A real deal with a real end date, shown truthfully, is fine.
Keep affiliate links marked rel="sponsored" or rel="nofollow" in both versions. Google's spam policies accept affiliate links "with links treated appropriately", meaning paid links carry one of those two values (Google Search Central: Spam policies).
How many visitors does an affiliate A/B test need?
Usually thousands per version. PostHog's documentation gives the standard rule of thumb: visitors per version = 16 × variance ÷ (the change you want to detect)², at 95% confidence and 80% power (PostHog Docs: Running time and sample size). For a click rate, variance is the rate × (1 − the rate).
Your click rate is the share of readers who click any affiliate link. Visitors needed per version:
| Click rate today | To detect a 20% lift (e.g. 10% to 12%) | To detect a 10% lift (e.g. 10% to 11%) |
|---|---|---|
| 5% | 7,600 | 30,400 |
| 10% | 3,600 | 14,400 |
| 20% | 1,600 | 6,400 |
Double it for the whole test. Sales are rarer than clicks, so commission takes even more visitors to read: call the test on clicks, then check that commission per visitor didn't drop.
Decide the length before you start and don't stop the moment one version leads. PostHog names that mistake the peeking problem:
"This is when you check the intermediate results for statistical significance, make decisions based on them, and end your experiment too early." (PostHog Docs: Experiments and A/B testing best practices)
Run whole weeks so weekday and weekend readers both count, and avoid big sales such as Black Friday. A win among phone readers only is an idea for a new test, not a result.
Too little traffic? Test a template, not a page: the same change on every review page at once, with readers split across all of them. Or skip the test: watch recordings and heatmaps in Microsoft Clarity, which its FAQ calls "a free service forever", fix what's plainly broken, then compare equal periods before and after.
How do I track affiliate sales for each version?
Your testing tool sees the click. The sale happens on the merchant's site, so only the merchant or network sees it. Tag each version's links with its own ID:
- Amazon Associates (US and Canada): create a second tracking ID under "Manage your Tracking ID" and use it on version B's links. Each account gets up to 100, and the Tracking ID Summary Report shows data per ID (Amazon Associates Help).
- Awin: add a click reference (ClickRef) to each version's links, up to 6 per link. Awin's Click References report shows clicks, transactions, sale value and commission per ClickRef (Awin help).
- Other networks: look for the same field, usually called a sub-ID.
Compare commission per 1,000 visitors for each version. Amazon credits orders placed within 24 hours of the click, and items added to the cart in that window if bought within 89 days (Amazon Associates: Program Policies, Commission Income Statement). Read earnings a few days after the test ends and expect small late additions.
Does A/B testing hurt affiliate SEO?
Little, if you follow Google's testing rules.
"Don't show one set of URLs to Googlebot, and a different set to humans." (Google Search Central: Minimize A/B testing impact in Google Search)
If version B lives at its own URL, Google asks you to:
- Point the copy at the original with rel="canonical", not noindex.
- Send readers there with a temporary 302 redirect, not a permanent 301. Google says "JavaScript-based redirects are also fine."
- End the test promptly and remove test URLs and scripts. Google warns that a test run "for an unnecessarily long time" may be read as "an attempt to deceive search engines."
Which A/B testing tools are free for an affiliate site?
Free tiers as of November 2026, from each vendor's own pages:
| Tool | Free tier | Fits |
|---|---|---|
| PostHog | 1 million feature flag requests a month (experiments are billed with flags), 5,000 recordings, no credit card; no-code editor | Owners who want point-and-click tests |
| GrowthBook | Starter: up to 3 users, unlimited experiments and traffic; no visual editor, so someone adds code | Sites with a developer; Pro is $40 USD per seat a month |
| Microsoft Clarity | Free recordings and heatmaps; it doesn't run the split itself | Finding what to test, or pages too small to test |
Example
A US kitchen-gear site on Amazon
The "best chef's knives" page gets 30,000 visitors a month; 10% click an Amazon link. Version B moves the top pick under the headline and uses its own tracking ID. A 20% lift needs 3,600 visitors per version, about a week; the owner runs two. B: 12.4% clicks and $200 USD commission per 1,000 visitors; A: 10% and $180. The owner publishes B and deletes the test.
A Canadian camping blog with 3,000 visitors a month
At a 5% click rate on Amazon.ca links, a 20% lift needs 15,200 visitors: five months. Instead, Clarity recordings show phone readers skipping the comparison table, so the owner moves a short pick list above it and compares four weeks before and after, checking last year's same weeks since camping is seasonal.
A software review site on a network
Version B shows three plans side by side instead of one recommended plan. Clicks rise from 8% to 9%, but commission per 1,000 visitors falls from $120 to $95 USD, because readers pick the cheapest plan. The owner keeps A.
FAQ
Can I A/B test Amazon affiliate links?
Yes. Give each version its own tracking ID, name Amazon on buttons, keep Amazon's required statement, and show prices only via Amazon's links or API.
What is a good click-through rate for affiliate links?
There's no standard: it depends on the topic, page type and where readers come from. Compare versions of the same page, measured the same way.
Can I test several changes at once?
Yes, in one version, but you won't know which change worked. Each extra version needs a full sample: three versions take 50% longer than two.
Should I test button colors?
Last. Smaller lifts need more visitors to detect, so start with bigger changes: the top pick, the product list, the button wording.
What if neither version wins?
Keep version A and test a bigger change. A tie means any difference was too small to detect; don't hunt through segments for a winner.
Different for you?
- Your pages don't get enough visitors to test: Is my SEO working?
- Your reviews need stronger first-hand proof: E-E-A-T for a small business website
- Traffic fell after a Google update: Traffic dropped after a core update
- You sell your own products instead of earning commission: Ecommerce SEO and Landing page best practices
- You want affiliate clicks counted in Google Analytics: GA4 key events
- You want pages built and tested for you: Website Agent