There is a number that should end the conversation, and it is this: only 10–15% of what Google ranks in its top results overlaps with the sources ChatGPT actually draws on. (source: Chen et al., University of Toronto — arXiv:2601.16858, January 2026 — corroborated across multiple 2026 cross-platform studies) Run the same search on Perplexity and on ChatGPT, then compare the sources each one pulls from: less than 1% overlap. (source: Roopesh Patel, BrandFeatured, “AI Search Ranking Factors Explained for 2026”)
If you have spent the last decade building Google rankings and you assume that translates into AI visibility, the data says otherwise. Decisively, not marginally.
This is not an edge case. It is the default condition of AI search in 2026. And most businesses have not yet reckoned with what it means.
The Assumption That Is Failing
The implicit logic goes: we rank highly, therefore we are a credible source, therefore AI will recommend us. It is a reasonable inference. It is wrong.
A Semrush study of 5,000 queries — the clearest cross-platform comparison currently available — produces an overlap hierarchy that maps directly to architecture, not content quality. Perplexity overlaps with organic top-10 domains at around 91%, because it retrieves heavily from the live web. Google AI Overviews overlap at around 86%, because they draw from Google’s own index. Google AI Mode drops to 51%. ChatGPT is the weakest of all platforms tested, with no comparable domain overlap figure published because the divergence is so pronounced as to make the comparison largely meaningless. (source: Semrush 5,000-query study, reported by Local Falcon, April 2026)
The academic data is even starker. A peer-reviewed University of Toronto study (Chen et al., arXiv:2601.16858, January 2026) measured how much each AI system’s chosen domains overlapped with Google’s top-10 results. GPT-4o shared just 4.0% — one in twenty-five domains in common. (source: Chen et al., University of Toronto, arXiv:2601.16858) Four percent is not a gap. It is a different world.
And it gets worse for local businesses. Local Falcon’s analysis of 190,000 ChatGPT restaurant results found that 83% of restaurants are completely invisible on ChatGPT — compared to 14% invisible on Google Search. Years of local SEO investment in Google Business Profiles, directory listings, and geographic ranking signals have produced almost no transfer to ChatGPT AI search presence. (source: Local Falcon, “How Much Overlap Is There Between AI Search and Traditional Search?”, April 2026)
What AI Is Actually Selecting For
The question, then, is what does predict whether AI recommends you, if not Google rank?
The evidence points to a coherent but unfamiliar set of signals. They are not the signals that SEO tools measure, and they are not what link-building campaigns are designed to produce.
Entity authority. AI engines weight how well-documented and consistently recognised your brand is as an entity across the web — Wikipedia entries, Wikidata records, consistent name/address/phone details across directories, persistent mentions in publications AI already trusts. In campaign data across 450+ clients, entity signals were the determining factor in 73% of direct comparisons. Only 3.3% of new brands appear in ChatGPT product discovery versus 99% recall for established brands. (source: OverTheTop SEO analysis of 450+ campaigns; GetMint) The gap is not subtle.
Earned media, not owned content. The arXiv study found that Claude draws on earned media — independent press and reviews — 65% of the time; GPT-4o, 57%. For consideration queries — “best of”, “compare X vs Y” — the earned-media share rises to 59–86% across AI platforms, versus 41% for Google. (source: Chen et al., University of Toronto, arXiv:2601.16858) Brand mentions correlate 3x more strongly with AI recommendations than backlinks do. The ConvertMate study of 12,500+ queries found that brands are named 6.5x more often via third-party sources than via their own domain. (source: GetMint; ConvertMate GEO benchmark, 12,500+ queries) Publishing more content on your own website is the wrong lever.
Answer structure, not depth for its own sake. Forty-four percent of what ChatGPT draws on comes from the first third of a page. Content that leads with its core claim gets picked up 2.3x more often than equivalent content with a traditional essay structure that builds toward a conclusion. Question-and-answer format text is twice as likely to be used as conventional paragraphs. The engine is extracting claims, not reading articles. (source: Kevin Indig analysis of 1.2M queries; OverTheTop SEO 450+ campaigns)
Freshness. AI engines draw on content that is, on average, 40–70% newer than what Google surfaces. The arXiv data gives Claude a median source age of 62 days in consumer electronics; Google’s equivalent is 130 days. (source: Chen et al., University of Toronto, arXiv:2601.16858) Coverage that is older than three to six months faces a structural disadvantage, regardless of its quality or how well it ranks in organic search.
Community presence. Reddit accounts for 46.7% of Perplexity’s most-used sources. ChatGPT pulls Reddit at enormous scale to understand topics — then names more institutional sources when formulating its response. Community discussion shapes what the model knows, even when it never appears as a named source. (source: Ahrefs study of 1.4M prompts; Perplexity citation analysis) Brands with no community presence are, at the knowledge level, less visible to AI than their organic rankings suggest.
The Pre-Training Factor: Why SMEs Have an Unexpected Advantage
There is a further complexity that the headline overlap numbers do not fully capture. How AI systems use their signals depends on whether they already know who you are.
For well-known brands, AI rankings are governed primarily by pre-training knowledge — the patterns baked in from training data. When University of Toronto researchers aggressively scrambled the retrieved evidence for popular entity queries, the rankings barely moved. The model already had a view; the retrieved evidence merely confirmed it. (source: Chen et al., University of Toronto, arXiv:2601.16858) For those brands, getting into training data — through Wikipedia, sustained authoritative publication, community discussion at scale — is the only lever that materially moves AI rankings. Content publication is largely irrelevant.
For niche and mid-sized brands, the opposite is true. The model has no stable internal hierarchy for them, so it follows wherever the retrieved evidence leads. When researchers applied the same evidence-scrambling to niche entity queries, the rankings shifted nearly twice as much. (source: Chen et al., University of Toronto, arXiv:2601.16858) Fresh earned coverage in trusted sources, well-structured content, consistent entity signals — these move the needle in ways that are simply not possible for a brand the model already has strong opinions about.
This is the structural argument for why AI search optimisation is not a large-enterprise game. The brands with the most Google authority have the least ability to shift their AI standing through content strategy. The brands with the most to gain from deliberate AI visibility work are exactly the mid-sized businesses that have been told their Google performance is good enough.
Two Channels, Two Strategies
None of this means Google rank is irrelevant. Eighty-eight percent of the pages ChatGPT draws on come from the search channel — pages that Google has indexed and can retrieve. Being in Google’s index is a prerequisite for being named on most platforms. (source: Ahrefs study of 1.4M prompts) The error is not investing in SEO. The error is stopping there.
What the data demands is a second, parallel strategy — one built around the signals that AI systems actually weight: entity documentation, earned third-party coverage, structured extractable content, freshness, and community presence. These are the things that determine whether an indexed, rankable page gets named in an AI answer or quietly discarded.
The two strategies share some foundations. Technical SEO, crawlability, schema markup — all prerequisite for both. But beyond that shared base, they diverge sharply. A backlink campaign moves Google rankings and does almost nothing for entity authority. A Wikipedia entry and Wikidata record do the opposite. Brand mentions in category media count 3x more than backlinks for getting recommended by AI. That inversion matters.
Google rank tells you where you sit in a list. Being recommended by AI tells you whether you’re worth including in an answer. They are different questions, with different answers, driven by different signals. The 10–15% overlap figure is not a calibration problem or a temporary quirk of early AI search. It is the signal that two fundamentally distinct visibility systems are now operating simultaneously — and assuming one covers the other will leave most businesses invisible on at least one of them.
Build for both. Understand that for AI, the question is never where you rank. It is whether you are known.
Sources
- Chen, M., Wang, X., Chen, K., & Koudas, N. (University of Toronto). “Navigating the Shift: A Comparative Analysis of Web Search and Generative AI Response Generation.” arXiv:2601.16858, January 2026.
- Local Falcon. “How Much Overlap Is There Between AI Search and Traditional Search?” April 2026 — reporting a Semrush 5,000-query study, Local Falcon’s own 190,000-result restaurant dataset, and the SOCi 2026 Local Visibility Index.
- Patel, Roopesh (BrandFeatured). “AI Search Ranking Factors Explained for 2026 — What Drives AI Citations.”
- OverTheTop SEO. “AI Search Ranking Factors: What We Know From Testing 450+ Campaigns.”
- GetMint. AI search ranking-factor research (new-vs-established brand recall; brand mentions vs backlinks).
- ConvertMate. GEO benchmark, 12,500+ queries across 8,000 domains, 2026.
- Indig, Kevin. Analysis of where within a page ChatGPT draws its quotes (1.2M queries; reported via Marketing4eCommerce).
- Ahrefs. “Why ChatGPT Cites One Page Over Another” — study of 1.4M prompts.
Leave a Reply