A critical survey of GEO reviewed 45 studies and found no technique with a stable, longitudinal, cross-platform causal effect

Does GEO work? The survey's own confidence ratings, claim by claim

Re-read at the source on September 29, 2026: arXiv still lists one version (v1, July 15, 2026), so nothing in the paper has changed. The survey answers the question directly in its Table 5, “Confidence in the field's principal claims.” The ten rows below are that table, in the paper's wording.

Read as a buyer: the four high-confidence rows describe what happens to a page an engine has already retrieved, and the fact that engines differ. The two claims most AEO retainers are sold on, durable discoverability across engines and citations that turn into revenue, are the rows rated low and very low.

Corrections made on September 29, 2026. Three statements in the first version of this post went further than the paper. (1) We wrote that 53% of domains cited in AI Overviews are absent from the organic top 100. The survey reports, from Kirsten et al., that 53% do not appear in the organic top 10 and 27% are absent from the top 100. (2) We wrote that Schulte et al. recommend “a minimum” of 7–8 repetitions per prompt. The survey calls seven to eight repetitions “a starting point” and adds that the number “is not a universal standard: it derives from a small universe of Swiss queries and at most ten repetitions.” (3) We wrote that the survey sorted the 45 studies into evidence levels and that “Level A is nearly empty.” The survey's hierarchy (its Table 7) grades a specific claim, not a paper, and the phrase was ours. What the paper says is that very few studies observe clicks, traffic or conversions, and it titles that section “The Weakest Evidence.” We also called this the first critical survey of the field; we have not verified that and have removed the word.

What the survey did

The primary source is arXiv:2607.14035v1, dated July 15, 2026 (cs.IR, cs.DL). It searched arXiv, the ACM Digital Library, the ACL Anthology, NeurIPS, PMLR and OpenReview with backward and forward citation tracking, and proposes a five-level framework for grading any GEO claim, from Level A (a randomized field trial or strong quasi-experiment with logs and controls) down to Level E (a fixed context, synthetic ranker, or LLM judge). The level grades a specific claim, not a whole paper. The finding that should give every buyer pause: very few studies in the corpus observe clicks, traffic or conversions at all, and the survey files that evidence under the heading “The Weakest Evidence.” The most-cited foundational result in this field was measured in a fixed five-document context, and most of what circulates as an “AEO playbook” descends from it.

The headline numbers, re-read at their source

None of this says AEO work is worthless. It says the effect sizes in circulation were measured under conditions — a supplied set of documents, a single engine, a short window — that do not describe a client's live visibility, and that the one attempt in the corpus at a controlled field measurement came back positive but tentative, with its own placebo test failing to clear significance.

The part we can corroborate from our own data

The survey's methodological core is a sentence we have been arguing from first-party measurements for six weeks: “Visibility is a distribution. It depends on the engine, date, location, query formulation, whether search is actually activated, and stochasticity in generation. A point estimate is not a stable indicator.” The evidence it cites for that: Schulte et al. (2026) measured source-overlap Jaccard of 0.34–0.42 across four engines over 45 days and propose seven to eight repetitions per prompt as a starting point, and found 57.8% of ChatGPT repetitions did not activate web search at all; Kirsten et al. (2026) found that even at temperature zero, repeated runs change 9–28% of decisions; URL-level overlap between Google organic, AI Overviews and Gemini is 0.11–0.18, and 53% of domains cited in AI Overviews do not appear in the organic top 10 (27% are absent from the top 100).

Our own run: 25 buyer questions × 4 engines × 3 repetitions, 258 completed answers, 681 brand-recommendation slots — and only 31.4% of recommended brands appeared in every repeat of the identical question on the identical engine (the full dataset). Different method, same conclusion: one answer is one draw from a distribution.

Where this cuts against us, stated plainly

Schulte's starting point of seven to eight repetitions is above what our own $79 audit does — it samples each of 25 questions three times per engine (300 answers), and the free check samples once. Three repetitions is enough to expose flicker and to report a rate instead of a verdict; it is not enough to pin a mention rate to a tight confidence interval, and we are not going to claim otherwise. The survey's own advice is to repeat a measurement until the interval around the estimate is narrow enough for the decision at hand. The honest framing of any single audit — ours included — is a floor with a date and an engine attached, re-measured over time.

Why this matters for businesses and agencies

• Treat any AEO uplift promise as un-evidenced until it names its level. The survey's grading is the useful buyer's tool: ask whether a claimed result came from a randomized field trial with logs (Level A) or from a fixed-document experiment (Level E). Most vendor decks — and most of the “40%” slides — are quoting Level E as though it were Level A.
• “Transfers poorly” is the operative phrase. Only topical relevance and positioning replicated across the corpus; general heuristics did not. A checklist that worked for one client's category is a hypothesis for the next one, not a deliverable.
• Visibility is indexed by engine and surface. With 0.11–0.18 URL overlap between Google organic, AI Overviews and Gemini, a rank report and an AI-citation report are measurements of different things. Reporting one as if it covers the other is the most common measurement error we see.
• Under broad adoption, gains compress. C-SEO Bench's “gains approach zero under broad adoption” is the strategic caveat: the techniques being sold today assume most competitors are not doing them. Measurement of where you actually stand keeps its value either way; the tactic-of-the-month does not.

The takeaway

This audit of the field's evidence base concluded that the interventions are real but narrow — they can change what happens to content the engine has already retrieved, and nobody has yet shown a technique that durably moves organic discoverability across engines. That is an argument for spending less on portable playbooks and more on repeated, per-engine measurement of the thing you are actually trying to move.

You can run the free 60-second check to see a business's mention rate across ChatGPT and Perplexity as separate numbers (the $79 audit adds Gemini and Claude, 25 questions sampled three times each — a rate, with its sample size stated). Agencies scoping AEO retainers across a book of clients can baseline five businesses at once with the $249 agency 5-pack — which is the measurement this survey says should come before any promise about what the work will move.