Publishing a second page about a topic you already cover usually makes you less visible in AI answers, not more. Retrieval selects passages, and two of your own passages saying nearly the same thing in nearly the same words split the signal rather than doubling it. SQSEO surfaces this the moment you map your prompt set against your own content: the questions where you rank worst are frequently the ones you have answered three times.
Cannibalisation moved from pages to passages
Classic SEO had a version of this problem. Two pages targeting one keyword competed, and the fix was consolidation or clear differentiation.
The AI search version operates one level down. Your pages are chunked into passages, and it is passages that compete. That means two pages can coexist perfectly well while a section inside each of them competes, and it means a single long page can compete with itself if it covers the same ground twice in different sections.
Retrieval works by comparing an encoded query against encoded passages, the mechanism established by dense passage retrieval and measured across models in benchmarks such as MTEB. When several candidate passages sit at almost the same point in that space, none is clearly the best match, and the practical outcome is that the engine may take one arbitrarily, take a competitor’s clearer passage instead, or blend two of yours into something you did not write.
What duplication actually costs you
| Situation | What happens in retrieval | Visible symptom | Fix |
|---|---|---|---|
| Two pages, same question, similar depth | Similar vectors, no clear winner | Inconsistent citation, different URL each run | Consolidate into one, link from the other |
| One long page, topic covered twice | Two near identical chunks | Vague answers citing your page | Delete one, deepen the other |
| Series of posts each restating basics | Many shallow matching passages | Cited for basics, never for specifics | Keep one primer, make the rest specific |
| Localised pages with same content | Duplicate across languages | Wrong language version cited | Write natively per market |
| Product page and blog post overlap | Commercial and editorial compete | Blog cited for buying queries | Differentiate by intent |
The row that surprises teams most is the third. Publishing consistently is good advice, and it turns into a problem when every post opens with three paragraphs defining the category. Those introductions are near identical passages across dozens of URLs, and collectively they teach retrieval that your site is a source of definitions rather than of specifics.
How to find your own competing passages
You do not need a tool. You need a list of questions and a list of pages.
Write out the questions your site actually answers, one line each, phrased as a buyer would ask them. Then for each question, list every URL on your site containing a section that answers it. Any question with more than one URL at similar depth is a candidate.
The second pass is to read the competing sections side by side. Sometimes they genuinely differ, one for beginners and one for practitioners, and the fix is to make that difference explicit in the opening sentence of each. Often they are the same answer written twice, and one should become a link.
A quick proxy when the site is large: search your own domain for a distinctive phrase from the question. If several pages return with substantially the same passage, you have found a cluster.
Consolidate at section level, not page level
The classic instinct is to merge two pages into one and redirect. That is often too blunt now, because the pages may serve different purposes even while one section overlaps.
The lighter move is to remove the overlapping section from the weaker page and replace it with a sentence and a link. The page keeps its purpose and its URL, and the topic now has one canonical passage on your site rather than two competing ones.
Choose the survivor by retrieval history rather than by which page you prefer. The page already being cited has accumulated the signals, and moving the canonical answer to a fresher page means competing with your own incumbent. This is the same reasoning behind updating old content rather than republishing it.
Differentiation is usually better than deletion
Deleting content feels decisive and is frequently wrong. Two pages covering adjacent ground can both be valuable if each is sharpened to a distinct question.
The test is whether you can write, in one sentence, who each page is for and what specific question it answers, without the two sentences overlapping. If you can, rewrite the opening of each to say exactly that, and trim the parts that stray into the other’s territory. If you cannot, one of them is redundant.
Sharpening usually means getting more specific rather than broader. Two general pages compete; a general page and a narrow one do not, because their passages land in different regions of the space. Specificity is what separates them, which is the same property that makes any single passage retrievable at all, and the same logic behind writing RAG friendly content.
The introduction problem
Most content management workflows encourage every article to stand alone, which produces a definitional preamble on every post. Across a hundred posts that is a hundred near identical passages.
The fix is not to remove context, since chunks do need to carry their own subject. It is to make the context specific to that article rather than generic. Instead of three paragraphs defining the category, one sentence naming the subject in the context of this article’s particular question. That preserves chunk independence without manufacturing duplicates.
A useful rule: context should orient the reader to this question, not to the field. If the opening paragraph could be moved to any other article on the site without editing, it is a duplicate passage.
Programmatic and templated pages need special care
Sites generating pages from a template, one per city, per integration, per use case, create duplication at scale by construction. The template text is identical across hundreds of URLs and only the variable parts differ.
Where the variable content is genuinely substantial, the pages are fine. Where it is a name swapped into boilerplate, every page produces effectively the same passage, and the set collectively matches nothing well. Search engines have long treated this as thin content, and Google’s own framing of scaled content abuse applies to the same behaviour.
The practical bar is whether each page contains at least one passage that could only have been written about that specific variant. If not, fewer and deeper pages will outperform the set.
Intent collision between commercial and editorial pages
A distinct flavour of self competition happens between a product page and a blog post covering the same ground. Both are yours, both are legitimate, and they answer the same question with different motives.
The blog post usually wins retrieval, because it is written to explain and therefore contains more complete claims, while the product page is written to persuade and contains more positioning. That sounds fine until you notice what it means: buying intent queries get answered from an educational page that does not mention pricing, availability or how to start.
The resolution is to split by intent explicitly. Let the educational page own the how and why, and make sure the commercial page owns the what it costs, who it suits and how to begin, in prose rather than in interface elements. Then link them, so a reader arriving at either can reach the other. That division also stops the two pages producing similar passages, because their claims genuinely differ.
Series, hubs and the pillar that repeats its spokes
Hub and spoke structures are good for navigation and risky for duplication. The failure mode is a pillar page that summarises every spoke in enough depth to answer the same questions the spokes answer.
A pillar that works is an index with orientation: it says what each spoke covers and who should read it, and it answers the one question none of the spokes answer, which is usually how the pieces fit together. A pillar that competes is a long article restating each spoke’s conclusions. The first produces one distinctive passage plus clear routes onward. The second produces a diluted duplicate of every spoke.
The same applies to series. If part four opens by recapping parts one to three, that recap is three duplicate passages. Link instead, and trust readers arriving mid series to follow a link, which they will.
Fixing duplication is cheaper than publishing more
The instinctive response to poor AI visibility is to publish more, and on a site that already has overlapping coverage that instinct makes things worse. Each additional page on a covered topic adds another near identical passage to the pool your own content is competing in.
Auditing existing coverage first is faster and cheaper than a content sprint, and it frequently frees capacity. The forty post example below took two weeks of editing against what would have been months of new writing, and it moved more. Research on generative engine optimization points the same way in a different register: what changes how engines use a source is the quality and explicitness of the claims inside it, not the number of documents carrying them.
The planning rule that falls out of this is to check whether a question is already answered on your site before commissioning anything, and if it is, decide whether to sharpen the existing answer rather than add a competing one.
Measuring whether consolidation worked
Track two things before and after: which URL is cited for each target question, and how consistently. The improvement from fixing duplication shows up first as consistency, because the failure mode was arbitrary selection among similar candidates.
Use equal run counts before and after, and hold the prompt set fixed, or the comparison is void. Expect a lag: the old passages remain in caches and indexes for weeks after you edit them, so a fair read needs a month. The full timeline is set out in how long it takes to appear in AI answers.
Also watch for the opposite outcome. If consolidation reduced coverage, because the removed section was the only thing matching a related query, citations for that adjacent question will drop. That is recoverable by restoring a sharpened version of the section, targeted at the adjacent question specifically.
A worked example: forty posts, one answer
A company had 40 blog posts on one subject area and was cited for general definitional queries but never for buying queries.
The audit mapped 18 buyer questions against those posts. Fourteen of the questions had three or more posts each containing a section that answered them, all at similar depth, all opening with the same category definition.
The consolidation took two weeks. Each question got one canonical section on one page, chosen by existing citation history. The other posts kept their unique material and replaced their overlapping sections with a sentence and a link. Generic introductions were replaced with article specific openers.
Nothing new was published. Across the following six weeks, citation consistency for the 14 questions rose sharply and four buying queries began citing the company for the first time, because the canonical passages were finally distinct enough to match a specific query rather than a definitional one.
Key takeaways
Cannibalisation now happens between passages, so two pages can coexist while their sections compete, and one long page can compete with itself. Near identical passages produce arbitrary selection, blended answers, or a competitor’s clearer passage winning instead. Map your buyer questions against the URLs that answer them, consolidate at section level rather than merging whole pages, and pick the survivor by citation history rather than preference. Replace generic introductions with article specific ones, since a preamble that could sit on any post is a duplicate passage by construction.
