As of September 2026, AI-search research has an evidence problem, not primarily a content problem: a handful of genuinely rigorous studies (Ahrefs, the Princeton GEO paper) carry most of the field's reliable findings, while most circulating statistics are vendor-reported from private corpora or, in several documented cases, simply don't reconcile with each other. This snapshot inventories what's solid, what's genuinely open, and what this site has built toward closing that gap so far.
- Ranking no longer strongly predicts citation. Overlap between AI citations and Google top-10 results runs from about 12% (ChatGPT/Gemini/Copilot grouped) to about 38% for AI Overviews, down from about 76% a year earlier (Ahrefs).
- Engines barely agree with each other. Reported ChatGPT/Perplexity domain overlap sits around 11%, so a strategy tuned for one engine may not transfer to the next.
- Schema and llms.txt both have real evidence and no clear effect. Observational data shows no lift for either; live-fetch tests find schema often ignored entirely.
- Three separate studies collapsed into one widely-quoted "58%" figure — the clearest documented case of statistic conflation in the field.
- Zero randomized causal tests of a GEO tactic existed in public before this year's pre-registrations. Every published finding prior to that is correlational.
- The five open questions: what Claude actually cites; how many sub-queries AI Mode generates; whether anything a site controls causally changes citation rate; how long a citation lasts; and whether AI crawlers render JavaScript.
- This is a dated snapshot, not a standing description. Every figure above is stated with the engine and source it came from, and each links to its own page below.
The landscape, in five findings
Google ranking no longer strongly predicts AI citation. Overlap ranges from ~12% (ChatGPT/Gemini/Copilot) to ~38% (AI Overviews, down from ~76% a year earlier) — see most-cited domains and Ahrefs' underlying measurement.
Engines barely agree with each other. Reported ChatGPT/Perplexity domain overlap sits around 11% — see cross-platform concordance. The same divergence is described qualitatively by Discovered Labs, Leapd and Profound.
The schema debate has real evidence on both sides, and it's unresolved. Observational data — Ahrefs' correlational look included — shows no clear lift; live-fetch tests show schema often ignored entirely — see the schema RCT design. The llms.txt story is the same shape: 300,000 domains, no clear effect, and Google calling it speculative.
Three separate studies collapsed into one "58%" figure that gets misapplied across contexts — the clearest documented case of statistic conflation in the field, per the provenance audit.
Not one randomized, causal test of a GEO tactic existed in public before this year's pre-registrations. Every published finding in the field prior to this site's Tier 3 program is correlational.
of AI citations on ChatGPT, Gemini and Copilot (grouped) also rank in Google's top 10 for the same query.
domain overlap reported between ChatGPT's and Perplexity's citation sets — two engines answering the same questions from largely different sources.
randomized, causal tests of a GEO tactic existed in public before this year's pre-registrations.
- Statistics & platform pages 20
- Pre-registered studies 15
- Reference & framework 9
- Definitions & glossary 6
What these five findings mean together
Read individually, each finding above is a discrete fact about one narrow slice of AI search. Read together, they describe something more structural. The traditional shortcut, rank well on Google, get cited by AI, no longer reliably works — even though Seer found one engine matching Bing's results at 87%, which is exactly the kind of single-engine result that does not generalise. A strategy tuned for one engine doesn't transfer to the next.
Even a widely recommended technical tactic, schema, has genuinely mixed evidence, not a settled answer. Sloppy statistic handling is common enough to produce a documented three-studies-collapsed-into-one case. And almost nothing has been tested with a method that could actually support a causal claim. Put plainly: the practical implication isn't "GEO doesn't work." It's "almost nobody has actually tested what works." That's precisely the gap this site's research program exists to close.
Fifty pages into building an independent AI-search research site, the finding holds: the field has an enormous amount of published conclusion and very little published evidence.
Share on XThe five biggest unanswered questions
- What does Claude actually cite? The least-measured major surface — see Claude citation statistics.
- How many sub-queries does AI Mode really generate per question? Folklore says 8–16; nobody has published the data, though Google has described the mechanism — see the Fan-Out Corpus design.
- Does anything a site controls causally change its citation rate? Nearly everything published is correlational — see the Tier 3 RCT program (schema, freshness, author bio).
- How long does a citation last? Every study is a snapshot; none track duration — see citation half-life.
- Do AI crawlers render JavaScript? A binary, foundational, and still-untested technical question — see the rendering experiment, and the bot registry for which crawlers would be affected.
What this site has published so far
Forty-nine research pages as of this snapshot. The Citation Index's pre-registration and its core metrics. Thirteen curated and cross-platform statistics pages. Ten pre-registered original studies. Six Claude-Code-for-SEO practitioner pages, covering failure modes and open tooling. And a reference layer: the bot registry, the tactic scoreboard, the Null-Results Registry, the glossary, and two definitional pillars. Everything is listed under studies, with tools alongside.
What comes next
Data collection for Index v1 begins Q1 2027. Every pre-registered study above moves from design to result on that same timeline. Two surfaces sit outside that schedule because they are too new to measure yet — MCP servers as a visibility channel and WebMCP — and both are tracked as open questions rather than tactics. The next State of AI Search edition, in Q4 2027, will report actual findings against every prediction registered here — including the ones we predicted would fail.
- Sep 2026This snapshot
49 research pages published; no first-party citation data yet.
Published deliberately before any results exist, so a later claim of progress is checkable against a stated baseline.
- Q1 2027Index v1 collection
Data collection begins for the Citation Index and every registered study.
Design to result on one shared query set and one collection window.
- Q4 2027Full annual report
Findings reported against every prediction registered here.
Including the predictions we expect to fail. That is the point of registering them.
Why the evidence layer is thin
This is not a story about laziness. The incentives explain most of it.
The data is the product. A visibility vendor's corpus is what customers pay for. Publishing raw files would give it away. So the method stays private, and the claim cannot be checked.
Null results do not market. A finding that a tactic did nothing sells no software and wins no links — which is why the Null-Results Registry exists as a standing commitment rather than a good intention. The llms.txt literature is the exception that proves it: adoption tracking and the finding that most files are never requested got far less circulation than the original proposal did. The publishing filter therefore selects for positive results.
Causal tests are expensive. A randomised design needs many comparable pages, a control group, and the willingness to leave the control group alone for months. Most sites will not do that.
The systems move. A result can be obsolete before it is written up. That discourages careful work and rewards fast commentary.
Nobody is required to disclose anything. There is no registry, no peer review, and no convention that a number should carry its method — the gap the measurement standard and the dataset strategy are both aimed at. So most numbers do not.
Hypothesis Our reading is that these incentives, not any shortage of skill, explain the shape of the field. We cannot prove that, and a different explanation may fit the same facts.
What does not transfer between engines
Low cross-engine overlap has a practical consequence that is often skipped. A finding about one engine is a finding about that engine only.
Retrieval differs. One system may lean on a live search index. Another may lean on its own crawl, or on material already inside the model. These produce different source sets from the same question.
Answer format differs. A system that shows numbered citations creates different click behaviour from one that mentions a brand in prose without a link.
Freshness policy differs — a registered experiment here, and separately how long a citation survives at all, measured in the half-life study. Some surfaces clearly prefer recent material. Others repeat the same source for months. A freshness result on one is not a freshness result on all.
User population differs — Similarweb's category usage data, Statista's ChatGPT user series and StatCounter's search share each describe a different population, as the market share page explains. The questions people bring to each product are not the same, so the query mix behind any aggregate is different too.
The rule this site follows is to report per engine, and to treat any blended figure as a description of the blend rather than of AI search in general.
Null results we would publish
- No effect from any registered tactic. If the randomised studies find nothing, that result gets the same prominence as a positive one.
- Higher engine overlap than reported. If our own measurement contradicts the low-overlap figures we have cited, we will say so and revise the finding above.
- No half-life effect. If citations turn out to be stable over time, the freshness framing loses its motivation and we will retire it.
- Our own traces being wrong. If a statistic we called untraceable turns out to have a solid origin we missed, the correction gets published with the same weight as the original criticism.
What would change this page
An operator publishing a real description of how sources are selected for an answer would change more of this page than anything else on this list.
A vendor releasing a raw, dated corpus with a documented sampling method would move several findings from vendor-reported to checkable overnight.
A large-sample observational study with a fully published corpus — the shape of Semrush's AI Overviews work or Ahrefs' click-loss measurement, but with the underlying data attached — would move several findings at once.
A second independent group running a randomised test would let us compare designs instead of relying on a single programme, which is itself a weakness of the current picture.
Any of these would be good news, and would be treated as such. This page is not invested in the field staying thin.
Open questions beyond the top five
- Does being cited in an AI answer produce measurable business value, or only visibility? Almost nothing public addresses this.
- How stable is an engine's source set for the same question asked twice, an hour apart?
- Do answers differ by country and language in ways that change which domains get named?
- What share of AI answers name no source at all, and is that share moving?
- Does a site being blocked to one crawler affect its appearance in a different company's answers?
- Does authorship or credential signalling change anything, as the registered author-bio RCT is designed to test, and do brand mentions outperform links as a correlate?
- Do agentic browsers, MCP servers and agent-readable endpoints become a separate visibility surface, or fold back into the ones measured here?
How this snapshot was built
A first-party inventory and synthesis of this site's own published research as of the date shown, cross- referenced against the provenance grades established throughout and the standards described in about. No new data collection underlies this specific page — it's a synthesis of work already published and linked above. This is a dated snapshot: it describes September 2026 and will be replaced rather than silently edited.
Next: the tactic-by-tactic version of finding three is the GEO tactic evidence scoreboard. To have the next snapshot and each study arrive as it reports, the newsletter carries them.
Namdev, R. (2026). State of AI search: September 2026 data report (v1). Retrieved from https://ritiknamdev.com/blog/state-of-ai-search Published under CC BY 4.0 — reuse freely with attribution.
For the traced, source-by-source figures behind this snapshot, see the AI SEO statistics hub, the parent page for every number quoted here. Start with the AI Citation Index if you're new to this research program.