AI Search Visibility Statistics 2026: The Engines Disagree on Sources and Agree on Brands

TL;DR:

The AI search visibility statistics published through 2026 point at one conclusion most brands measure past. Across five engines, BrightEdge found source agreement ranging from 16% to 59% while brand agreement stayed in a tighter 36% to 55% band, and Ahrefs reached the same conclusion from 75,000 brands with engine-pair brand correlations of 0.749 to 0.821. The engines pull from very different parts of the web and still name largely the same companies. At the same time the link between Google rankings and AI citations has weakened sharply, with Ahrefs' analysis of 863,000 SERPs putting AI Overview citations from top-10 pages at 37.9%, down from 76.1% in its earlier study.

The most common reaction to an AI visibility report is defensiveness

When we show a founder that their company does not appear in AI answers for their own category, the first response is almost never curiosity. It is a printout of their rankings.

They are not wrong about the rankings. They rank for the keywords they paid to rank for, often in the top three. They are also invisible in AI, and both things are true at once because ranking and citation are now two different competitions with different rules.

The published generative engine optimization statistics explain why, and the gap is wider than most people running SEO programmes have priced in.

AI search visibility is a brand measurement, not a page measurement

AI search visibility is the rate at which a brand appears in AI-generated answers, either as a named mention in the response text or as a linked citation. It is not a ranking position, because AI engines synthesise one answer instead of returning a list.

Two numbers get confused constantly. Citation share measures which domains an engine draws source material from. Brand share of voice measures which companies the engine actually names. A brand can be recommended in an answer assembled entirely from sources it does not own.

That distinction decides what is worth tracking. Domain citations are cheap to measure, which is why most tools measure them. Whether the engine says your name is what determines if you make a buyer's shortlist.

Engines disagree about sources and agree about brands

BrightEdge analysed citations and brand mentions across ChatGPT, Perplexity, Gemini, Google AI Mode and Google AI Overviews in April 2026, spanning ten industries including B2B technology, finance, healthcare, insurance and travel.

Pairwise top-100 overlap in cited sources ranged from 16% to 59%, a 43-point spread. Pairwise top-100 overlap in named brands ranged from 36% to 55%, a 19-point spread. In every pairwise comparison, brand agreement fell in a tighter band than source agreement.

Ahrefs arrived at the same place by a different method. Its December 2025 study of 75,000 brands put brand-mention correlation between engine pairs at 0.821 for AI Overviews and Google AI Mode, 0.769 for AI Mode and ChatGPT, and 0.749 for AI Overviews and ChatGPT.

AI engines are far more likely to disagree about where to look than about who to recommend.

We see this on live accounts. For a brokerage client, ChatGPT leaned heavily on Reddit threads to answer category questions, while Perplexity and Claude pulled from the company's own website for the same questions. Same brand, same category, three engines reading three different parts of the internet to reach a similar answer.

That is the practical shape of the finding. If you audit one engine and treat the result as your AI visibility, you have measured one of several unrelated retrieval paths.

The sourcing differences underneath are severe. BrightEdge's breakdown shows roughly a 90x spread in how much user-generated content each engine relies on.

Source: BrightEdge AI Catalyst, April 2026. BrightEdge does not publish a sample size for this analysis.

Note that the two Google search surfaces share roughly 59% of their top-cited sources, while Gemini, also a Google product, overlaps with Google AI Mode at only 27%. Treating "Google AI" as one system misreads three different sourcing behaviours.

Ranking in Google no longer predicts getting cited

This is the section to put in front of the founder holding the rankings printout.

Ahrefs analysed 863,000 keyword SERPs and 4 million AI Overview URLs, publishing in March 2026. It found 37.9% of AI Overview citations came from pages appearing in the first ten result blocks, with the remainder split almost evenly between positions 11 to 100 at 31.2% and outside the top 100 at 31.0%.

The comparison point is what stings. Ahrefs' earlier study of 1.9 million citations put the same figure at 76.1%. The company also improved its citation parsing between the two studies and does not quantify how much of the drop that accounts for, so read the direction rather than the exact delta.

For AI assistants rather than AI Overviews, the overlap was thin from the start. Across 15,000 long-tail prompts tested in July 2025, an average of 11.9% of assistant citations ranked in Google's top 10 for the original prompt. Perplexity was the outlier at 28.6%. ChatGPT, Gemini and Copilot sat near 8%, and roughly 80% of their citations came from pages ranking nowhere in Google for that query.

Semrush found the same thing from the opposite direction in July 2025, reporting that ChatGPT cites pages ranking in organic position 21 or lower almost 90% of the time. That study was scoped to digital marketing and SEO topics, so it describes one industry's results rather than the web generally.

Ranking well and being cited are correlated activities, not the same activity. A brand can own the top three positions for its money keywords and still be absent from every answer a buyer sees.

Three platforms absorb most of the citation supply

Citation share is heavily concentrated, and the concentration sits on platforms most brands do not own.

In Ahrefs' July 2026 snapshot covering more than 3.1 million US queries, YouTube held 31.2% mention share among Perplexity's fifty most-cited domains, Reddit 13.9% and English Wikipedia 7.2%. No other domain exceeded 3.3%. Mention share here is calculated against the top fifty sources rather than all citations, so those percentages overstate each domain's share of the full pool.

YouTube's position in Google AI Overviews is more striking. Among AI Overview cited pages that did not rank in Google's top 100 for the same keyword, 18.2% were YouTube URLs, and YouTube accounted for 5.6% of all AI Overview URLs in the dataset.

The picture also moves fast enough that every figure needs a date on it. Semrush's July 2025 study named Quora the most-cited domain in Google AI Overviews with Reddit second. Ahrefs' March 2026 analysis names YouTube. The most likely explanation is that the AI Overview source mix genuinely shifted when Google moved AI Overviews to Gemini 3 in January 2026.

What correlates with AI visibility, and what does not

Ahrefs tested search and brand metrics against AI visibility across 75,000 brands using Spearman correlation. The ordering reshuffles the standard SEO priority list.

• YouTube mentions correlated most strongly at approximately 0.737, measured as brand appearances in video titles, transcripts and descriptions. Mention volume mattered slightly more than reach, with view-weighted impressions at roughly 0.717.

• Branded web mentions held between 0.66 and 0.71 across ChatGPT, Google AI Mode and AI Overviews. Being discussed in many places predicted visibility better than domain strength did.

• Domain Rating peaked at 0.326 on AI Overviews, the highest of the three surfaces and still well below the branded signals.

• Number of site pages correlated at approximately 0.194, effectively no relationship. Publishing volume for its own sake did not predict AI visibility here.

• ChatGPT showed the weakest correlations with traditional authority metrics, including branded search volume at 0.352, which makes it the least gated entry point for brands without established recognition.

Our own view on the YouTube finding is that it is real and that most agencies quietly ignore it because acting on it is expensive. Video is a different production line from written content, with different people, different turnaround and different cost. We run it for clients who ask for it rather than bolting it onto every retainer, and anyone telling you it is a quick win has not built the pipeline.

Two caveats belong on the correlation data. Ahrefs states the correlation-is-not-causation disclaimer itself. Its sample was built from domains with Domain Rating above 40 whose top keyword had at least 800 monthly searches, which skews toward established brands and plausibly inflates the brand-signal correlations.

AI answers change every two days, which breaks most reporting

Any AI visibility figure taken from a single run is close to meaningless, and there is now data on exactly how meaningless.

Ahrefs tracked 43,000 keywords, each with at least sixteen recorded AI Overviews, over one month. Content changed between consecutive observations 70% of the time, with average persistence of 2.15 days. Only 54.5% of cited URLs carried over, meaning 45.5% of sources were new each time. Entity overlap sat at 54%.

The underlying meaning barely moved. Semantic similarity between consecutive responses averaged 0.95 cosine, where 1.0 is a perfect match. The engines rephrase a stable answer using rotating sources rather than changing their opinion.

Ahrefs notes its checks were not daily, so the true change rate is likely higher and 2.15 days is an upper bound on stability. Search volume showed no relationship to change rate at -0.014, which rules out the theory that popular queries are cached.

A single snapshot of AI visibility measures the weather, not the climate. Anything reported to a client needs repeated dated runs against a fixed prompt set, aggregated across many prompts rather than tracked one prompt at a time.

What these statistics cannot tell you

Every figure above carries four limitations, and any strategy built on them should account for all four.

1. No dataset covers every engine. The published studies span three to five surfaces. Claude and Grok appear in none of them, which is notable given that in our own client work Claude behaved differently from both ChatGPT and Perplexity on identical questions.

2. No dataset covers professional services or provider-selection prompts. All of this is consumer and broad commercial data. Nobody has published citation behaviour for prompts like "best AI SEO agency for a B2B SaaS company," which is the question that decides shortlists.

3. All of it is US-scoped. Ahrefs' Perplexity snapshot states US-only explicitly, and there is no published UK or Australia equivalent.

4. All of it is vendor-produced. Every study cited here comes from a company selling AI visibility software. The methodologies are disclosed and the numbers are usable, but no independent benchmark exists in this category yet.

The standard that would close those gaps is not complicated: a frozen prompt set that does not change between runs, dated and archived snapshots so figures can be compared over time, coverage of every engine a buyer actually uses, and scoping to a category rather than the open web. Until that exists for professional services, brands in this category are making decisions from consumer data.

Talk to us about your AI visibility →

Read more about what we do.

^{[1][2][3][4][5][6][7]}