Ahrefs’ AI SEO Data Breaks the One-Playbook Myth

Ahrefs’ AI-search research remains useful, but the idea that it delivered one definitive formula no longer survives close inspection. The findings come from separate studies with different samples, dates and search products, while newer evidence shows that conventional search visibility still affects how pages enter at least some AI citation pipelines.

The practical answer is a portfolio strategy: preserve crawlability and strong editorial work, then measure ChatGPT, Google AI Overviews and AI Mode independently. What remains valid is the need to understand citations as a new visibility layer; what changes is the claim that this layer follows one settled set of rules.

The “definitive study” was really a research program

Several striking statistics circulated together as though they described one experiment. In fact, analyses of cited pages, brand mentions, query types, click-through rates and Google’s AI products answer different questions. Combining them can provide a broad view of the market, but it cannot turn their correlations and snapshots into universal ranking factors.

This distinction matters because a page-level citation study does not establish what makes a brand appear in an answer, and a correlation across brands does not prove that adding one channel caused the observed visibility. Results drawn from Google AI Overviews also cannot automatically be applied to ChatGPT, whose retrieval system, interfaces and source selection can differ.

Ahrefs itself supplied an important refinement after many of its earlier findings. Its April 2026 ChatGPT citation study analyzed 1.4 million ChatGPT 5.2 prompts in a desktop dataset and reported that roughly half of retrieved URLs became citations. The researchers warned that retrieval did not necessarily mean every candidate page was opened or read in full.

The more useful finding was not a magic percentage but the structure of the selection process. Citation rates differed sharply by retrieval channel, and pages surfaced through general search behaved differently from URLs supplied through dedicated Reddit, YouTube, news or academic channels. Ahrefs also found stronger title relevance among cited search results, but explicitly cautioned that aggregate comparisons became misleading when the uncited pool was dominated by Reddit data.

Traditional rankings are not irrelevant to AI visibility

The newer ChatGPT analysis complicates the earlier claim that Google visibility had become largely irrelevant. In Ahrefs’ dataset, the general search retrieval category produced most citations and had a much higher citation rate than the specialized categories. That does not mean a Google position directly determines a ChatGPT citation, but it does show why discoverability in a search selection pool can still be consequential.

There is also no contradiction between finding some cited pages with little estimated organic traffic and finding that search-based retrieval matters. Organic visibility estimates, rank for the user’s original wording and eligibility for an AI system’s internally generated subqueries are different measurements. A page may fail to rank for the initial prompt yet match one of the narrower questions generated during retrieval.

Freshness needs similar restraint. The study found that cited pages in its search category had a median age of about 500 days, while relatively established pages often beat newer candidates inside the same retrieval set. News was different: when relevance was similar, younger pages had an advantage. “Always publish newer content” is therefore too crude; relevance, retrieval channel and query type must be considered together.

Google’s two AI surfaces need separate scorecards

Even products operated by the same company do not behave as a single channel. An Ahrefs comparison of AI Mode and AI Overviews examined September 2025 US data, using 540,000 paired queries for citation analysis and 730,000 pairs for response similarity. Only 13.7% of cited URLs overlapped, although the answers had an average semantic similarity of 86%.

That contrast is more actionable than a generic instruction to “optimize for Google AI.” The two surfaces often reached broadly similar conclusions while selecting different supporting pages. A site can therefore gain visibility in one without securing the same citation in the other, and a combined success metric can conceal that difference.

The study also compared only one generation from each product for a query. Since generative answers and their sources can change between runs, the result should be read as a snapshot of observed behavior, not a permanent source-allocation rule. Measurement needs repeated observations, a stable prompt or query set and separate reporting for mentions, citations and visits.

Google still treats core SEO as the foundation

Google’s current generative-search guidance says pages must be indexed and eligible to appear with a snippet before they can be considered for its AI features. It recommends crawlable pages, a clear technical structure and original, useful content, while stating that no special schema markup or llms.txt file is required for generative search.

This official position places an important boundary around the schema debate. Structured data is not a dedicated switch for AI citations, but Google still recommends accurate markup where it supports established search features such as rich results. Removing valid structured data solely because an AI-citation experiment showed a small average effect would confuse two different purposes.

The same guidance rejects mass-producing pages for every possible fan-out query. A better interpretation of query fan-out is to cover the meaningful subquestions of a topic within coherent, useful resources. Exact-match page multiplication creates duplication and maintenance costs without guaranteeing selection by an AI system.

A defensible AI-search workflow

The evidence supports a measured operating model rather than a list of hacks. Teams can organize the work in four layers:

1. Protect eligibility. Keep important pages indexable, crawlable, technically reliable and eligible for conventional search presentation.
2. Build citable substance. Publish original evidence, clear explanations and titles that accurately describe the questions answered. Refresh time-sensitive facts without discarding useful history merely to change a date.
3. Measure each surface separately. Track brand mentions, linked citations and referral visits as different outcomes. Maintain distinct views for ChatGPT, AI Overviews and AI Mode instead of merging them into one “AI rank.”
4. Test against business results. Compare citation gains with qualified visits, subscriptions, leads or sales. Visibility without attribution may still influence discovery, but its value should not be assumed from citation counts alone.

This approach also clarifies the role of channels such as video, editorial coverage and reference sites. They may expand the contexts in which a brand is discussed, but observational correlations do not justify manufacturing mentions or treating one platform as a guaranteed ranking lever. Investment should follow audience relevance and measured outcomes.

The lasting lesson from Ahrefs’ work is not that SEO has been replaced. AI systems have added volatile, product-specific selection stages on top of familiar discovery and publishing fundamentals. The hard part is now maintaining those fundamentals while measuring several answer engines whose citations can diverge even when their conclusions agree.

Also read:

• Understanding and Implementing Structured Data Markup for SEO Benefits
• No One Talks About The Growing Importance of Data Integration
• One in Three TikToks Shown to New Users Is Low-Quality AI Slop, Study Finds