A client portfolio had five pages targeting “how to choose an AEO agency” and adjacent prompts. A pillar, a blog post, a use-case page, a comparison piece, and a case study, all reasonably well-written, all targeting variants of the same buyer question.

Across 60 days of tracking, the engines cited them — but they rotated. Week 1, Perplexity quoted the pillar. Week 2, the blog post. Week 3, the use-case. Citation count looked OK on the rollup. Citation consistency on any single URL looked terrible.

We have written before in AI source selection about how LLMs almost never cite two URLs from the same domain in one answer. That observation has a corollary nobody warned us about — five pages competing for one slot on a prompt cluster are not five times more visible. They are one slot, distributed five ways. This is internal citation cannibalization, and it is more common than backlink cannibalization ever was in classical SEO.

The hard constraint

Across 240 AI answers we measured over 90 days, fewer than 7% contained two URLs from the same domain. Zero contained three. The pattern is consistent across all five major engines.

What this means in practice. If you have ten pages on your site targeting variations of the prompt “best AEO agency for SaaS,” all ten of them are not competing against your competitors’ pages on that prompt. They are competing against each other for one citation slot on your own domain. The competitor field is a separate competition.

Classical SEO did not have this constraint with the same severity. Google might rank multiple URLs from one domain on the same SERP (especially when sitelinks expand a top-ranking result). LLMs do not. The slot is single-occupancy.

How cannibalization shapes up

Five common shapes we see across client portfolios.

The pillar-plus-blogs cluster. A 4,000-word pillar page on a topic, plus 8 to 12 blog posts each covering a sub-topic. All of them target related prompts. The pillar should be the canonical citation home, but the blogs are individually sharper on specific sub-prompts and steal the citation 40% of the time.

The use-case sprawl. A SaaS with a use-case page per buyer segment — “for product teams,” “for ops teams,” “for marketing teams” — all targeting “best [tool] for [segment]” style prompts. Each use-case page is fine individually. As a group, they fragment the citation pool for the parent “best [tool] for B2B” prompt.

The case study orbit. A pillar page on a methodology, plus 5 case studies that apply the methodology. The case studies frequently outrank the pillar on specific tactical prompts because they have concrete numbers and named clients. The pillar starves.

The blog cluster on one keyword. Three or four blog posts on essentially the same topic, written over 18 months as the topic evolved. None of them links to a canonical winner. Engines rotate.

The translation collision. EN and UA versions of the same page both appearing in the same engine’s pool when the prompt is language-agnostic. Less common, but we have seen it on technical topics where the EN version has more authority but the UA version has fresher data.

How to diagnose

The signal is rotation, not absence. If your tracking shows that for a given prompt cluster you have multiple URLs each cited maybe 10-15% of the time, never above 30%, that is cannibalization. A non-cannibalized page either holds 50%+ citation share or is not in the pool at all.

The check.

For each tracked prompt cluster, pull which of your URLs got cited across the last 60 days. If three or more URLs share the citation pool with no clear winner, you have a cannibalization problem.

The deeper check — look at the rerank scores in the candidate pool, not just the final citations. If you can pull Perplexity’s source list per query, you will often see four of your URLs in the candidate pool for one prompt, with the model picking different ones on different days based on small reranker shifts. That is the smoking gun.

The consolidation play

Pick a canonical winner. Reroute everything else. This is the move, and it is uncomfortable because it usually means de-emphasising pages your team is proud of.

Step 1 — pick the winner. For each cannibalized cluster, decide which URL deserves to be the canonical citation home. Criteria — strongest snippet structure, best author byline, freshest dateModified, highest existing inbound link authority. Usually it is the pillar, but not always. Sometimes the blog post is genuinely sharper than the pillar.

Step 2 — decide the fate of the others. Three options per page. Redirect (301) if the page truly duplicates the winner’s purpose. Canonical-tag (rel=canonical pointing to the winner) if the page has secondary value but should not earn the citation. Reposition (rewrite to target a different, non-cannibalizing prompt cluster) if the page is solid but mistargeted.

Step 3 — reroute internal links. Every internal link on the site that pointed to one of the consolidated pages now points to the canonical winner. Anchor text aligned with the winner’s targeted prompt. This is the move that signals to the retriever that the canonical winner is the authoritative answer for the cluster.

Step 4 — wait six weeks. Citation consolidation is not instant. The retrieval indices need to refresh. Engines need to re-evaluate which URL deserves the slot. In our portfolio, six weeks is the median time from rerouting to measurable citation lift on the canonical winner.

What we measure after consolidation

Across 12 client portfolios where we executed consolidation between 2025 Q3 and 2026 Q1, the median outcome.

Citation share on the canonical winner — from a fragmented 15-25% (split across multiple URLs) to a consolidated 45-65% on the same prompt cluster. The total citation share for the brand on the cluster typically also lifts by 20-40%, because the consolidated page wins more reranks against competitors when it is not fighting itself.

The pages that got redirected or canonical-tagged — they lose their citations, predictably. The trade-off is favourable because the winner gains more than the losers lose, and citation consistency on the canonical winner improves downstream traffic and trust signals over time.

What not to consolidate

Two cases where consolidation is the wrong move.

When the pages target genuinely different prompt clusters. “How to do X” and “When to hire someone to do X” feel adjacent but pull different buyer intents. Engines treat them as separate clusters. Consolidating these collapses two independent citation surfaces into one and loses coverage.

When the secondary pages exist for other reasons. A case study that mostly serves sales enablement, even if it incidentally targets a prompt cluster, should not be canonical-tagged into invisibility — its primary job is the sales conversation, and the AEO loss is acceptable.

The rule of thumb — consolidate when the only reason multiple pages exist is keyword/prompt coverage. Preserve when the pages have distinct purposes for their primary audience.

Operationalising consolidation as part of the rhythm

Consolidation is a one-quarter project the first time. After that it folds into the quarterly tier of the refresh cadence rhythm — once per quarter, the practice lead reviews the citation distribution per prompt cluster and flags any new cannibalization.

The shape that wins long-term — one canonical page per cluster, internal links coherently pointing at the canonical, supporting pages that exist for non-AEO reasons clearly labelled as supporting. Brands that drift toward this structure consistently outperform brands that keep accumulating pages around the same prompts.

LLMs decided the rule. Five pages on one prompt is one citation, not five. Build for the constraint instead of fighting it.