Do GEO Tactics Survive Replication? What 2026 Re-Tests Found

Short answer: the three tactics every GEO guide recommends — adding statistics, adding expert quotations, and citing sources — did not produce a measurable lift when re-tested at scale in September 2026. An arXiv preprint by Elisha Bajemon and Andre-Louis Rochet (7 September 2026) applied them as volume-controlled paired edits and measured no positive pooled effect: quotations −0.325 percentage points, statistics −0.276, and cite-sources −0.793 with a confidence interval excluding zero. That does not mean GEO is dead. It means the industry has been treating context-dependent tactics as universal laws, and 2026 is the year the measurement caught up.

This page is the honest version of the evidence. If you have built a content programme on statistics, quotations and citations, most of what you are doing still helps — but probably not for the reason you were told, and almost certainly not as much as a +33% or +41% headline suggests. What follows is what the replications actually tested, what they did not test, and how to restructure a GEO programme around what survives.

What the Original Princeton Study Claimed

The Princeton GEO paper (Aggarwal et al., arXiv:2311.09735, KDD 2024) tested nine content interventions and reported visibility lifts of +33% for statistics, +41% for expert quotations, +28% for source citations, and a −8% penalty for keyword stuffing. Those four numbers became the founding citation of commercial GEO, quoted in nearly every vendor deck and most guides on this site included. They were measured on a purpose-built benchmark, not on live engines.

That distinction is not a technicality. A benchmark fixes the query, the candidate set and the retrieval conditions. Live AI search fixes none of them. The gap between the two is exactly where the 2026 replications went looking — and where the numbers moved.

"Three popular GEO tactics showed no positive pooled lift when re-tested with volume-controlled edits. That does not mean GEO does not work. It means the query-blind score weakly predicted citation ordering, while the three historical interventions did not transfer positively."

What the September 2026 Replication Actually Tested

It tested two separate objects: a deterministic 0–100 query-blind page score built from 11 text features, and a query-conditioned predictor that adds query-source relevance. The score used entropy, information and entity density, semantic coherence, self-containment, citation F1, statistic density, MMR, nDCG, redundancy and quotable density. The outcome measured was a visibility share in generated answers, not Google rank, traffic, leads or revenue.

Source: Bajemon & Rochet, arXiv preprint, 7 September 2026. Ranking data contain 229 unique queries and 1,269 query-engine groups with five candidate sources per group. This is a preprint and has not been peer reviewed.

Why the Tactics Failed — and Why That Is Not the Whole Story

The replications edited volume, not relevance. A statistic pasted into a passage that was never about that statistic does not improve query fit; it adds tokens. The study design makes this explicit: the edits were paired and volume-controlled, applied to existing pages, which is a very different intervention from commissioning original research on the question your buyers actually ask.

The confidence intervals matter here, and most summaries of this paper skip them. The quotation interval ran from −0.841 to +0.171 and the statistics interval from −0.970 to +0.425. Both include zero, so the correct reading is "no detected effect," not "a proven negative effect." Only cite-sources produced an interval that excluded zero on the negative side.

There is also a scale question nobody has resolved. The paper measured visibility share among five candidate sources for the same query. It did not measure rankings, traffic, leads or revenue — the layers that decide whether a content programme is worth funding. Weak ordering correlation among a fixed candidate set is not the same as zero commercial effect across a portfolio of pages.

What Actually Predicts AI Citations in 2026

Query-context fit, off-page entity presence, and freshness — in that order. The preprint's own ablation is the cleanest evidence: adding query-source relevance lifted correlation from 0.114 to roughly 0.37, and correcting query leakage deflated a fitted model from 0.5724 to 0.3625. Anything that looks like a big win without query conditioning should be treated as leakage until proven otherwise.

The off-page evidence points the same direction. Ahrefs' Brand Radar study of 75,000 brands (3 September 2026) correlated every major marketing signal against how often ChatGPT, Google AI Mode and AI Overviews mention a brand:

The practical translation: being named in more places beats being optimised harder. A page-level formatting edit is a small lever; getting your brand, product and people discussed across YouTube, podcasts, newsletters and comparison content is a large one. The 2026 evidence consistently ranks off-page entity presence above on-page tactics.

The Measurement Problem Nobody Wants to Talk About

Most GEO reporting is noise, because most GEO tooling samples once. Across 536 prompt-and-engine combinations with at least five checks, 80% of appearance cases were inconsistent, and a single check matched the majority result only 72.2% of the time. The working rule from practitioners is 30 to 40 runs per prompt before volatility flattens; SignalForge runs 40 runs per prompt for under 2% margin of error.

The second reporting failure is conflating mention with recommendation. Latent Space's Frontier AI visibility tracker, covering 6,762 answers across 161 categories, found Kysely appearing in 42 of 42 answers but named first choice exactly once. Category leaders also differed between GPT-5.6 Sol and GPT-6 Astra in 33 of 121 comparable categories. Share of voice built on mentions will systematically overstate commercial position.

"If AI is already talking buyers out of purchases, does your GEO reporting still stop at whether you were mentioned?"

That question is now empirically grounded. A Semrush and Exploding Topics survey of 2,338 US adults (7 September 2026) found 57.5% of AI users said chatbot information caused them not to buy, while 65% said AI had at least partially replaced Google for product research. Among weekly AI users, 73.6% bought from an organic AI recommendation. AI answers both create and destroy demand, and a dashboard that only counts presence cannot see the difference.

A GEO Programme Built on What Survives

Six changes, ordered by the strength of the evidence behind them. None of them require abandoning the tactics that failed to replicate — they require repositioning those tactics from growth levers to hygiene.

1. 1.
   Sample 30 to 40 runs per prompt before reporting anything

   With 80% of appearance cases inconsistent across five checks, single-run dashboards are measuring volatility, not visibility. Budget the runs, or report ranges instead of point estimates.

2. 2.
   Separate appearance, citation, recommendation and sentiment

   A brand can be mentioned in every answer and chosen in none. Track the recommendation role and the sentiment of the mention separately from raw presence, then align reporting to pipeline rather than share of voice.

3. 3.
   Optimise for query fit, not page score

   Query conditioning roughly tripled the correlation in the preprint (0.114 to ~0.37). Write each section to answer one specific sub-query, and verify it against the actual prompt your buyer uses.

4. 4.
   Move budget from on-page formatting to off-page entity presence

   YouTube mentions correlate 0.737 and unlinked web mentions 0.656 to 0.709, against roughly zero for backlink count. Podcasts, comparison content, review sites and community threads are now higher-yield than another round of on-page editing.

5. 5.
   Keep statistics, quotations and citations — as credibility, not as lift

   They did not replicate as visibility levers, but they remain how a passage earns trust with a human reader and signals sourcing to a retriever. Just stop forecasting +33% from them.

6. 6.
   Track each engine separately

   Intender found only 8% source overlap across five engines and 71% of domains appearing on one engine alone. A blended share of voice hides more than it reveals.

Where We Expect the Evidence to Move Next

Three open questions will decide whether the 2026 replications hold. First, whether the interventions behave differently when the added statistics and quotations are genuinely query-relevant rather than volume-matched — the preprint lists this as unknown. Second, whether the results transfer to live consumer surfaces rather than API output, since UI and API results are known to diverge. Third, whether model-version dependence swamps content effects entirely: category leaders already differ between GPT-5.6 Sol and GPT-6 Astra in 33 of 121 categories, and Google began rolling Gemini 3.8 Flash into AI Mode in September 2026.

Until those are settled, the defensible position is uncomfortable but simple: keep doing the tactics, stop selling them as percentages.

Frequently Asked Questions