AI Search Statistics 2026, the Cited Library

A running, primary-sourced library of AI search statistics for 2026. Every figure links to its original study or document, with the caveats intact.

Abstract violet and magenta light filaments on a dark background, the AxiomAI brand motif

TopicData & Benchmarks

This is a running library of AI search statistics for 2026, every one traced to its primary source. We update it as new studies land. Use it to settle arguments, brief a board, or check whether a claim you read somewhere actually holds up. Where studies disagree, we show both numbers and explain why, because a stat without its caveat is just a rumour with a decimal point.

We have watched a lot of statistics get laundered. A vendor publishes a number, a blog quotes the blog that quoted the vendor, and three months later it is a fact everyone repeats and nobody can source. This page exists to break that chain for AI search. Every figure below links to the study or document it came from. If we cannot find the primary source, it does not go on the page.

How much AI is in the search results now

AI answers now appear above the ordinary results, and AI writing has spread through the results underneath them. Originality.ai's ongoing study of 500 keywords found that 17.31 percent of top-20 Google results showed signs of AI-generated content, a share that dipped around the March 2024 core update and then climbed again. The way Google builds those answers has changed too. Its AI features use query fan-out, running multiple related sub-searches to build a single response, so one answer can draw on several pages rather than a single ranked result, according to Google Search Central.

Does structured data help you get cited by AI

The honest answer is that the evidence conflicts. Anyone giving you a confident yes or no has picked one study and ignored the rest. Google states plainly that no special structured data is required to appear in AI Overviews or AI Mode, per Google Search Central. Ahrefs then tested it directly, tracking 1,885 pages that added JSON-LD between August 2025 and March 2026 against 4,000 matched controls, and found schema produced no measurable lift in AI citations for pages that were already heavily cited. Their own caveat matters as much as the finding. The treated pages already had strong citation rates, so the null result does not necessarily apply to an unknown site starting from zero.

Other studies point the other way, so read them together before you settle on one. The table below lays the main ones side by side.

Source Evidence Headline finding
Ahrefs 1,885 pages adding JSON-LD vs 4,000 controls No lift for already-cited pages, but cited pages are around three times more likely to carry schema
AirOps and Kevin Indig 16,851 queries, 353,799 pages Pages with JSON-LD showed a 38.5 percent citation rate
Otterly Seven-platform experiment, Dec 2025 to Mar 2026 Only Gemini could read raw schema, and only when asked to look at it directly. Information placed only in schema never surfaced in a normal answer
Microsoft, on record Statement by Fabrice Canel at SMX Munich Schema helps Bing and Copilot understand pages for their LLMs

Two things reconcile the mess. First, correlation is not cause. The sites that get cited are also the sites that bother with schema, a point Ahrefs made itself. Second, Otterly ran the cleanest experiment on this, planting unique information only in schema and never in visible text, and no AI platform ever surfaced it. In practice, schema is worth doing for search and comprehension, but it is not a hidden channel that feeds AI facts your readers cannot see. Our guide to schema and entity SEO for AI search goes into more detail.

How strong is your own AI marketing? Take the free two minute scorecard and get an honest read plus the next steps that matter most.

Take the scorecard

Can AI-written content actually rank on Google

Yes, and the largest datasets say the production method barely matters. Ahrefs analysed 600,000 top-20 pages and found 4.6 percent were pure AI, 13.5 percent pure human and 81.9 percent a mix, with effectively zero correlation between AI-content share and ranking position. Semrush's separate look at 42,000 blog posts found human-classified content tended to hold slightly stronger positions. Those two findings are closer than they look. A slight edge for human-classified posts in the smaller dataset is still a small effect, and the much larger sample finds no relationship at all. Both studies also rest on probabilistic AI detectors that produce false positives, so treat the exact percentages as directional.

Google's own guidance carries more weight than either study. It focuses on content quality rather than how the content was produced. Appropriate AI use is permitted, and AI confers no ranking advantage on its own, per the Google Search Central Blog. What Google does police is scaled content abuse, and its spam policy explicitly names generative AI page generation without added value as an example, per Google Search Central. The rule targets the behaviour, not the tool.

What the platforms themselves say you must do

Less than most GEO guides claim, which is worth knowing before you spend money on any of it. Google's generative-AI optimisation guide reiterates that structured data is optional for AI search, and it confirms directly that llms.txt files are not used by Google Search, which those same guides still get wrong, per Google Search Central. For AI-generated images, Google now asks for IPTC metadata and separate labelling, per its generative-AI content guidance. For limiting what AI features show from your pages, Google still points to the standard snippet controls.

Where the standards are heading

The specs underneath all of this keep changing, so dates matter. The schema.org vocabulary reached version 30.0 on 19 March 2026, adding equivalence annotations and EU Digital Markets Act terms, per schema.org. The llms.txt convention was authored by Jeremy Howard and published on 3 September 2024, and its only required element is a single H1 with the site name, per the official spec. Our llms.txt definitive guide tracks whether publishing one does anything yet.

How to use these numbers without getting burned

  • Quote the primary source, not the blog that quoted it.
  • Carry the caveat with the number. A detector-based study is directional, a vendor study measuring its own platform has a stake in the result, and a controlled study still has a scope limit. All three appear above.
  • When two studies disagree, ask what each actually measured. Ahrefs and AirOps reach different schema conclusions because one tested lift on already-cited pages and the other measured citation rates across a broad sample.
  • Go to the platform's own documentation. Google, Microsoft and schema.org publish their own rules, and those beat any agency claim about them.
  • Check the date. AI search guidance from early 2025 can already be wrong. Where a source carries a date, we keep it with the figure.

If you want to know where your own site stands, run the free AI marketing scorecard for a plain-English read, or the deeper entity audit for how AI systems see the entities on your site.

Where do these AI search statistics come from

Every figure links to its primary source, whether that is a platform document, a named research study or an official spec. We do not include a number we cannot trace to the original, and where a study relies on an AI detector or measures a vendor's own product, we say so next to the figure.

Does schema markup get you cited by AI

The evidence is mixed. Ahrefs found no citation lift from adding schema to pages that were already cited, while other studies report higher citation rates for pages that carry schema. In the cleanest test, no platform picked up information that existed only in the markup. Mark your pages up for search and comprehension, and put anything you want an AI to repeat in the visible text.

Will Google penalise AI-written content

No, not for being AI-written. Google judges pages on quality and permits appropriate AI use. What it penalises is scaled content abuse, mass low-value pages produced to manipulate rankings, whether a human or a machine made them.

Does llms.txt help with Google

Not with Google Search. Google states directly in its generative-AI optimisation guide that it does not use llms.txt files. Some other tools may read it, and the cost of publishing one is small, but anyone selling llms.txt as a way to rank higher in Google is ahead of the evidence.

How often is this page updated

We add new studies as they are published, and revise figures when better data lands. AI search moves fast enough that a statistic from early 2025 can already be out of date, so we update or remove figures rather than leave them standing because they were once true.

Put this to work on your business

We build exactly this, SEO, AI visibility, PPC, email and automation, as one measured system. Talk it through with a real person, no obligation.

Book a free strategy call or take the two minute scorecard