Documentation
Cannibalization Entropy: find queries split across pages
Measures how evenly each query's impressions are shared between your pages, in bits of Shannon entropy, and bands the results from Healthy to Critical.
Cannibalization Entropy measures, for every query that reaches two or more of your pages, how evenly its impressions are shared between them. When one page takes nearly all the impressions the score is close to zero; when several pages share them equally it is high. The score is Shannon entropy in bits, and a threshold you set decides which queries are listed as high-entropy and worth a closer look.
How the score is measured
For each query, every page that earned impressions gets a share: its impressions divided by the query's total. The entropy of those shares is
H = −Σ p × log₂(p)
Only queries with impressions on at least two pages and at least 10 impressions in total are analysed. Evenness divides H by the most it could be for that many pages, log₂ of the page count, so 100% means a perfectly even split whatever the number of pages. Top page share is the biggest single share.
| How the impressions are split | Entropy | Evenness |
|---|---|---|
| 90% / 10% over two pages | 0.47 bits | 47% |
| 70% / 30% over two pages | 0.88 bits | 88% |
| 50% / 50% over two pages | 1.00 bit | 100% |
| 60% / 30% / 10% over three pages | 1.30 bits | 82% |
| 25% each over four pages | 2.00 bits | 100% |
Entropy grows with the number of pages, so when two queries reach different numbers of pages, compare their evenness rather than their entropy.
Threshold and bands
The Entropy threshold slider runs from 0 to 3 bits in steps of 0.1, with 1.0 as the default. Every band is set relative to it:
- Healthy: below half the threshold.
- Moderate: from half the threshold up to the threshold.
- High: from the threshold up to one and a half times it.
- Critical: one and a half times the threshold or more.
At the default, a 50/50 split between two pages is High and a 90/10 split is Healthy. Moving the slider shifts every boundary, the histogram and the High-entropy tab at once, without fetching the data again.
Reading the results
- Cards: Queries analysed (with how many were seen and excluded), High-entropy queries at or above the threshold, Impressions at risk on those queries (with their clicks) and Average entropy.
- Entropy distribution counts queries in 0.1-bit bins, coloured by band.
- The table lists Query, Pages, Impressions, Clicks, Entropy (bits), Evenness, Top page share, Band and Impression share by page, a bar of up to six pages. Hover a segment for the page, its share and its average position. Tabs switch between High-entropy and All analysed, and the search box matches a query or its top page.
- The CSV export carries every page's share as text, not only the six in the bar.
When a high score is not a problem
When Search Console groups data by page, each of your pages shown for a search collects its own impression. Two of your pages listed on the same results page therefore both score, and holding two listings can be a strength rather than a fight. Entropy over a date range cannot tell that case apart from Google switching between the pages from one day to the next.
Open the query's detail page before acting. If both pages hold good positions at the same time, leave them. If the positions are weak and the leading page keeps changing, the pages are competing.
Fixing pages that compete
- Decide which page should rank for the query, usually the one that answers it most completely.
- Move the useful material from the other page into it, then either redirect the weaker page or point it at the stronger one with a canonical link, as Google recommends for consolidating duplicate URLs.
- Update internal links so they point at the page you chose.
The Cannibalization report shows the same problem by URL, with the impressions and clicks that went to the weaker pages.
How AutoPSEO avoids creating it
The content engine checks the topics it finds against the pages your site already has, so it does not write a second article for a search an existing page answers. Two of its own articles that turn out to cover the same searches can be merged: one is taken down and your site answers its old address with a 301 to the one kept. See the content engine.
Limits
- The report reads up to 25,000 query and page rows for the period. Rare pairs can be missing, and queries Google anonymises for privacy never appear.
- Impressions are the only input. A page that collects impressions far down the results counts the same per impression as one at the top, so check positions in the bar before you merge anything.