# SEO Split Test Analyzer

> Free, in-browser tool that tells you whether an SEO change worked, by comparing pages you changed
> (test) against similar pages you left alone (control), before and after the change date.
> URL: https://seo-split-test.skillsafe.ai/

## Input
- Daily totals as CSV: one file with a date column plus test and control columns
  (e.g. date,test_clicks,control_clicks); or one row per day per group (date,group,clicks);
  or two files, one per group, such as two Google Search Console "Dates" exports.
- The date the change went live, and an optional settling window (0 to 28 days) that is drawn
  but not counted while search engines recrawl.
- Any count metric works: clicks, sessions, impressions, visits, pageviews.

## Method
- Counterfactual: Bayesian linear regression of the test series on the control series, fitted on
  the days before the change, on a log scale by default, with a weekday pattern and AR(1) errors.
  The control coefficient has a prior centred on proportional movement (elasticity 1), which stops
  noise in the control from biasing the estimate. 1,000 simulated forecasts give the 95% band.
- Effect: actual minus expected per day and in total, the relative lift with a 95% interval, and
  the share of simulations below the actual total (chance the effect is positive).
- Cross-check: difference-in-differences on daily log(test/control) with Newey-West errors.
- Placebo checks: the same analysis at up to six fake change dates before the real one.
- Verdict: positive effect, negative effect, leaning positive, leaning negative or no clear effect,
  with cautions (short baseline, loose fit, low volume, false-alarm-prone setup) and a rough
  estimate of how many more days an inconclusive test needs.

## Output
- Verdict sentence, lift, extra or lost clicks, chance positive, days counted.
- Charts: actual vs expected with 95% band, daily difference, running total. Table view.
- Downloads: Markdown report, daily results CSV, chart PNG; copy summary.

## Privacy and cost
Everything runs in the browser. No upload, no sign-in, no charge.

## Examples
Three built-in synthetic examples: a title-tag rewrite that worked (+14% injected), a change with
no effect, and a change that hurt (-11% injected).