Which GEO studies are actually worth citing?

Prioritize peer-reviewed work with published methodology and reproducible setups — the Princeton GEO-bench paper (Aggarwal et al., KDD 2024) is the clearest example. Below that, large-scale industry datasets with disclosed sample sizes are usable directionally. Vendor blog posts with no stated methodology should not appear in a strategy document at all.

How do I tell whether a GEO statistic has been laundered?

Trace it back three hops. A laundered statistic is one where every source cites another blog post rather than a primary document, the number drifts slightly between retellings, and no version links to a paper, dataset, or methodology page. If you cannot reach a primary source in three clicks, drop the number.

What is an evidence tier and why does it matter for content strategy?

An evidence tier is a grade assigned to a source based on methodology strength — peer-reviewed experiment, large-scale observational dataset, small-sample scrape, or unsourced claim. It matters because AI answer engines increasingly reward content that attributes claims to verifiable primary sources, and because building a roadmap on a Tier 4 statistic wastes real budget.

How much of the GEO literature do I actually need to read?

Less than you think. Roughly six kinds of source carry the practical weight: the Princeton tactic experiments, one large-scale CTR dataset, one behavioral panel study, one positional-retrieval paper, one platform-architecture disclosure, and your own first-party analytics. The remaining literature is useful for depth and for defending recommendations, not for deciding what to do first.

How often does the GEO literature go stale?

Faster than most fields. Platform-behavior findings — trigger rates, citation formats, crawler behavior — can be invalidated within a quarter by a product change. Mechanism findings about retrieval and ranking age far more slowly. Re-verify anything platform-specific every 90 days and treat architecture papers as durable.

How to Grade GEO Research Evidence and Spot Laundered Statistics | MarketLens

Read twenty articles about generative engine optimization and you will encounter roughly the same six statistics, quoted with escalating confidence and decreasing accuracy. A percentage from a peer-reviewed experiment gets reproduced without its experimental conditions. A vendor’s internal sample becomes “research shows.” A number invented for a conference slide acquires a fake citation on its fourth retelling.

This matters more in GEO than in traditional SEO for a specific reason: answer engines are trained to prefer content that grounds claims in verifiable sources, and they are increasingly capable of noticing when a claim has no source. Building a strategy on laundered numbers is bad for the strategy and bad for the citations.

What follows is deliberately not a summary of the literature. It is a grading system for reading it — how to sort sources into tiers, which handful genuinely drives decisions, and how to maintain a research ledger so your own content survives scrutiny.

Four tiers, and why the boundary between them is where money is lost

Every claim in the GEO literature falls into one of four buckets. Assign the tier before you assign the budget.

Tier	What it is	Example characteristics	How to use it
1 — Controlled experiment	Peer-reviewed, published methodology, reproducible	Named authors, venue, dataset released, effect sizes with conditions	Safe to build a program on
2 — Large-scale observational	Big real dataset, disclosed sample, no control group	Millions of impressions, stated time window, portfolio bias acknowledged	Directional; validate on your own data
3 — Small-sample or scraped	Hundreds to low thousands of SERPs, one snapshot	No position control, no time series, often vendor-published	Hypothesis generation only
4 — Unsourced assertion	Confident number, no traceable origin	“Studies show”, no link, number varies across retellings	Do not cite; do not plan against

The costly mistake is not citing a Tier 4 number in a blog post. It is treating a Tier 3 finding as if it were Tier 1 when allocating a quarter of content budget. A single-snapshot scrape of 2,000 keywords cannot support a claim about causation, no matter how confidently it is charted.

The six sources that actually drive decisions

Of the broad GEO literature, a small core does the real work. Everything else is depth.

Controlled tactic experiments. The Princeton GEO-bench work (Aggarwal et al., KDD 2024) remains the only widely available Tier 1 evidence that specific content edits change generative visibility. Read the paper itself rather than a summary of it, and transcribe the effect sizes from the PDF — the per-tactic numbers are among the most frequently garbled in the whole field. What survives retelling reliably is the ranking of tactics, and the ranking is the actionable part, because it tells you what to do first. Why those particular tactics work is the subject of measuring information gain and content novelty.

Positional retrieval research. The finding that retrieved passages cluster disproportionately in the opening portion of a document is a structural mechanism, not a platform quirk, which is why it ages well. It converts directly into a layout rule.

Large-scale CTR observation. One good portfolio-scale impression dataset to establish the direction and rough magnitude of AI Overview erosion.

Behavioral panel data. One instrumented-browser study to explain the mechanism behind the CTR numbers — specifically session termination, which server-side data cannot see. The Pew Research Center’s instrumented study of US adults’ Google sessions is the clearest publicly available example, and it is worth reading for the behaviour it captures rather than for a headline percentage.

Platform architecture disclosures. Engineering blogs and conference talks from the retrieval vendors themselves. These are primary sources about how the system works, and they are chronically under-read by SEO practitioners.

Your own analytics. Tier 1 for your site specifically, and the only source that can tell you effect size in your vertical.

Everything beyond these six adds nuance, defensibility, and depth — which matters most when you are writing for a regulated or high-scrutiny vertical, where an unsourced number is a liability rather than merely an embarrassment.

How to trace a statistic in three hops

Take any number you are about to publish and follow it backwards.

Hop one: open the source your source cited. If it is another blog post, continue.
Hop two: open that post’s source. If it is also a blog post, you are already in laundering territory.
Hop three: if you have not reached a paper, a dataset, a methodology page, or a first-party platform document, stop and delete the claim.

Two additional tells worth knowing. Number drift — the same statistic appearing as 41%, “over 40%”, and “roughly 45%” across retellings — indicates nobody in the chain read the original. And condition stripping is the most common failure of all: an effect measured under specific experimental conditions gets quoted as a universal law. A tactic that produced a large lift in a controlled benchmark on a particular query set is evidence that the tactic can work, not a guarantee of that number on your pages.

Maintaining a research ledger

The practical artifact is a single versioned file that every content brief references. Keep it in the repo next to your content so it gets reviewed like code.

# research-ledger.yml — every externally-sourced number lives here.
# No entry, no number on the site.
claims:
  - id: geo-quotation-lift
    claim: "Adding authoritative quotations increases generative visibility"
    value: TRANSCRIBE_FROM_PDF     # never copy this figure from a blog retelling
    tier: 1
    primary_source: "Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024"
    conditions: "GEO-bench query set; visibility measured inside generated responses;
                 effect size varies by query domain"
    review_cycle_days: 365
    last_verified: 2026-08-31

  - id: dentalclinica-clicks-yoy
    claim: "Organic clicks, 1 May - 30 Aug, year over year"
    value: "8,122 -> 18,747 (+130.8%)"
    tier: 1                        # first-party: reproducible from the API on demand
    primary_source: "GSC search_analytics, dimensions=[device],
                     2025-05-01..2025-08-30 vs 2026-05-01..2026-08-30"
    conditions: "One site, one vertical, recovering from a two-year decline —
                 not a general effect size"
    review_cycle_days: 90
    last_verified: 2026-08-31

  - id: aio-healthcare-penetration
    claim: "AI Overview trigger rate in healthcare queries"
    value: null
    tier: 4                        # traced three hops, never reached a primary source
    primary_source: null
    action: "removed from all published pages 2026-08-31"

Two fields earn their keep. review_cycle_days encodes decay: platform-behavior claims go stale in about a quarter, while mechanism claims about retrieval and ranking hold for a year or more. And the Tier 4 entry is kept deliberately rather than deleted — a record of a number you used to publish stops someone reintroducing it next quarter from a stale draft.

Reading a study you have not read before

Four questions, in order, applied to any GEO paper or industry report before you decide whether it changes anything you do.

What was the unit of measurement? “Visibility” means different things across the literature — appearance in a generated response, position within it, word count attributed to a source, citation link presence. Two studies reporting a 30% improvement may be measuring quantities that are not comparable at all. Find the operational definition before comparing numbers.

What was the control? A study measuring content with a tactic against content without it is measuring the tactic. A study measuring optimized pages against a random sample of the web is measuring quality in general. The second is far more common and far less useful.

Over what window? A single-day SERP capture cannot distinguish an effect from ordinary volatility. Anything claiming causation over less than a few weeks should be read as a hypothesis.

Who benefits from the result? Not disqualifying — vendors run some of the most rigorous available research, and they have the data nobody else does. But a study whose conclusion is precisely the service its publisher sells warrants a closer look at methodology than one that does not.

Running these four questions takes about ten minutes per source and eliminates most of what would otherwise end up in a strategy deck.

What the literature still does not answer

Honest inventory of the open questions, because claiming otherwise is how Tier 4 statistics get created:

Attribution at the query level. No public method reliably connects a specific content change to a specific citation appearing in a specific engine.
Cross-engine transfer. Whether a tactic that lifts visibility in one retrieval architecture lifts it in another is largely untested.
Durability. Almost every published effect is measured over weeks. Nobody has published a multi-year decay curve.
Interaction effects. Tactics are measured in isolation. Whether stacking four of them compounds or saturates is unknown.

Naming these gaps costs you nothing and buys real credibility. It is also, in practice, the sort of passage that gets quoted verbatim by answer engines summarizing “limitations of GEO research” — a small demonstration of the principle that specific, verifiable, well-attributed prose is what gets extracted. The structural side of that argument, about where in a document such prose has to sit, is in structural extractability and positional retrieval bias.

Start here

Audit the numbers already live on your site. Pull every statistic you have published, trace each one three hops, and delete or re-attribute whatever fails. Expect the pass to be more destructive than you assume going in — and the version that survives it is the one an answer engine can safely quote.

Then keep the ledger. It takes ten minutes a month and it is the difference between a strategy built on evidence and one built on repetition.