Cannibalization audit: which page wins, and why

Two pages ranking for the same query costs you nothing by itself. What costs you is instability: Google alternating between the URLs, each one relearning the query from a worse starting position every time. My working definition is a query where the ranking URL has swapped at least three times in ninety days and neither URL has held a position better than the other's best. On a five-hundred-page site that test usually returns somewhere between eight and twenty real clusters, not the several hundred that a text-similarity tool reports. This engagement identifies the real ones, names the page that should own each query, and specifies what happens to the pages that lose. How the map is built: Search Console query-to-URL data over sixteen months, pulled through the API or the BigQuery bulk export rather than the interface, so the sampling and the 1,000-row cap stop distorting the picture. Position history per URL from Ahrefs to see the swaps rather than a single snapshot. A crawl for the internal link graph, because the most common cause of a swap is not similar copy, it is that the site's own navigation points at the weaker page. Similarity of wording is checked last and only as corroboration. It is the signal everyone starts with and the one that produces the most false positives. Deciding which page wins: The page that converts wins, not the page with more clicks. This is where the engagement earns its fee and where clients most often disagree with me. A guide pulling four thousand monthly visits and no pipeline loses to a comparison page pulling six hundred visits and eleven demo requests, and the guide gets rewritten to feed it rather than compete with it. The other inputs, in order: referring domains that would be lost or carried by a redirect, the page's fit to the SERP's dominant format, and whether the URL sits inside a template your team can maintain. Traffic history is a tiebreaker, not a criterion. What happens to the pages that lose: Four outcomes, and the plan names one per URL with the reason. Merge and redirect, the default only when the losing URL has referring domains worth carrying across. De-optimise and keep, when the page serves a real audience at a different stage and simply needs its title, headings and internal anchors pointed away from the contested query. Noindex and keep, for pages that exist for logged-in users or sales, not search. Leave alone, which applies more often than sellers of this service admit, because a redirect chain built to fix a query that was never unstable is a self-inflicted wound. The uncomfortable finding: On sites with programmatic or templated content, the honest answer is frequently that consolidation is treating a symptom. If your taxonomy generates one page per keyword variant, you will regenerate the same overlap within two quarters and pay for this audit again. In that case the deliverable changes: a rule for what earns its own URL, applied to your existing set, plus the redirect map for what should never have been created. It is a harder conversation with the team that built the template, and I will have it with them present rather than sending you in alone. Scope boundaries: I do not write the surviving pages. I do not implement redirects, edit templates or deploy anything. I specify; your team ships, and I review what shipped in the packages that include a follow-up. No link building, no outreach, no disavow work. No promise about position for a consolidated URL, because that depends on execution I do not control and on competitors I cannot see. Who should not buy this: Anyone who wants a list of near-duplicate pages. A crawler licence costs a fraction of this and does that job. Sites in the middle of a replatform or a domain move: finish it, wait six weeks, then look, because everything will read as cannibalisation while redirects settle. Sites with under roughly 150 indexed pages, where the overlap is usually visible without an audit. And teams whose real question is why traffic dropped last month. That is a different diagnosis, and cannibalisation is the answer people reach for when they have not found the actual one yet. Timing: Do not schedule this to overlap a core update rollout or a seasonal peak. Position volatility during either makes the swap test unreliable, and I would rather move the start date by three weeks than hand you a map built on noise. Work is delivered in English, Portuguese or Spanish, and multilingual sites are analysed per market, since the same two URLs can be stable in one language and fighting in another.