How to Run a GEO Page-Collision Audit When AI Systems Cite the Wrong URL

Sometimes the brand is present, but the page that wins is still wrong.

That sounds like a minor problem until you look closer.

A pricing prompt cites an old educational blog post. A software-evaluation prompt cites a generic integrations directory instead of the specific connector page. A comparison prompt keeps pulling a support article because the commercial page never answers the trade-off cleanly near the top.

The dashboard says your brand appeared. The business result says the wrong asset carried the answer.

I see teams miss this all the time because they track visibility at the domain level, not the URL level. That makes page collisions look harmless. They are not harmless. They confuse buyers, weaken the follow-up journey, and hide real architecture problems inside what looks like healthy coverage.

This guide is deliberately narrower than our work on citation-loss RCA, the GEO content map, the internal-linking audit, and the HTML parity audit. Those posts help you diagnose prompt loss, map page types, inspect support links, or catch rendering gaps. This one is about a specific failure mode in the middle: AI systems are citing your brand, but they keep choosing the wrong internal URL.

What a page collision actually is

A page collision happens when two or more internal URLs can plausibly answer the same prompt family, but the one AI systems keep reusing is not the one you want to win.

That can happen across:

• •blog post versus money page
• •old page versus refreshed page
• •generic page versus specific page
• •support doc versus commercial page
• •parent hub versus detailed child page

The problem is not always duplication in the classic SEO sense.

Sometimes both pages are useful. The real issue is that one page answers the prompt more directly, proves the point more clearly, or sits in a stronger internal-routing position than the page that should own the commercial or strategic outcome.

Why this matters more in GEO than in classic organic reporting

In classic SEO, ranking overlap usually shows up in your page reports. In GEO and AEO work, the failure can stay hidden because the model still names your brand.

That creates three bad habits:

If a prospect asks, "Does this integrate with Salesforce?" and the model cites a vague directory page instead of the specific connector page, you did not really win that prompt. You rented it.

The four collision patterns I see most often

Blog post versus money page

An educational article wins because it explains the topic more directly than the commercial page.

This is common when the money page leads with positioning language, while the blog post gives a blunt answer in the first screen.

Old page versus new page

A legacy page keeps getting cited because it has more internal links, cleaner headings, or older backlinks that still reinforce it.

The newer page may be better. It may also be much harder to retrieve and classify.

Generic page versus specific page

A broad hub wins instead of the exact page the buyer needs.

This shows up on integration, pricing, service, and category pages all the time. The broader page has stronger routing, but the narrower page has the truth. AI systems often pick whichever page makes the answer easiest to assemble.

Support page versus commercial page

A help doc wins because it contains the crispest implementation detail, definition, or setup answer on the site.

That is not always bad. It becomes a problem when the support page steals prompts that should feed evaluation, shortlisting, or sales conversations.

Start the audit with a URL-level prompt sheet

Do not begin by reading pages and guessing where the overlap might be.

Start with the actual prompts that matter and log what URL gets cited today.

A simple sheet is enough:

This does two important things.

First, it stops the audit from becoming abstract. Second, it forces the team to define the page that should own the prompt before anyone touches copy or links.

Compare page jobs before you compare wording

This is the step most teams skip.

They open both pages, notice overlapping language, and conclude they have a copy problem. Usually they have a page-job problem.

Ask what each page is supposed to do.

If the wrong page is winning, the fix is often about clarifying the page roles and rebalancing routing, not just deleting overlapping sentences.

The audit itself: what to inspect side by side

Take the intended page and the page that keeps getting cited. Review them against the same grid.

This is where our HTML parity audit often connects. If the intended page only exposes the best answer after hydration, and the wrong page exposes it in raw HTML, the model's choice stops looking mysterious.

Name the collision type before you assign the fix

You need one label. Without it, the team will throw mixed fixes at the problem.

I like five practical labels:

That label becomes the handoff sentence.

"This is a routing collision." "This is an answer-shape collision." "This is a page-role collision between the comparison page and the blog post."

That is much more useful than saying, "these pages overlap a bit."

The most common fixes for each collision pattern

If it is a routing collision

Strengthen the intended page's support system.

That usually means:

• •adding internal links from adjacent evaluation pages
• •tightening anchor language
• •linking from the page that keeps stealing the prompt
• •reviewing nav, breadcrumbs, and hub-to-child relationships

This is where the internal-linking audit matters. The right page often loses because the site keeps telling the model another URL is more central.

If it is an answer-shape collision

Rebuild the top of the intended page so it answers the prompt as directly as the wrong page does.

Do not hide the actual answer under brand language, soft positioning, or a vague hero. Give the page a first-screen answer block that states the truth in plain language.

This is especially common on pricing pages, comparison pages, and integration pages.

If it is an evidence collision

Move the proof closer to the claim.

The wrong page may be winning because it has one specific benchmark, example, control detail, or setup note that makes the answer feel safer to quote. If that is the case, a broad rewrite will not help much. A proof upgrade will.

If it is a retrieval collision

Check canonicals, raw HTML, breadcrumb output, tabs, accordions, and template behavior before you rewrite anything.

A technically weaker page can lose even when the editorial plan is right. That is why the release checklist and change log matter. Template changes often create wrong-page wins that get misread as editorial drift.

If it is a page-role collision

Decide which page should own the prompt family, then make the other page support it.

That can mean:

• •narrowing the stealing page's scope
• •adding a stronger handoff link
• •rewriting the intro so the page frames rather than owns the answer
• •consolidating sections if the split no longer makes sense

The goal is not to let every page compete. The goal is to make the right page easiest to choose.

A practical teardown example

Imagine a software company that wants /integrations/salesforce to win prompts like:

• •does this integrate with Salesforce
• •is the Salesforce integration native
• •what syncs between this tool and Salesforce

Instead, AI systems keep citing /integrations.

Here is the wrong way to interpret that result:

Great, the brand is still visible.

Here is the useful diagnosis:

That is a fix-ready result.

What not to do when you find a collision

A few bad responses show up again and again.

Do not strengthen both pages equally

If both URLs keep getting more content, more proof, and more routing support, you can intensify the collision instead of resolving it.

Do not canonicalize everything out of fear

Page collisions are not always duplicate-content issues. Sometimes both pages need to exist. The problem is role clarity and prompt ownership, not indexation alone.

Do not force the commercial page to mimic the blog post word for word

The lesson is usually about answer clarity, proof placement, or routing. Copying the same content onto every page creates new overlap.

Do not stop at one prompt

If the wrong URL only appears once, that may be noise. If it repeats across a cluster, it is a pattern.

How this connects to the rest of your GEO operating system

A page-collision audit works best when it plugs into the systems around it.

• •Use the content map to define which page type should own each prompt family.
• •Use the citation-loss RCA when the issue may be broader than a wrong-URL result.
• •Use the internal-linking audit when routing support looks weak.
• •Use the HTML parity audit when the intended page may be technically harder to retrieve.
• •Use the release checklist and change log to catch new collisions after template or proof changes ship.

That is why I like this audit for evening-slot operator work. It sits between measurement and remediation. It tells you which page should own the answer and why it is losing today.

A compact weekly workflow for teams that keep seeing wrong-URL citations

That loop is simple on purpose. If you need a giant workflow board to fix wrong-page citations, the process is probably hiding weak diagnosis.

FAQ

Is a page collision the same thing as keyword cannibalization?

Not exactly.

Classic keyword cannibalization usually describes multiple pages competing in search results. A GEO page collision is narrower. It describes multiple internal pages that can answer the same prompt family, where AI systems keep citing the URL that is not supposed to own the prompt.

Should I merge the pages when I find a collision?

Sometimes, but not by default.

If both pages serve different jobs, keep both and make the ownership clearer through routing, answer shape, and scope. Merge only when the split no longer serves a real user or business purpose.

What if the wrong page is a support article and it genuinely answers the question best?

Then decide whether that prompt should belong to support or to the evaluation cluster.

If the prompt is truly implementation or troubleshooting oriented, support may be the right winner. If it is a buyer-stage fit prompt, the commercial page usually needs a clearer answer and better proof.

How many prompts should I use in the audit?

Use a compact cluster, not one isolated query and not fifty random ones.

I prefer a small family of closely related prompts that reflect the same buyer question. That is enough to show whether the wrong URL is winning by accident or by pattern.

The bottom line

When AI systems cite the wrong internal page, the brand is not actually winning cleanly.

It is leaking authority to the wrong asset.

The fix starts when you stop asking, "Did we appear?" and start asking, "Which URL carried the answer, and was it the one built for that job?"

If your team keeps seeing brand mentions tied to the wrong page, that is exactly the kind of problem we diagnose in our technical GEO reviews. We map prompt clusters to URLs, identify the collision pattern, and turn it into fix-ready work your content, SEO, and engineering teams can actually ship.