What Results Did Mass Publishing Produce for GEO?

This article was first published in Korean by Intrix Lab on August 26, 2026.

Publishing more did not lead to increases in both search and AI visibility. K-DermZIP's automatically generated hospital detail pages received 0 PerplexityBot crawls over 30 days. At one site in a YMYL field where we manage GEO, a period of publishing 203 articles over 25 days coincided with clicks for a fixed cohort of 488 articles falling from 44 to 19. intrix.kr also published 56 articles in 2026-07, but the indexing rate for articles 31–60 days after publication was 52%. The observations collected in the ledger as of 2026-08-26 point toward designing each page's value and structure before focusing on mass production.

What emerges when we compare the observations?

The first case is K-DermZIP. After filtering out impersonating requests from bot logs for 2026-07-11–2026-08-10, we found 0 PerplexityBot crawls of Korean-language hospital detail pages and 0% coverage. During the same period, it crawled the same site's blog 146 times and procedure guides 82 times. It did not skip the entire site; its activity differed by page type.

Other bots read the same detail pages. ClaudeBot coverage was 93.3%, OAI-SearchBot coverage was 66.6%, and GPTBot coverage was 44.4%. This contrast does not permit the conclusion that all AI crawlers evaluate mass-generated pages in the same way. For PerplexityBot, however, the selection clearly differed between human-written sections and automatically generated sections produced at scale.

The second case is one site in a YMYL field where we manage GEO. After publishing 203 articles over 25 days, about 8 per day, we compared search clicks for a fixed cohort of 488 articles. Clicks fell from 44 on 2026-07-26 to 19 on 2026-08-20, a 57% decrease.

The third case is intrix.kr. In 2026-07, we published 56 articles at a rate of 2.1 per day. Among articles 31–60 days after publication, indexing was confirmed for only 24 of 46, an indexing rate of 52%. By contrast, indexing was confirmed for all 7 articles at least 61 days old. The pattern does not show that waiting longer produces the same result. It shows a lower indexing rate in the period when publishing volume was concentrated.

Can cannibalization alone explain the decline in clicks?

If duplicate keywords were the cause, clusters with several articles targeting the same keyword should have declined more, while articles with unique keywords should have held up relatively well. The actual breakdown differed. Clicks for the 158 articles in clusters with 5 or more articles sharing a keyword fell from 15 to 6, a 60% decrease. For the 84 articles in clusters with 2–4 articles sharing a keyword, clicks fell from 12 to 5, a 58% decrease. The 246 articles with unique keywords or no keywords also fell from 17 to 8, a 53% decrease.

Average position also moved in a similar direction across all groups. The group with 5 or more articles sharing a keyword moved from 25.4 to 32.3, the 2–4 article group from 26.3 to 30.7, and the unique-keyword or no-keyword group from 20.1 to 26.2. This makes it difficult to explain the overall decline as a problem with just the largest keyword cluster of 40 articles.

This breakdown does not eliminate the cannibalization hypothesis, but it weakens it as the main explanation. The group without duplicates declined similarly. The main hypothesis remaining after the breakdown is a decline in site-level evaluation. However, this is an observation that mass publishing and the decline coincided, not a causal determination that mass publishing caused the decline in evaluation.

What worked in the opposite direction?

There is also an observation in the opposite direction, where careful design rather than volume worked. One page that ChatGPT read most often for queries in this industry was read 260 times at the URL level in tracking data from 640 GPT calls. The page contained 8 JSON-LD types, 3 FAQ questions, 4 question-form H2 headings, and a first sentence that put the bottom line up front (BLUF). Unlike the cases of producing many similar documents, this content closely aligned a single page's search intent and information structure.

This observation means that a small number of carefully designed pages and volume-driven mass production should not be treated as the same strategy. The latter is what we describe as having failed in these observations. The fact that a carefully designed page was read shows the potential of the former. However, these data do not isolate and prove JSON-LD or question-based structure as the cause. It also remains possible that the title and path matched the model's rewritten search queries.

How far do these observations go?

The results for all three properties are observational, not intervention experiments. They did not rule out factors other than mass publishing, such as domain age, topic, content density, or changes in search engines over time. Figures from different populations must not be combined and interpreted as a single effect size.

Because keyword clusters were defined strictly, keywords with similar wording may have been placed in the unique-keyword group. This limitation could lead to underestimating cannibalization.

The indexing rate for intrix.kr is a proxy measured at that point in time; Search Console is the authoritative source. The K-DermZIP figures exclude impersonating requests, so they cannot be interpreted as raw request counts or the total number of hospital records.

How should publishing plans change?

These results favor defining the question a page will answer, its evidence, structure, and internal links before setting publishing volume and filling the slots. After publication, look beyond the total article count and examine crawls, clicks, and indexing separately by page type. This helps distinguish pages that were not crawled from pages that were read but not selected.

Intrix's free SEO and GEO audit checks your site's current crawlability and structure using actual responses. Before starting mass publishing, you can identify whether the bottleneck lies in access, indexing, or content structure.

Frequently asked questions (FAQ)

Q. Should we stop mass publishing based on these observations alone?
The same conclusion cannot be applied to every site. Decisions should follow separate measurements of existing article quality and search intent, publishing speed, and crawl and indexing status.

Q. Does this mean cannibalization was not the cause?
No. Breaking the data down by keyword cluster showed clicks and rankings declining in a similar direction regardless of duplication. This means cannibalization alone has difficulty explaining the overall decline.

Q. What does careful page design check first?
It checks whether the direct answer in the opening paragraph, question-form subheadings, verifiable evidence, structured data, and internal links match actual search intent. Connecting answers and evidence within a page comes before counting elements.