The ABWatcher blog

How to Find Your Conversion Rate Baseline: A Step-by-Step Guide

A numerate walkthrough for beginners: how to calculate your real conversion rate baseline, avoid selection bias, and set a defensible starting point before you run your first A/B test.

Sam Lee

Data Analyst · · Updated

TL;DR

  • Conversion rate = (conversions / total visitors) × 100 — measured over a minimum of 2 full business cycles, not a lucky week.
  • Excluding "bad" traffic (bot filters too aggressive, paid traffic cut, mobile web dropped) is the #1 cause of inflated baselines we see in beginner CRO programs.
  • ConversionTeam's benchmark data is built on 8,000+ A/B tests with a 61.1% win rate — even experienced teams lose roughly 4 in 10 tests, so your baseline needs to be solid enough to detect real lift, not noise.
  • ABWatcher detected live A/B tests on hero copy at Buffer (26 Aug 2026) and Lyft (8 Jul 2026) — both run on top of an established baseline, not a guess.
  • A 2% baseline conversion rate needs roughly 15,000+ monthly visitors per arm to detect a 20% relative lift at 95% confidence within a few weeks — know your math before you start testing.
  • Lucky Orange's CRO guide frames it well: the goal isn't a universal benchmark, it's beating your own number over time.

Why does a conversion rate baseline matter before you test anything?

A baseline is the reference point every future "win" gets measured against — without it, you can't tell signal from noise.

Here's the uncomfortable truth: most teams that jump straight into A/B testing without a clean baseline end up chasing variance. If your signup rate bounces between 1.8% and 2.6% week to week just from normal traffic mix changes, a test that "wins" by lifting conversion from 2.0% to 2.3% might just be sampling noise, not a real effect. You need at least 2-4 weeks of stable, unfiltered historical data before you can even calculate the sample size for your first test.

The formula itself is simple: Conversion Rate = (Number of Conversions / Total Visitors) × 100. If 15,000 people hit a landing page in a month and 300 sign up, that's a 2% baseline, per Lucky Orange's framework. The hard part isn't the division — it's making sure the numerator and denominator actually represent your real traffic.

What data should you collect to calculate it correctly?

Collect full-funnel visitor and conversion counts across every traffic source and device for at least one full business cycle before touching a test.

At minimum, pull:

  • Total sessions/visitors segmented by source (organic, paid, direct, referral, email)
  • Conversions defined precisely — one macro conversion (signup, purchase) and 2-3 micro conversions (add-to-cart, demo request, pricing page view)
  • Device and browser breakdown — mobile web often converts 20-40% lower than desktop in SaaS and ecommerce; blending them hides the real picture
  • New vs. returning visitor split — returning visitors frequently convert 2-3x higher and will inflate a baseline if not separated
  • Time-of-day and day-of-week patterns — B2B SaaS traffic can dip 30%+ on weekends, skewing a short sampling window

A baseline calculated on 10 days of data during a product launch spike, or one that silently excludes mobile traffic because "the mobile experience isn't done yet," is not a baseline — it's a cherry-picked snapshot that will make every subsequent test look artificially successful or unsuccessful.

What's the biggest mistake beginners make when measuring baseline?

The most common error is quietly excluding traffic segments that make the number look worse, which breaks the baseline's validity for every test that follows.

Three patterns we see repeatedly:

  1. Excluding bounce traffic or bot-filtered sessions inconsistently. If your analytics tool's bot filter changes between the baseline period and the test period, you're comparing two different populations.
  2. Measuring only "qualified" or paid traffic. Dropping organic or referral traffic because it "doesn't count the same way" shrinks your sample and introduces selection bias — the paid traffic that converts best isn't representative of your full funnel.
  3. Self-selection bias from opt-in sampling. If your baseline period happens to overlap with an email campaign to your most engaged list, or a press mention driving unusually high-intent visitors, your baseline conversion rate will be inflated relative to normal traffic. Any test run afterward will look like a loser purely from regression to the mean.

This is the same logic that makes multi-variant production tests legible when you track them long enough. ABWatcher caught Attio running an A/B test on its homepage layout — detected 11 July 2026 — swapping the position of the logo bar and pricing table, and showing two different hero product demos (a search-style "Find experienced marketing agencies" UI versus a workflow-style "Automatically qualify leads" AI card). A test like that only produces a trustworthy readout if Attio's baseline conversion rate was measured across comparable traffic mixes for both arms — otherwise a shift in paid vs. organic split between variants would get misattributed to the hero creative.

How much traffic do you need before your baseline is statistically meaningful?

You need enough volume that your baseline's natural week-to-week variance is smaller than the lift you're trying to detect.

Here's the napkin math: if your baseline is 2% and you want to detect a 20% relative lift (2% → 2.4%) at 95% confidence and 80% power, you need roughly 15,000-20,000 visitors per arm, depending on your conversion variance. If your baseline is thinner — say 500 visitors/month — you simply cannot detect anything smaller than a massive 50%+ relative lift without months of data collection. That's a real constraint, not a hypothetical one: a lot of early-stage CRO programs burn a quarter running tests that were underpowered from day one because nobody checked the baseline sample size first.

Baseline conversion rate Monthly visitors needed (per arm) to detect 20% relative lift Monthly visitors needed to detect 10% relative lift
1% ~31,000 ~124,000
2% ~15,000 ~61,000
5% ~5,900 ~23,500
10% ~2,800 ~11,200

ConversionTeam's data, drawn from over 8,000 A/B tests, puts the average test win rate at 61.1% — meaning even well-powered tests from an experienced practice lose nearly 40% of the time. If your baseline sample size can't support detecting a real effect, you'll see a much higher false "win" rate purely from noise, not genuine lift.

How do live test detections help you sanity-check your own baseline?

Watching how established brands structure ongoing tests shows you what a mature, properly segmented baseline supports.

Three patterns worth modeling:

  • Nationwide — detected 25 June 2026 — is running an A/B test on its middle feature block, alternating between "Online account access" messaging and a "Find a local agent" ZIP-code lookup CTA, varying by browser in the sample but clearly segment-targeted rather than randomly rendered. A test like this implies Nationwide has a baseline granular enough to separate self-service-intent visitors from agent-seeking visitors — exactly the segmentation beginners skip.
  • Ally — detected 17 September 2026 — is testing two full hero/banner combinations ("More bank for your buck" + WNBA promo vs. "The bank fans bank on" + a different promo), and the variants appear across multiple browser engines, confirming genuine randomization rather than a caching artifact. That cross-browser consistency check is the same discipline you should apply to your own baseline data: if your "baseline" conversion rate differs suspiciously by browser or device, you likely have a tracking bug, not a real signal.
  • Buffer — detected 26 August 2026 — is testing hero subheadline copy ("Connected to every platform and tool you use" vs. "Works with every platform you post to, and plugs into your favorite tools"), again split across both Chrome and Firefox sessions, ruling out rendering as the explanation. The lesson: a real test produces a difference that survives a segment cut. Your baseline should be stable enough that it does too.

Per Otter's CRO guide, the most fundamental test structure is presence-vs-absence — establishing whether an element matters at all before testing its variants. That only works if you've already nailed down what "normal" (baseline) performance looks like without the element.

Common questions

What's a good conversion rate to start from?

There's no universal benchmark worth chasing — ecommerce product pages often run 1-3%, SaaS trial signups 2-5%, and fintech account openings can be under 1% due to higher-friction compliance steps. The right move is calculating your own number cleanly, then improving it over time rather than comparing against an industry average that doesn't share your traffic mix.

How long should I measure before I trust my baseline?

Use a minimum of 2-4 full weeks, ideally spanning 2 complete business cycles (including weekends, billing cycles, or any recurring traffic spikes like newsletters). Shorter windows risk capturing one-off events — a product launch, a PR spike, a holiday — that inflate or deflate the number in ways that won't repeat during your actual test period.

Should I exclude bot traffic from my baseline?

Yes, but consistently — apply the same bot-filtering logic to both your baseline period and every subsequent test period. The mistake isn't filtering bots; it's changing your filtering rules between measurement periods, which silently shifts what counts as a "visitor" and makes before/after comparisons invalid.

Does mobile traffic need a separate baseline?

Generally yes. Mobile web conversion rates commonly run 20-40% lower than desktop for complex funnels like checkout or multi-step signup, so a blended baseline can mask real segment-level performance. Calculate device-level baselines separately, then decide whether to test device segments independently.

How do I know if my baseline is biased by self-selection?

Check whether your measurement window overlaps with any non-organic traffic driver — an email blast, a press mention, a paid campaign surge — that would skew visitor intent higher or lower than normal. If a campaign-driven spike represents more than roughly 15-20% of your baseline sample, re-measure during a quieter period before locking in the number.

Your next sprint

Before you spin up your first A/B test, spend one sprint purely on baseline hygiene: pull 2-4 weeks of unfiltered, device-segmented, source-segmented conversion data, confirm your bot-filtering rules are locked, and run the sample-size math for the lift size you actually care about. A clean baseline is cheaper to build now than a quarter of mistrusted test results later.

Read next

More from the ABWatcher blog

From the glossary

Vocabulary the article uses. See all terms →

  • Shrinkage Estimator— A statistical adjustment that pulls noisy per-variant or per-segment estimates toward an overall average to reduce overfitting to randomness.
  • Flicker Test— A diagnostic check to confirm whether users briefly see the original page before a variant renders in client-side experiments.
  • Pre-Experiment Bucketing Audit— A review process confirming users are being randomly and evenly assigned to experiment variants before launch.
  • Above-Fold Exclusion Rule— A test-design rule that deliberately prevents changes to above-the-fold elements so an experiment isolates below-the-fold effects.
  • Flicker-Free Testing— A technical setup that prevents users from briefly seeing the original page before a test variant loads.
  • Orthogonal Test Design— A test structure where multiple variables are changed independently so their individual effects can be measured without interference.

See more like this

ABWatcher catches A/B tests like this across 1,000+ brands.

Watch live experiments at 1,000+ high-converting brands, complete with hypothesis and takeaway. 7 days of Pro, free, no card, then Free forever for ten watched companies.