Your Own Pages Are Stealing Your AI Citations

You published thirty posts last year, and four of them answer the same customer question in slightly different words. That overlap is not neutral storage. In the best public data we have on how ChatGPT picks its sources, a domain with two tightly matched pages on one question earns a citation about 6.2% of the time, and a domain with six or more pages on that same question falls to about 1.7%. Your own pages are suppressing your own citations, and the repair is subtraction.

Two pages on one question get cited. Six cut your odds by three quarters.

Two tightly matched pages from one domain is the highest rate the study observed. At six or more pages on the same intent, the rate falls to roughly a quarter of that. Read it as a workload number instead of a percentage and it stings more: at 1.7% you need roughly three and a half times as much retrieval to earn the same single mention you were getting for free at 6.2%.

The numbers come from Suganthan Mohanadasan, co-founder of Snippet Digital, who logged 57 ChatGPT conversations in July 2026 and hand-labeled all 3,554 pages the model retrieved to see which ones actually earned a citation in the final answer. He published the analysis on his own site and in Search Engine Journal on August 14, 2026. He is careful about the limits of it, and so am I: the percentages are directional, drawn from one ChatGPT Plus account over two days, not a census of the internet.

You have read the "quality over quantity" advice before, and it stops one step short. It tells you not to churn. It does not tell you that the pages you already published are actively taxing each other, which of your four overlapping pages should survive, or how to kill the other three without losing the rankings those pages hold. That is the whole job of this article, and it comes from running my own archive under a strict one-page-per-intent rule, a pattern I have watched repeat across 300-plus businesses in the USA, Canada, the UK, Singapore, Australia and New Zealand over seventeen years.

Owners were sold a page count, not an answer

Google rewards freshness and consistency, the agency said, so four posts a month shipped and the invoice made sense. The instruction was not wrong a decade ago. It aged badly, and it aged badly in a specific way: the unit that gets rewarded stopped being the page and became the question. Carolyn Shelby made the same argument in Search Engine Journal on June 17, 2026, that AI retrieval rewards semantic precision rather than publishing volume. Precision means one page that unmistakably answers one thing.

The mechanism underneath is not mysterious. When somebody asks a real question, the model does not run one search, it runs several, because one question fans out into many searches, each phrased slightly differently. Every one of those searches can pull in a batch of your pages. If four of them look like near-copies of each other, the model has to spend its judgment picking between your own pages instead of picking you over a competitor. It settles that by taking one, or by taking none.

The scale of the filter explains why. In Suganthan's labeled set, only 110 of the 3,554 retrieved pages were cited, a 3.1% survival rate, and he describes an answer as involving reading on the order of 600 pages while crediting around 30 of them. Roughly twenty pages are read for every one that gets named. Being retrieved is not the win. Being the page that survives that cut is the win, and near-duplicates get cut first because a second version of something already read adds nothing.

ChatGPT often picks the brand before it reads a page

Before ChatGPT searches, it writes its own search query, and that query already contains brand names of its own. Brands named inside that self-generated query landed in the final answer 68.9% of the time. Brands that were merely fetched during the search, with no name in the query, landed 2.1% of the time. That is roughly a 33x gap, decided before a single page of yours is read.

Read that alongside the 68.9% number and the practical instruction writes itself. You want your name attached to one clear question in the model's head, not smeared thinly across four pages that each half-own it. That is where I would place my bet on consolidation: one URL that unambiguously answers "how much does an ADHD assessment cost in London" is a far stronger signal than four pages that each partially answer it and mutually contradict each other on price.

Position matters too, and it compounds the same problem. When several pages from one domain arrive in the same retrieved group, the first one is cited 5.2% of the time, the second 4.6%, the third 2.4%, the fifth 0.6%, and the sixth or later 0.3%. Your sixth page on the topic is not a second chance at the answer. It is a page that gets read and discarded, at a rate seventeen times worse than your first.

Deciding which page lives is a business call, not an SEO one

Picture a private ADHD clinic in London with four pages that all touch adult assessment: the service page written when the site launched, a blog post titled "What to expect from an adult ADHD assessment," a second post on assessment costs, and a landing page an agency built for a paid campaign that never got switched off. All four rank for scraps of the same query. None of them ranks properly. When a parent asks an AI assistant where to get assessed in London, the clinic gets read and skipped, because the model has four half-answers from one domain and one complete answer from a competitor. I have booked a London ADHD clinic solid for three straight months, so I know exactly how much demand sits behind that query. The four-page mess is the hypothetical. The demand is not.

Choosing the survivor is where teams stall, because they argue about which page they like rather than which page has assets. Assets are countable. Run these five checks in order and the argument ends.

One nuance saves good pages from the bin. Two pages can look like duplicates and genuinely not be, if they split cleanly on an axis a customer would recognize: which thing to do versus how to do it, choosing versus using, owner versus agency. When that split is real, keep both and link them to each other so they read as a deliberate pair. When the split only exists in a keyword tool, you have two pages fighting.

Merging done right keeps the rankings you already have

The fear that stops most cleanups is losing rankings. The answer is procedural. A merge is not a deletion, it is a transfer, and the transfer has five moving parts.

Three of those steps trip people up in practice. Harvesting means opening each losing page and lifting anything the winner does not already say, a paragraph a customer quoted back to you, an FAQ answer, a photo or diagram the winner lacks, then pasting it into the winning page where it fits the flow. To find your own links pointing at a page you are retiring, search your site for the old web address the way you would search for a phrase, or open your CMS link report, then change each link to the winner. And a 301 redirect, if you have never touched one, is a permanent change-of-address notice: anyone who asks for the old address gets sent to the new one and told the move is permanent. Your web host or a CMS redirect plugin sets one in a few minutes, one line per retired URL.

I run my own archive on exactly this rule, and it is not theoretical housekeeping. One question, one page, and where two angles deserve to live separately, they carry reciprocal links to each other so they reinforce instead of compete. A documented internal-linking pass took this site from fourteen orphan posts, pages nothing else linked to, down to zero. The pages that survived did not lose anything. They stopped competing with their own siblings for the same answer slot.

That rule is why this article and my piece on how to tell when a page is decaying and whether to refresh or prune it are a deliberate pair rather than rivals. That one is about time: a page that used to work and is fading, and what to do about it. This one is about intent: several pages alive at once, all chasing one question, and which of them should own it. Different trigger, different decision, no overlap. Once you have picked the survivor, the next job is making that page quotable, which is a question of how to structure a page so AI systems can lift a clean answer from it.

Measure the question, not the page count

After a merge, traffic to the surviving URL will wobble while Google reprocesses redirects and reconsiders which page to show. Judging the cleanup in week two is how good work gets reversed. Set a light check at four weeks to confirm the redirects fire and nothing 404s, then judge the outcome at twelve weeks on the numbers that pay you.

The right scoreboard is the question, not the URL. Add up clicks and inquiries across all the old pages before the merge, compare that total to the survivor after, and you have an honest answer. Compare the survivor against one of the four old pages and you have a story that flatters or panics you at random. Then go ask the assistants themselves: type the customer question into ChatGPT the way a customer would phrase it, and note whether your page appears and which one. That five-minute check tells you more than a month of impression charts.

One warning about the cleanup itself. Merging is not an excuse to gut the archive. Pages that answer different questions are not duplicates just because they sit in the same category, and deleting them shrinks the number of distinct questions you can win. The study's best observed rate came from domains with two matched pages, and every question you delete outright is a question you can no longer be cited for. Precision is the goal. Emptiness is not.

Frequently Asked Questions

Does publishing more blog posts still help SEO?

More posts help only when each one answers a question none of your other pages already answer. Once two or more of your pages chase the same question, they compete instead of compounding, and in Suganthan Mohanadasan's July 2026 ChatGPT retrieval data the citation rate falls from about 6.2% at two tightly matched pages to about 1.7% at six or more. Volume is not the input that pays. Coverage of distinct questions is. Before you brief the next post, check whether an existing page already owns that question, and if it does, strengthen that page instead.

How many pages should I have on the same topic?

One page per question, not per topic. A topic like ADHD assessment holds many separate questions: what it costs, how long the wait is, whether the report is accepted by schools, what happens on the day. Each of those deserves its own page, and none of them deserves two. In the same dataset, the first page from a domain in a retrieved group was cited 5.2% of the time and the sixth or later 0.3%, so the extra near-duplicates are not neutral filler, they are dead weight sitting on your best page.

Will merging two pages lose my Google rankings?

Not if you merge properly. Move the unique sections of the losing page into the winner, then 301 redirect the old URL to the winner so its links and history transfer rather than disappear. Never delete a page that has backlinks or rankings without a redirect, and never redirect to your homepage, which tells Google the answer moved nowhere. Rankings wobble for a few weeks while Google reprocesses the redirect. Judge the result at twelve weeks on the combined clicks and conversions, not on the traffic of the single surviving URL in week two.

Most content plans are written as a list of things to add. Try writing next quarter's plan as a list of questions your business should own, then counting how many pages you currently point at each one. If any question has more than two, your next content investment is not a new page. If you want a second pair of eyes on which of your pages should win, book a call and we will go through your archive together. The four posts you paid for last quarter may be the reason the one good page you already had stopped getting quoted.