Growtika/Unsplash
A study of 2.4 million AI citation records published today by Somantra AI found that 57.2% of websites referenced by ChatGPT and Google appear in exactly one month of results and are never surfaced again — and that the content format most commonly blamed for the disappearance is the one content teams continue to fund the most heavily: the long-form “complete guide.”
Somantra AI, a Sydney-based answer engine optimization platform, tracked 2,437,107 citation records across 28,725 domains on ChatGPT and Google AI Overviews over seven months (November 2025 through July 2026) in the Australian insurance sector. The study defines the zone between single-citation domains and persistent ones as the “content graveyard” — and its most consequential finding is what separates the two groups. It is not brand size. It is format.
Pages built around comparison tables, FAQ structures, and discount or pricing language were cited persistently at roughly twice the rate of one-citation domains across the observation window. Government explainers and comparison sites dominated the list of domains cited in all seven months. Meanwhile, content labeled a complete guide was 3.5 times more common among domains that vanished after a single citation than among those cited across the full seven months.
“If a content team is still producing long-form guides because that is what won on Google in 2020, the data suggests those pages are being cited once and then left behind,” said Arun Prasad, Somantra’s founder. “Discount pages, comparison pages and FAQs are the formats holding their citations. Brands that make that switch keep their visibility. Brands that do not lose it one citation cycle at a time.”
Format Beats Authority in AI Citation Survival
The finding inverts a central assumption of the SEO era: that domain authority and brand recognition are the dominant signals of content ranking. In Somantra’s dataset, the domains that held citations across all seven months were not necessarily the biggest or most recognized names in Australian insurance. Only 2.7% of the 28,725 tracked domains — approximately 779 sites — earned citations in every month of the study.
The persistence advantage of structured formats was consistent enough that Somantra argues smaller brands have a realistic route to durable AI citation without first matching competitor domain authority. Because survival tracks format rather than size, the study claims the approach is reproducible by any brand willing to change what it publishes.
A broader, independently run study published last month by DeltaV Digital — tracking 25,337 citations across ChatGPT, Perplexity, Gemini, Google AI Overviews, and Google AI Mode over 90 days and spanning eight industries — reached a finding consistent with Somantra’s on the comparison-page question. Comparison pages earned only 4.1% of total citations across DeltaV’s dataset, but posted the highest citation rate of any page type at 1.87 citations per retrieval — 45% above the portfolio average of 1.29. When an AI engine retrieved a comparison page, it relied on it heavily.
DeltaV’s data adds an important caveat, however. Across the eight industries tracked, no single format dominated universally. Listicles captured 61% of citations in B2B technology services; homepages captured 55% for a local services brand; program pages captured 53% in higher education. The practical implication for content teams is not “build comparison pages instead of guides” as a universal rule, but rather to understand the citation fingerprint of their specific industry before reallocating budget.
Why RAG Rewards Short, Specific Content
The mechanism behind the format finding lies in how modern AI search engines are built. Most major platforms — including ChatGPT’s search functionality and Google AI Overviews — use a technique called retrieval-augmented generation, or RAG. Rather than generating answers solely from training data, RAG systems retrieve specific text passages from indexed web pages at the moment a query is received, then synthesize those passages into a response and cite the sources they used.
The critical step is passage retrieval. RAG systems break web content into chunks, embed those chunks as vectors, and pull the most semantically relevant chunk in response to any given query. A comparison table, a FAQ block, or a pricing page concentrates its relevance in a single, self-contained passage: one chunk answers one discrete question completely. An AI engine can retrieve that chunk cleanly, cite it, and return to it whenever the same type of query appears again.
A 12,000-word comprehensive guide works differently. Its relevance is dispersed across dozens of sections; no single chunk is the best answer to any specific query. An AI engine may retrieve the guide once when a broad query broadly matches it, cite it once, and then replace it with more precisely structured sources as its index improves. The guide’s comprehensiveness — the quality that made it valuable in traditional SEO — works against it under the RAG retrieval model.
This is the structural reason the Somantra finding is unlikely to be an artifact of the insurance vertical specifically. RAG architecture is consistent across ChatGPT, Google AI Overviews, and Perplexity. A content format that produces low-density, diffuse chunks in one sector produces low-density chunks in every sector. The citation-mortality pattern documented in Australian insurance is a consequence of how AI search systems process any content, not of how insurance queries happen to work.
Spam Domains and the Quality Problem AI Search Hasn’t Solved
The same underlying dataset surfaced a separate and more troubling finding that Somantra published in July. Of the 28,725 domains in the citation record, 38 were identified as confirmed spam or gray-area “parasite SEO” operators — sites constructed to simulate legitimate financial guidance while generating revenue through AI citation, none of which held an Australian financial services license. One flagged domain alone was cited 5,366 times in ChatGPT’s Australian insurance responses, briefly ranking as the 13th most-cited source in the category.
The disparity between platforms was pronounced: gray-area and spam domains accounted for 1.97% of ChatGPT citations in the dataset, compared with 0.1% of Google citations — a 19-fold difference. The pattern suggests ChatGPT’s citation mechanisms are substantially more susceptible to format-optimized content, regardless of its accuracy or regulatory standing, than Google AI Overviews.
For brands and content teams, this finding carries an edge. The same format signals that help legitimate content survive in AI citation — concise structure, direct answers, comparison formatting — also make it easier for low-quality and fraudulent content to hold AI visibility. The formats that win in AI search do not distinguish between authoritative sources and well-structured spam.
A New Metric to Replace Traffic
Somantra’s practical recommendation centers on a shift in how content performance is measured. The company proposes “content survival rate” — the share of a domain’s cited pages that remain in AI citations after three or more months — as a standing metric in content reporting, separate from and in addition to traffic, keyword rankings, and output volume. Somantra defines this metric as a way to distinguish pages that earn one-time citations from pages that hold consistent placement over time.
The argument is that traffic figures increasingly obscure whether a brand is actually maintaining AI visibility. A page can generate traffic from traditional search while being cited once and abandoned by AI engines. Only a longitudinal citation metric captures whether a brand’s content is being treated as a durable source or as a novelty.
“Marketing teams are being measured on what they published this quarter,” said Prasad. “The question that matters now is what is still being cited two quarters later.”
The DeltaV study’s findings on citation rate versus citation share support a similar distinction. Comparison pages, which earned only 4.1% of total citations, posted a citation rate of 1.87 per retrieval — meaning the pages AI engines did retrieve were trusted heavily. Volume and rate are different signals, and current content analytics capture primarily volume.
Scope and Caveats
The Somantra study was conducted entirely within one vertical — Australian insurance — and covers a seven-month window. Somantra acknowledges the findings establish correlation rather than causation, and that the pattern is consistent with how conversational engines assemble answers, not a proof that switching formats will mechanistically improve any individual domain’s citation longevity.
Disclosure: Somantra is both the author of the study and a commercial provider of AEO and GEO services and brand audit tools. The study’s conclusions — that brands should track citation survival and shift budget from long-form guides toward structured formats — are directly aligned with Somantra’s commercial offering. Readers should weigh the findings accordingly, alongside the independent DeltaV Digital dataset and the growing body of practitioner research that broadly corroborates the format-advantage finding, if not the specific magnitude of the guide-mortality effect.
The Australian insurance sector is a useful proxy for high-consideration, competitive-research markets: dense competition, high consumer research intent, a mix of large institutional players and comparison intermediaries. But markets where AI citations concentrate in fundamentally different formats — B2B services, local services, higher education — may not see the same guide-versus-FAQ divergence in citation survival.
The full AI Search Ranking Factors for Australian Insurance study, including source-type breakdowns and slug-level citation data, is available at somantra.ai.
Frequently Asked QuestionsWhy do long-form guides fail in AI search while FAQs and comparison pages survive?
The mechanism is retrieval-augmented generation, the architecture most major AI search platforms use. RAG systems retrieve specific content passages — chunks — from web pages and synthesize them into answers. A FAQ block or comparison table concentrates its relevance in a self-contained chunk that cleanly answers a discrete query, so the engine can cite it repeatedly for similar questions. A long-form guide disperses its value across dozens of sections; no single chunk is the best answer to any specific query. The guide may be retrieved once when it broadly matches a broad query, but more precisely structured sources progressively replace it as the index improves. The Somantra study documents this pattern empirically; the RAG mechanism explains why it holds structurally, not just within one sector.
Does the citation-survival finding apply only to insurance, or to other industries?
Somantra’s data comes from one vertical over seven months, and the company acknowledges its findings describe correlation in the Australian insurance market, not a universal law. However, an independent 2026 study by DeltaV Digital — tracking 25,337 citations across eight industries and five AI platforms — found that comparison pages posted the highest citation rate of any page type at 1.87 citations per retrieval, 45% above the portfolio average. That broad consistency suggests the format advantage is not insurance-specific, even if the exact magnitude varies by sector. DeltaV’s data also found that every industry has a distinct citation “fingerprint” — the dominant format differs — so the practical question for any content team is to identify the format that works in their specific category, not to assume comparison pages universally win.
Should content teams stop producing long-form guides entirely?
Not necessarily, but the data supports redirecting at least some of the budget that goes into comprehensive guides toward comparison pages, FAQ sections, and structured pricing or discount content — particularly for brands whose primary growth channel is AI-mediated discovery. Long-form guides still serve functions: they support traditional SEO, demonstrate topical depth, and build E-E-A-T signals that influence AI engine trust over time. The Somantra study’s practical implication is about resource allocation at the margin, not complete elimination of a content type. The more actionable shift may be measuring survival rate alongside output volume: a content team that tracks how many of its pages are still being cited three months after publication has a feedback signal that traffic volume alone cannot provide.
What is “content survival rate” and how is it measured?
Content survival rate is a metric proposed by Somantra as the share of a brand’s AI-cited pages that remain in citation records after three or more months. It is designed to distinguish between pages that earn a one-time citation — what Somantra calls the “content graveyard” — and pages that hold consistent placement in AI-generated answers over time. Measuring it requires AI citation tracking tools that monitor brand mentions and source citations across platforms like ChatGPT, Google AI Overviews, and Perplexity at regular intervals, rather than relying on web traffic as a proxy for AI visibility.